Documente Academic
Documente Profesional
Documente Cultură
Second Edition
Volume 4
Measuring, Modelling, and Control
Biotechnology
Second Edition
Fundamentals Special Topics
Volume 1 Volume 9
Biological and Biochemical Fundamentals Enzymes, Biomass, Food and Feed
Volume 2 Volume 10
Genetic Fundamentals and Special Processes
Genetic Engineering
Volume 11
Volume 3 Environmental Processes
Bioprocessing
Volume 12
Volume 4 Patents, Legislation, Information Sources,
Measuring, Modelling, and Control General Index
Products
Volume 5
Genetically Engineered Proteins and
Monoclonal Antibodies
Volume 6
Products of Primary Metabolism
Volume 7
Products of Secondary Metabolism
Volume 8
Biotransformations
Distribution:
VCH, P. 0. Box 101161, D-6940 Weinheim (Federal Republic of Germany)
Switzerland: VCH, P. 0. Box, CH-4020 Base1 (Switzerland)
United Kingdom and Ireland: VCH (UK) Ltd., 8 Wellington Court, Cambridge CB1 1HZ (England)
USA and Canada: VCH, Suite 909,220 East 23rd Street, New York, NY 10010-4606 (USA)
Biotechnology
Second, Completely Revised Edition
Edited by
H.-J. Rehm and G. Reed
in cooperation with
A. Puhler and P. Stadler
Volume 4
Measuring, Modelling,
and Control
Edited by
K. Schugerl
- -
Weinheim New York . Base1 Cambridge
Series Editors: Volume Editor:
Prof. Dr. H.-J. Rehm Dr. G. Reed Prof. Dr. K. Schiigerl
Institut fur Mikrobiologie 1016 Monmouth Ave. Institut fiir Technische Chemie
Universitiit Miinster Durham, NC 27701 Universitiit Hannover
CorrensstraBe 3 USA CallinstraBe 3
D-4400 Miinster D-3000 Hannover 1
Dr. P. J. W. Stadler
Prof. Dr. A. Piihler Bayer AG
Biologie VI (Genetik) Verfahrensentwicklung Biochemie
Universitiit Bielefeld Leitung
P.O. Box 8640 Friedrich-Ebert-StraBe 217
D-4800 Bielefeld 1 D-5600 Wuppertal 1
This book was carefully produced. Nevertheless, authors, editors and publisher do not warrant the information con-
tained therein to be free of errors. Readers are advised to keep in mind that statements, data, illustrations, procedural
details or other items may inadvertently be inaccurate.
Published jointly by
VCH Verlagsgesellschaft mbH, Weinheim (Federal Republic of Germany)
VCH Publishers Inc., New York, NY (USA)
Editorial Director: Dr. Hans-Joachim Kraus
Editorial Manager: Christa Maria Schultz
Production Director: Maximilian Montkowski
Production Manager: Peter J. Biel
In recognition of the enormous advances in The present volume, fourth in the series, re-
biotechnology in recent years, we are pleased flects the enormous impact of computer tech-
to present this Second Edition of “Biotech- nology on biotechnology, especially in the
nology” relatively soon after the introduction areas of measurement and control. It describes
of the First Edition of this multi-volume com- monitoring of the biotechnological process
prehensive treatise. Since this series was ex- with sophisticated analytical techniques, use of
tremely well accepted by the scientific commu- the resulting data by means of mathematical
nity, we have maintained the overall goal of models, and computer-aided closed loop con-
creating a number of volumes, each devoted to trol for improvement of the productivity of
a certain topic, which provide scientists in biotechnological processes. While Volume 4
academia, industry, and public institutions can be used independently, Volume 3 “Bio-
with a well-balanced and comprehensive over- processing” is recommended as a companion
view of this growing field. We have fully re- volume.
vised the Second Edition and expanded it from The volume editors and the authors of the
ten to twelve volumes in order to take all re- individual chapters have been chosen for their
cent developments into account. recognized expertise and their contributions to
These twelve volumes are organized into the various fields of biotechnology. Their will-
three sections. The first four volumes consider ingness to impart this knowledge to their col-
the fundamentals of biotechnology from bio- leagues forms the basis of “Biotechnology”
logical, biochemical, molecular biological, and and is gratefully acknowledged. Moreover,
chemical engineering perspectives. The next this work could not have been brought to fru-
four volumes are devoted to products of indus- ition without the foresight and the constant
trial relevance. Special attention is given here and diligent support of the publisher. We are
to products derived from genetically engi- grateful to VCH for publishing “Biotechnolo-
neered microorganisms and mammalian cells. gy” with their customary excellence. Special
The last four volumes are dedicated to the de- thanks are due Dr. Hans-Joachim Kraus and
scription of special topics. Christa Schultz, without whose constant ef-
The new “Biotechnology” is a reference forts the series could not be published. Finally,
work, a comprehensive description of the the editors wish to thank the members of the
state-of-the-art, and a guide to the original Scientific Advisory Board for their encourage-
literature. It is specifically directed to micro- ment, their helpful suggestions, and their con-
biologists, biochemists, molecular biologists, structive criticism.
bioengineers, chemical engineers, and food
and pharmaceutical chemists working in indus- May 1991 H.-J. Rehm
try, at universities or at public institutions. G. Reed
A carefully selected and distinguished Scien- A. Puhler
tific Advisory Board stands behind the series. P. Stadler
Its members come from key institutions repre-
senting scientific input from about twenty
countries.
Scientific Advisory Board
KARL S C H ~ G E R L
Hannover, Federal Republic of Germany
The fourth volume of the second edition of ess and to effect process optimization and con-
“Biotechnology” presents a survey on an in- trol. Therefore, the Series Editors decided to
creasingly important field of biotechnology: add to Biotechnology a separate volume with
monitoring of the biotechnological process the title “Measuring, Modelling, and Con-
with sophisticated analysis techniques, use of trol”, and they asked the Volume Editor to or-
the resulting data by means of mathematical ganize it.
models, and (computer-aided) closed loop con-
trol for improvement of the productivity of This volume consists of four main parts:
biotechnological processes.
The bottleneck in biotechnological process 0 instruments,
control is the on-line measurement of con- 0 measuring techniques,
trolled process variables. Except for tempera- 0 modelling, and
ture, impeller speed (for stirred tank reactors), 0 control/automation.
aeration rate (for aerobic microorganisms),
pH, po,, which are usually controlled process Cell growth and product formation/sub-
variables, and the composition of the outlet strate conversion is at the focus of attention,
gas (0, and C 0 2 content), which sometimes is and here the measuring and control techniques
a controlled process variable (respiration quo- are well-developed and generally applicable.
tient, RQ, C 0 2 production rate, O2 consump- “Modelling, design, and control” of down-
tion rate), no other process variables are usual- stream processes are also considered because
ly measured on-line in commercial equip- of the great importance of downstream proc-
ment. essing. However, because of their broad scope,
However, manufacturers have recently they are too heterogeneous and not yet suffi-
made great efforts to improve process analysis ciently developed for treatment in the same
and control. In several laboratories, on-line way as processes for growth and product for-
systems for the analysis of the chemical me- mation.
dium composition are used to gain more infor- The instruments are subdivided into three
mation about the process and to control the groups:
concentrations of key components. Further-
more, mathematical models have been devel- 0 common instruments for medium analy-
oped in order to describe the production proc- sis,
2 Introduction
0 instruments for gas analysis, and Models for animal and plant tissue cultures
0 biosensors. have not yet been included because no reliable
kinetic data are available for mathematical
The last group of instruments is still being de- modelling of these cultures.
veloped. Modern control techniques are increasingly
Only instruments that are (or can be) used applied to closed loop control of bioreactor
for process control are considered in detail. systems. Therefore, different types of closed
However, modern off-line techniques (e.g., loop control techniques, including computer-
NMR) are also taken into account. aided control, are considered in detail.
The measuring techniques are subdivided Instrumental control is much more reliable
into four groups: than control by a human operator. Further-
more, long-range (many weeks or months)
0 physical techniques for the characteriza- runs are only possible in the laboratories of re-
tion of fluid dynamics, search institutes and universities if automated
0 chemical methods for the analysis of equipment is used. Thus, automation of bio-
broth composition, reactors is also taken into account.
0 physical methods for the determination Chapter 18 on modelling, design, and con-
of cell concentration, and trol of downstream processing covers only the
0 physical/biochemical methods for the most important downstream processes. Devel-
characterization of the biological state opment in this field is still limited, so this
of the cells. chapter is not as extensive as the potential im-
portance of downstream processing would
Most of these are on-line techniques; others warrant.
can only be carried out in a quasi on-line Expert systems are being developed in dif-
mode. All of them are used to characterize the ferent aspects of technology, medicine, and
reactor/medium/cell system. the natural sciences. Their use in biotechnolo-
There are interesting new developments in gy is desirable, since they would permit the
the characterization of such systems by mod- identification of equipment failures and their
ern mathematical methods, with optimization eventual elimination. Furthermore, by means
of sampling to gain maximal possible informa- of expert systems, large amounts of informa-
tion. These new techniques are also included in tion from on-line and off-line measurements
this volume. as well as from the literature and from heuris-
The models are subdivided into tic knowledge can be used with high efficiency.
The reviews consider only the most impor-
0 cell models, tant techniques and omit some detail because
0 reactor models, and of limited space. Further information can be
0 process models. gained through the reference notations.
It is hoped that information given in this
The process models cover volume will help students, engineers, and
scientists at universities, members of research
0 product formation, institutes, and those in industry to increase
0 aerobic wastewater treatment, their knowledge of this important and fast-
0 anaerobic wastewater treatment, and growing field.
0 models for recombinant microorgan-
isms. Hannover, March 1991 K. Schiigerl
I. Instruments
1 Common Instruments
for Process Analysis and Control
KARL SCHUGERL
Hannover, Federal Republic of Germany
1 Introduction 6
2 Use of the Variables Temperature, pH, po,, &, and pco, for Process Analysis, Optimization,
and Control 6
2.1 Temperature and pH 6
2.2 Dissolved Oxygen Partial Pressure, po, 7
2.3 Redox Potential, Eh,and Dissolved C 0 2 Partial Pressure, pcO2 9
3 Instruments for Determination of Physical System Properties 10
3.1 Temperature 10
3.2 Pressure 10
3.3 Liquid Level and Holdup 10
3.4 Liquid Throughput 11
3.5 Power Input 13
3.6 Liquid Viscosity 13
3.7 Foaminess 14
4 Instruments for Determination of Chemical System Properties 15
4.1 pH Value 15
4.2 Dissolved Oxygen Partial Pressure, po, 16
4.3 Redox Potential, E h 18
4.4 Dissolved COz Partial Pressure, pCo2 19
5 Performance of Instruments for Process Control 20
6 Future Developments 22
7 References 23
6 I Common Instruments for Process Analysis and Control
control the growth or production process of gen and substrate uptake rates, this control of
aerobic microorganisms. the substrate feed is very popular.
Furthermore, Po,-electrodes are used as re- Under steady-state conditions, the oxygen
search. tools for determining oxygen transfer uptake rate, OUR, and the oxygen transfer
rates (OTR) in pioreactors, biofilms, pellets, rate, OTR, are identical. Knowing the driving
and cells immobilized in beads. force for the oxygen transfer, (0,- OZ), the
The use of oxygen balancing for real-time volumetric mass transfer coefficient, KLa, can
estimation of the biomass concentration was be calculated:
recommended first by HOSPODKA (1966). ZA-
BRISKIE and HUMPHREY(1978) worked out 0TR
this technique of observation in detail. Using KLa =
(02 - 02)
the relationship between oxygen uptake rate
(OUR), the yield coefficient of the cell growth where 0, and 03 are the concentrations of the
with regard to the oxygen consumption, dissolved oxygen in the bulk and at the inter-
Yx/02, the maintenance coefficient with regard face (in equilibrium with the gas phase). By
to the oxygen consumption, moZ/X, and the measuring the oxygen balance during cell culti-
growth rate dX/dt: vation, the volumetric mass transfer coeffi-
cient can be calculated in real time.
1 d x In cell-free systems, KLa can be determined
OUR=mo2,X+ -- (3)
yx/o, dt by non-stationary or stationary measurements.
The non-stationary method is based on the re-
The cell mass concentration X can be calcu- lationship :
lated:
do2
-= KLa(Oj- 0,) (7)
dt
in ideal systems.
By measuring the variation of the dissolved
oxygen concentration in the bulk as a function
where X, is the initial biomass concentration. of time, and calculating the dissolved oxygen
Eq. (4) forms the basis for estimating the concentration at the interface from the oxygen
biomass concentration X(t) from the oxygen concentration in the gas phase, KLa can be
uptake rate. evaluated from Eq. (7). However, the interre-
The growth rate may be approximated using lationships between sorption rate and driving
Eqs. (3) and (4): force are in practice more complex. Several re-
lationships have been recommended for this
calculation.
A good review of these methods is given in a
‘Report of a Working Party on Mixing’ of the
With this method the biomass concentration European Federation of Chemical Engineering
of Saccharomyces cerevisiae was estimated. (LINEKand VACEK, 1986) and in the review
Several other authors used this technique in article of LINEKet al. (1987).
combination with the respiration coefficient, Several papers consider the mass transfer of
RQ, to estimate the biomass (e.g., COONEYet dissolved oxygen into biofilms, pellets, and
al., 1977; WANG et al., 1977; PERINGER and cells immobilized in beads. The dissolved
BLACHERE, 1979; TAKAMATSU et al., 1981). oxygen concentration profiles are determined
SQUIRES(1972) reported that po2was used to by means of micro-oxygen electrodes (BUN-
control the sugar addition to the broth of GAY and HAROLD,1971; CHENand BUNGAY,
Penicillium chrysogenum during penicillin pro- 1981; BUNGAYand CHEN, 1981; BUNGAYet
duction. The sugar feed was increased at a al., 1969, 1983; WITTLERet al., 1986).
high po2value; at a low po2it was reduced.
Since a close relationship exists between oxy-
Use of the Valriables Temperature, pH, p o 2 ,Eh, and pco, 9
2.3 Redox Potential, Eh,and The combination of Eqs. (10) and (11)
Dissolved C 0 2 Partial Pressure, gives
F 1
Pcoz
(
r H = 2 pH+--
R T 2.303 E h )
The oxidation and reduction of a compound
is controlled by the redox potential of its envi- This rH value may vary from rH = 0, corre-
ronment. sponding to a solution in which pH2 = 1 bar
The oxidation-reduction potential of a pair and pH = 0, to rH >42, corresponding to a so-
of reversible, oxidizable-reducible compounds lution with po, = l bar and pH = 0. The rH val-
is related to the equilibrium between the oxi- ue is a function of the pH value. Therefore,
dized (ox) and reduced forms (red) and the for the measurement of the redox potential,
number of electrons involved in the reaction both rH and pH values are needed.
(ne-) (KJAERGAARD, 1977; KJAERGAARD The redox potential is used in practice for
and JOERGENSEN, 1979; THOMPSON and GER- microaerobic cultivations, i.e., at very low dis-
SON, in KJAERGAARD, 1977): solved oxygen concentrations, which cannot be
measured by standard oxygen electrodes. An
ox + n e - P red (8) example is the production of exoenzymes by
Bacillus amyloliquefaciens in continuous cul-
The redox potential of this reaction is given by ture at 0.5% oxygen saturation by means of
the Nernst equation: redox-potential control (MEMMERT and WAN-
DREY, 1987).
RT activity of ox In small stirred tank reactors, the dissolved
E h = E o+ - In (9) COz concentration in the broth can be calcu-
nF activity of red
lated from the gas composition by assuming an
where Eh is the redox potential referred to the equilibrium between the phases. In tower reac-
normal hydrogen electrode, tors and large commercial units, no equili-
Eo is the standard potential of the sys- brium distribution of COz exists between the
tem at 25 "C, when all activities of phases; therefore, the direct measurement of
any reactants are at unity, pco2 can be useful.
R the gas constant, The driving force, (Pco2-pEo2), can be
T the absolute temperature, evaluated from the calculated pEo, at the in-
n the number of electrons involved in terface and the measuredpco2 in the bulk. The
the reaction, CO, production rate, CPR, can be determined
F the Faraday constant. from the evolved gas stream and the gas com-
position.
JOERGENSEN (1941) introduced a concept ana- The volumetric mass transfer coefficient of
logous to the pH, namely the rH, which is de- the C 0 2 desorption is given by
fined as
CPR
(KLa)co2=
rH = - logaH2 (10) (Pco2-PEo2)
where aH2is the activity of hydrogen in the hy- CPR can also be used for the calculation of
drogen-hydrogen ion redox system according the cell mass concentration and the specific
to Nernst. For hydrogen growth rate, ,u. The instantaneous specific
growth rate of Penicillium chrysogenum was
E h =R
- T lnaH2--RT2.303 pH
calculated by Mou and COONEY(1983) by
2F F measuring the CPR during the growth phase.
By monitoring the O2 and/or COz concen-
is obtained. trations in the outlet gas and its flow rate, O2
and/or CO, balances can be calculated and
used for state estimation of biochemical reac-
10 I Common Instruments for Process Analysis and Control
tors (e.g., STEPHANOPOULOS and SAN, 1982). less than 2% from the technical and physical
However, because this state estimation method atmosphere.
is based on measurements of the gas composi- Pressure measurements are necessary for the
tion, it will be discussed in Chapter 2. control of the sterilization and the state of the
outlet gas filter as well as for the evaluation of
the holdup and the partial pressures of the ga-
seous components in the gas and liquid phases.
Membrane pressure gauges are commonly used
in biotechnology, because they are particularly
3 Instruments for suited to aseptic operations. Numerous pres-
Determination of Physical sure gauges are used in the chemical industry
(HIRTE, 1980; ANDREWand MILLER, 1979).
System Properties In biotechnology, the commonly employed
pressure gauges are based on strain and/or
capacitance measurements. The capacitance
3.1 Temperature pressure gauges can measure very small pres-
sure differences; therefore, they are used for
Temperature is the most important control liquid level measurements. For the construction
variable for most biotechnological processes, of the different pressure meters, see HIRTE
including sterilization as well as cell growth (1980) and ANDREWand MILLER(1979).
and product formation. In general, a precision
of f0.5 "C is necessary in the temperature
range from +20 to + 130 "C. Only a few types 3.3 Liquid Level and Holdup
of the various industrial thermometers are
suitable because of this prerequisite (BUSING Measurement of the liquid volume is impor-
and ARNOLD,1980). Most popular are the Pt- tant for filling bioreactors with nutrient solu-
100 (100 ohm at 0 "C and 123.2 ohm at 60 "C) tions, for continuous and for fed-batch culti-
resistance thermometers which are encased in a vations. It can be performed (OEDEKOVEN,
protective steel tube fixed with a sealing com- 1980; ELFERS, 1964; ANDREW and RHEA,
pound of high heat conductivity. 1970) as follows:
According to DIN 43 760 (German Stand-
ard) the resistance of these instruments is guar- 0 by measuring the hydrostatic pressure
anteed with the following precision: 1OOf 0.1 difference between the bottom of the
ohm at 0 "C, which corresponds to an error of reactor, P b , and the head space, P h , by
k 0.26 "C. Therefore, these resistance ther- means of pressure gauges. The pressure
mometers can be used without calibration. difference is proportional to the weight
However, the resistances of all electrical con- of the liquid in the reactor:
nections must be controlled. These instruments
are steam-sterilizable at 121 "C. Thermometers
with short response times for fluid dynamical
measurements are described in Chapter 4. where h is the liquid height above the
bottom,
p the density of the broth, and
3.2 Pressure g the acceleration of gravity,
0 by measuring the total weight of the
The absolute pressure is measured with re- reactor by load cells. The accuracy of
spect to zero pressure. Gauge pressure is meas- the volume measurement is +0.2% for
ured with respect to that of the atmosphere. large reactors and + 1% for laboratory
The SI unit of the pressure is Newton per reactors.
square meter (N/m2) called Pascal (Pa).
(1 bar=0.1 M P a = 10' N/m2; 1 mbar = 100 The measurement of the volume of an aer-
P a = 100 N/m2.) Bar and millibar deviate with ated broth is accomplished with a level con-
Instruments for Determination of Physical System Properties 11
troller. The common liquid level meters are Since cultivation broths have adequate elec-
based on the variation of the capacitance C of trical conductivity, the liquid level can also be
the sensor with the composition of the dielec- measured by inexpensive electrical conductivi-
tricum. For plate condensers, the capacitance ty probes. Their application is restricted to
is given by aqueous broths. In the presence of a second
(organic) liquid phase, their application cannot
A be recommended.
C= E O E , - For other level control instruments, see,
d
e.g., OEDEKOVEN (1980), ELFERS(1964), AN-
where A is the area of the plates, DREW and RHEA(1970).
d the distance between the plates,
E o the absolute dielectric constant of
vacuum, and 3.4 Liquid Throughput
E, the relative dielectric constant of the
aerated broth between the plates. Gas and liquid flow rates are important con-
trol variables for biotechnological processes;
The value of E, of the broth and of the air dif- they must be known for reactor operation and
fer by a factor of about 80. In the case of non- for component balancing. In this chapter, only
aerated broth the capacitance of the condens- instruments for liquid throughput measure-
er, C, is given by ment are taken into account. Instruments for
gas throughput measurements are considered
in Chapter 2. Special techniques for measure-
ments of local liquid velocities are treated in
where Co is the capacity of the condenser Chapter 4.
with air, Of the large number of available instru-
AC the capacity difference due to ments (SCHR~DER, 1980; ANDREW et al.,
broth per unit height, and 1979; ERICSON,1979) only three types are im-
h the height of the liquid in the ca- portant in biotechnological practice:
pacitor.
- floating body flowmeters,
The capacity of the condenser with aerated - differential pressure flowmeters, and
broth is given by - magnetic-inductive flowmeters.
+
C = Co A C h (1 - E ) (17) The floating body flowmeter or rotameter
consists of a conical tube and a floating body
where E is the gas holdup in the aerated broth. with the upper diameter D,, mass M,, and den-
Analogous relationships hold true for cylindri- sity ps (Fig. 1). In the upstreaming fluid, the
cal condensers (OEDEKOVEN, 1980). lifting force, which is produced by the differ-
The accuracy of the level control amounts ential pressure across the slot between the tube
*
to 2-4% depending on the uniformity of the
liquid level. In the case of large reactors, the
wall and the floating body, is balanced by the
weight of the floating body minus its buoyan-
level variation can be extremely large. There- cy. The position of the float is a function of
fore, only the level of the broth can be meas- the flow rate and the density of the fluid, p.
ured in the reactor, not that of the aerated The volumetric throughput qv is given by
broth, which is measured outside of the reac-
tor, e.g., in a non-aerated section. Also in the
case of foam formation, the measurement of
the aerated broth level by capacitance instru-
ments becomes difficult. Under these condi- The flow coefficient a is a function of the Rey-
tions floating bodies can be used as level con- nolds number and the diameter ratio Dk/D,,
trollers. where Dk is the diameter of the tube at the up-
per edge of the floating body.
12 I Common Instruments for Process Analy>pis and Control
3.5 Power Input In the power inputs, Eqs. (24) and (25), the en-
ergy losses due to mechanical energy (e.g., due
In an agitated reactor, the power input, P, to gas compression) are not considered.
can be calculated by Eq. (23) by measuring the
torque on the shaft, MN, and the speed of ro-
tation, N 3.6 Liquid Viscosity
The viscosity of the broth influences the op-
eration of bioreactors considerably. At a high
The torque is measured by torsion dynamome- viscosity, a high specific power input is neces-
ters or strain gauges and the impeller speed by sary to increase the intensity of transfer proc-
an electronic tachometer. In large-scale reac- esses: oxygen transfer (into the broth of aero-
tors, the consumed electrical energy, as meas- bic microorganisms) and mixing. Since at high
ured by the wattmeter, yields useful data on specific power input the energy dissipation rate
power input, if the mechanical losses in gear, in the reactor is high, improvement of heat
seals, etc., are taken into account. transfer is also important to keep the tempera-
In small laboratory reactors, the mechanical ture constant.
losses are considerable in comparison with the High viscosity can be caused by high sub-
power input into the broth. Therefore, power strate concentration (e.g., starch), high prod-
input measurements are inaccurate and are not uct concentration (e.g., xanthan), high cell
recommended. concentration (e.g., penicillin), high solid con-
In bubble columns, P can be calculated by tent (e.g., peanut flour), or by their combina-
tion. The most general description of the rheo-
logical properties of fluids is given by the rela-
tionship between the velocity gradient dv/dx
and the stress, T, the so-called flow equation:
dv
f
- = (7)
dx
where MG is the gas mass flow,
R the gas constant, as long as viscoelastic behavior is not present
T the absolute temperature, or very slight. This flow equation can be calcu-
pin the gas pressure at the column lated from the experimentally measured shear
inlet, diagrams (shear rate versus shearing stress). It
pout the gas pressure at the column should be noted, however, that such a calcula-
outlet, tion is not always possible. In contrast to the
H the height of the bubbling shear diagram, the flow equation is indepen-
layer, dent of the experimental conditions (e.g., the
w ~ the, linear
~ ~gas velocity at the in- type of viscosimeter) used for the determina-
let, tion of the viscosity.
w ~ the linear
, ~ gas ~velocity
~ at the There are many methods available to esti-
outlet, and mate the rheological behavior of fluids, but
g acceleration of gravity. there are only a few that furnish true fluidity
values. These include the capillary, the falling
However, since the second and third terms to- sphere, the Couette, the Searle, and the tor-
gether make up only 0.2% of the overall pow- sional pendulum methods. Until now, the eval-
er input, the power input due to the gas expan- uation of the flow equation from the shear
sion dominates: diagram has only been possible for the capil-
lary, Couette, and Searle methods (MUSCHEL-
Pin KNAUTZ and HECKENBACH, 1980).
P = M G R Tln - The capillary viscosimeter cannot be em-
Pout
ployed for cultivation broths because of ad-
14 1 Common Instruments for Process Analysis and Control
verse wall effects in the capillary. As for the where Riand R, are the radii of the inner and
falling sphere and torsional pendulum viscosi- outer cylinders,
meters, the flow equation cannot be calculated ti and t, are the shear stresses at the in-
from the shear diagram (only partial solutions ner and outer cylinders.
are known).
The Couette and Searle viscosimeters can The relationship between the angular velocity
only be used if the following conditions are of the rotating cylinder SZ and t is experimen-
fulfilled: the annular slit between inner and tally determined to obtain the shear diagram.
outer cylinders must be large enough to reduce The relationship dv/dx=f(z) (flow equation)
the wall effects, and measurements must be can be calculated from Eq. (30). For this eval-
made using different cylinder lengths to elimi- uation, see MUSCHELKNAUTZ and HECKEN-
nate the end effects. In a Searle viscosimeter, BACH (1980) and DINSDALE and MOORE
the speed of rotation is limited by the occur- (1962).
rence of Taylor instabilities. For Newtonian fluids, the following rela-
By measuring the torque MN on the shaft of tionship is valid:
different types of stirrers at differing stirrer
speeds, N is suited for the evaluation of the T= -9-
dv
power input but not for the viscosity. These dx
techniques, which are commonly used accord-
ing to the literature, are not suitable for the where q is the dynamical viscosity.
evaluation of the shear diagram and the abso- In practice, relative viscosities are frequent-
lute viscosity. ly determined. The shear stress is measured for
Only the coaxial cylinder viscosimeters, different shear rates with fluids of known
Couette with rotation outer cylinder and (oils) and unknown (broth) viscosities, and the
Searle with rotation inner cylinder, are consid- relative viscosity of the broth can be calculated
ered here, since they are the most popular from the ratio of their shear stresses at the
ones. same shear velocity, if the broth has Newton-
The velocity gradient at distance r is ian behavior.
On-line determination of the broth viscosity
dv
-=--
do is sometimes useful for controlling a process.
dx dr The on-line techniques only yield relative vis-
cosities. The viscosity of the Aspergillus niger
while the shear stress is broth was measured on-line by means of a
tube viscosimeter by BLAKEBROUGHet al.
(1978). PERLEYet al. (1979) used an on-line
capillary technique for the measurement of the
viscosity of the Hansenula polymorpha broth.
where w is the angular speed, LANGERand WERNER(1981) and NEUHAUS
Mi the torque exerted on the inner cy- et al. (1983) developed an on-line slot-type vis-
linder, and cosimeter and measured the viscosity of the
L the length of the inner cylinder. Penicillium chrysogenum broth. KEMBLOWSKI
et al. (1985) used an on-line impeller type vis-
From Eqs. (27) and (28) it follows that cosimeter to determine the viscosity of the
A ureobasidium pullulans broth.
1 dt
d o = - f ( t )7
2
3.7 Foaminess
Integration of Eq. (29) with s2 = R?/Ri = tilta
gives: Several cultivation broth components, espe-
cially a combination of different surfactants
with proteins, may cause stable foams in aer-
ated bioreactors. Foam control is necessary to
Instruments for Determination of Chemical System Properties 15
avoid the loss of broth, the clogging of the gas At [H '1 = [OH-], the hydrogen ion con-
analyzers, and infections caused by foam centration is [ H + ]= lo-', therefore, at the
carry-out. neutral point pH =pOH = 7.
Foam can be suppressed by antifoam agents The pH can be measured with a galvanic cell
(BEROVICand CIMERMAN,1979; SIE and (chain). The potential E of the cell is given by
SCHOGERL,1983; SCHUGERL,1986; PRINS the Nernst equation:
and VAN'T RIET, 1987; VIESTURS et al., 1982)
or destroyed with mechanical foam breakers RT
(VIESTURSet al., 1982). Foam can be detected E=Eo + 2.3 -log [H '1 (33)
F
by an electrical conductivity probe, capaci-
tance probe, heat conductivity probe, or light where Eo is the standard potential and
scattering probe (HALLet al., 1973; VIESTURS F the Faraday constant.
et al., 1982). Antifoam and mechanical foam
breakers are frequently combined, if the foam In this definition the thermodynamic activities
is very stable. of the ions were replaced by their concentra-
The presence of an antifoam agent in the tions since the activities cannot be measured.
broth may influence cell growth and product The absolute potential cannot be measured
formation as well as downstream processing. either, only the potential difference U between
Mechanical foam breaking may exert stress the indicator electrode and a reference elec-
and selection pressure on the cells. trode.
Silver-silver chloride electrodes are used in
the galvanic chain for sterilizable electrodes.
Fig. 2 shows the schematic assembly of a pH
electrode (INGOLDI). In this figure El is the
4 Instruments for potential on the outer surface of the glass
Determination of Chemical membrane, which depends on the pH value of
the sample solution. E2 is the asymmetry (bias)
System Properties potential, i.e., the potential of the glass mem-
brane with the same solutions on both sides.
E3 is the potential on the inner surface of the
4.1 pH Value glass membrane, which is a function of the pH
value of the internal buffer solution. E4 is the
The dissociation constant K, of the purest potential of the internal Ag/AgCl lead-out
water is very low (10-'5.74 at 25 "C). The con- electrode, dependent on the KCl concentration
centration of water can be considered as con- in the internal buffer solution. E5 is the poten-
stant because of the low K, value. Thus, only tial of the reference AgCl/Ag electrode, which
the ion product K, is taken into account:
depends on the KC1 concentration in the refer- For more information on pH electrodes, their
ence buffer solution, E6 is the diaphragm or use, storage, aging, etc., see INGOLDI, PE-
diffusion potential. TERSEN (1980), MELZNER and JAENICKE
Since El is the potential which we want to (1980), and MCCULLOUGHand ANDREW
measure, the individual potentials E2-E6 (1979).
should be kept constant. These are included in
the standard potential U",which has to be de-
termined by calibration. 4.2 Dissolved Oxygen Partial
In modern pH electrodes, U" varies only in
a narrow range (e.g., Type U 402-K7 elec- Pressure, po,
trodes of Ingold AG have a potential of
- 10.4k3.8 mV at pH 7.02 and 20 "C). The dissolved oxygen concentration is also
The potential difference between the indica- measured by electrochemical methods. Two
tor and reference electrodes U is also given by types of electrodes are in use:
the Nernst equation:
- polarographic electrodes
U = @+ U , log [H '1 (34) - galvanic electrodes.
cathodic reaction:
O2+ 2 H 2 0 + 2e- H202+ 2 0 H -
+
H202+2e- +20H-
anodic reaction:
Ag + C1- + AgCl + e -
overall reaction:
4Ag + 0 2 + 2H20 + 4C1- + 4AgCl+ 4 0 H -
overall reaction
02+2Pb+2H20+2Pb(OH)2
mentation Laboratory Inc., Lexington, Mass., The dissolved oxygen concentration [O,] is
USA, and also produced by Dr. W. Ingold calculated by the relationship:
AG, Urdorf, Switzerland (INGOLD11). This
type of electrode has a fairly long response 1 0 2 1 =Po2/H (38)
time (45 to 90 s to attain 98% of the final sig-
nal). where H is the Henry coefficient
During steam sterilization, the membrane
thickness and shape change irreversibly. An 22.4(760)
H=
improved construction of BAUERMEISTER 1000 a
(1981) enables the electrodes to endure many
(ca. 20) sterilizations without any change in the and CY is the Bunsen coefficient.
membranes. Bunsen coefficients CY of oxygen for some
The temperature of the calibrations and simple aqueous solutions and a few cultivation
measurements must be controlled closely broths have been given by SCHUMPE(1985).
(k 0.1 "C) because of the temperature sensitivi- For more details, see MELZNER and JAENICKE
ty of the signal (temperature coefficient 3%/ (1980), INGOLD11, LEE and TSAO (1979),
"C). Since the electrode measures the partial FRITZE(1980), BUEHLERand INGOLD(1976),
pressure of oxygen, the signal is independent and SCHINDLER and SCHINDLER (1983).
of the 02-solubility in the broth. The calibra-
tion should be performed in the reactor under
the same fluid dynamic conditions (stirrer 4.3 Redox Potential, Eh
speed) as those that prevail during cultivation
to avoid errors due to differences in diffusion The definition of the redox potential is given
resistance at the surface of the membrane. by Eq. (9). To determine E, the potential be-
The calibration is carried out with nitrogen- tween the redox electrode and a standard refer-
and air-saturated broth by setting these values ence electrode is measured. The universal ref-
at 0 and 100%. The partial pressure of oxygen erence reaction is the oxidation of hydrogen:
is expressed as follows:
H2-+2H++2e-
po2= [PB-p(HzO)] x 0.2095 (37)
Pco, I H f l [HCO;I
K=
[CO2l
The presence of dissolved C 0 2 in the broth
influences cell growth and product formation according to Henry's law
(Ho et al., 1987). Therefore, thepco2 can be
an important variable. The pco, can be meas- [C02l =Pco,H (42)
ured in-line using thepcoz meter of Dr. W. In-
gold AG (INGOLDIV). The instrument con- where H is the Henry coefficient.
sists of a pH meter and a hydrogen carbonate Since the hydrogen carbonate concentration
solution, which is separated from the broth by in the electrolyte is high, it can be assumed to
a gas-permeable membrane. Fig. 5 illustrates be constant. Thus, Eq. (41) can be simplified
the main features of the electrode. to
The dissolved Cot diffuses through the
membrane into the hydrogen carbonate solu- [ H + ]=pco2const. (43)
tion. The equilibrium of the reaction
The potential of the inner pH electrode is a
CO,+H,O + HCO, + H + function of [H'I:
20 I Common Instruments for Process Analysis and Control
F
product are relevant. Accuracy of the meas-
where 2.3 R T/F=59.16 mV at 25 "C, ured data is expressed as the difference be-
E is the measured potential, and tween the observed value of the variable and
Eo is the standard potential. its true value, which is usually determined by
calibration.
From relationships (43) and (44) The precision of the data relates to the
probability that repeated measurements of the
RT same system will produce the same values. The
E = Eo + 2.3 -10g(p,02) (45)
F distribution of the values around their mean is
usually characterized by the variance and/or
is obtained. standard deviation, or, e.g., the 95% confi-
The response time is fairly long (one to sev- dence interval.
eral minutes) and is influenced by the thickness The most important property of a sensor is
of the membrane and by the electrolyte solu- its reliability, which is made up of factors such
tion as well as by the response time of the pH as failure rate, failure mode, ease of preventive
electrode. maintenance, ease of breakdown maintenance,
The measuring range of the electrode is 1 to physical robustness, and its credibility in the
1000 mbar COz. The deviation is f 2 % , if the mind of process operators (FLY", 1982). The
electrode is calibrated with gas mixtures. If the latter plays a role only if the data are used by
inner pH electrode is calibrated by buffer solu- the operator and not by a n automated sys-
tions, the deviation is f 10%. To avoid errors tem.
due to the complex temperature dependence of Based on information from three chemical
the reading, the calibration should be carried works, LEES (1976) published data on the re-
out at broth temperature. liability of the instruments important in the
The electrode is sterilizable. The steriliza- fermentation industry (Tab. 1).
tion is performed after the reduction of the One can observe that at the time of investi-
pressure of the p H electrode on the stainless- gation (1970-1975) the pH meter and the 0,
steel reinforced plastic membrane. The p H and COz analyzers were the least reliable in-
electrode is calibrated by buffer solutions after struments. During the last ten years, the relia-
sterilization. Then the electrode is filled with bility of these instruments has been improved
the electrolyte and put into the measuring posi- considerably, provided an accurate flowmeter
tion. Fig. 5 shows the electrode during calibra- is used for the 0, and CO, instruments.
tion and measurement. For more information, FLY" (1982) gave detailed results on the
see INGOLDIV. performance of the instruments in a 1 m 3pilot
plant bioreactor (Tab. 2).
One can see that the po2 measurement had
the lowest accuracy, the po2 and air flow con-
5 Performance trol the lowest precision, and the volume meas-
urement the lowest resolution. In the mean-
of Instruments time, the air-flow control should have attained
a much higher accuracy and precision, pro-
for Process Control vided the right instrument is used (e.g., mass
flowmeter). In recent years, the accuracy of
the po2 measurement has not been improved
According to FLY" (1982) the relevance, markedly, it can, however, be achieved by fre-
accuracy, and precision of the measured data quent calibration. The accuracy and precision
and the reliability, accuracy, precision, resolu- of the po2 control is much better if one uses
tion, specificity, response, sensitivity, availa- three sensors and parameter-adaptive control.
bility, and costs of the sensors/instruments are
important for their use in process control. All
Performance of Instruments for Process Control 21
Temperature
Measurement kO.1 w 0.02 "C
Control 0.02 0.08
PH
Measurement f O . l 4H 0.001
Control 0.1 0.01
PO,
Measurement 6.0 R 0.05 mbar
Control 0.52 9.95
Pressure
Measurement 0.004 psig
Control 0.41 R 0.07
Volume
Measurement 0.14 L
Control 0.4 R 3.4
Air Flow
Measurement 0.02 L/m
Control -7.0 R 23.0
Exit COz
Measurement kO.1 R 0.0005%
Exit 0,
Measurement kO.l R 0.0005%
Notes: W, weekly; H, hourly; R, per run, which relates to the frequency with which the measuring instru-
ments are recalibrated
22 1 Common Instruments for Process Analysis and Control
6 Future Developments a b c d
technological Processes, pp. 5-12. Helsinki: KJAERGAARD, L., JOERGENSEN, B. B. (1979), Re-
IFAC. dox potential as a state variable in fermentation
FORAGE,D. E., HARRISON,D. E. F., PITT, D. E. systems, Biotechnol. Bioeng. Symp. 9, 85-94.
(1985), Effect of environment on microbial activ- LANGER,G., WERNER,U. (1981), Measurements of
ity, in: Comprehensive Biotechnology (BULL,A. viscosity of suspensions in different viscometer
T., DALTON,H., Eds.), Vol. 1, pp. 251-280. flows and stirring systems, Ger. Chem. Eng. 4,
Oxford-Toronto-Sydney-Frankfurt: Pergamon 226-241.
Press. LEE, Y. H., TSAO, G. T. (1979), Dissolved oxygen
FRITZE,U. (1980a), Die Technische Messung der electrodes, Adv. Biochem. Eng. 13, 35-86.
Redoxspannung, in: Messen, Steuern und Regeln LEELASART, B., BONALY,R. (1988), Effect of con-
in der Chemischen Technik (HENGSTENBERG, J., trol of pH on the excretion of acid phosphatase
STURM,B., WINKLER,O., Eds.), 3rd Ed., Vol. by Rhodotorula glutinis, Appl. Microbiol. Bio-
2, pp. 367-385. Berlin-Heidelberg-New York: technol. 29, 579-585.
Springer Verlag. LEES, F. P . (1976), The reliability of instrumenta-
FRITZE,U. (1980b), Weitere potentiometrische und tion, Chem. Ind. (9,195-205.
amperometrische ProzeRanalysenmethoden, in: LEHN, J. M. (1988), Supramolekulare Chemie -
Messen, Steuern und Regeln in der Chemischen Molekiile, Ubermolekiile und molekulare Funk-
Technik (HENGSTENBERG,J., STURM, B., tionseinheiten, Angew. Chem. 100, 91-1 16.
WINKLER, O., Eds.), 3rd Ed., Vol. 2, pp. 383-402. LINEK,V., VACEK,V. (1986), Recommended proce-
Berlin-Heidelberg-New York: Springer Verlag. dure for the measurement of the volumetric mass
HALL,M. J., DICKINSON, S. D., PRITCHARD,R., transfer coefficient kLa in aerated agitated ves-
EVENS,J. I. (1973), Foams and foam control in sels. Report of a Working Party on Mixing of the
fermentation processes, Prog. Ind. Microbiol. European Federation of Chemical Engineering,
12, 170-234. Prague.
HIRTE, W. (1980), Druckmessung, in: Messen, LINEK,V., VACEK,V., BENES,P. (1987), A critical
Steuern und Regeln in der Chemischen Technik review and experimental verification of the cor-
(HENGSTENBERG, J., STURM,B., WINKLER,O., rect use of the dynamic method for the determi-
Eds.), 3rd Ed., Vol. 1, pp. 101-190. Berlin-Hei- nation of oxygen transfer in aerated agitated ves-
delberg-New York: Springer Verlag. sels to water, electrolyte solutions and viscous
Ho, C. H . S., SMITH, M. D., SHANAHAN,J. F. liquids, Chem. Eng. J. 34, 11-34.
(1987), Carbon dioxide transfer in biochemical MCCULLOUGH,R. L., ANDREW,W. G. (1979),
reactors, Adv. Biochem. Eng. Biotechnol. 35, Analytical instruments, in: Applied Instrumenta-
83- 126. tion in the Process Industries (ANDREW,W. G.,
HOSPODKA,J. (1966), Oxygen-absorption rate-con- WILLIAMS,B., Eds.), 2nd Ed., Vol. 1, p. 209.
trolled feeding of substrate into aerobic microbial Houston: Gulf Publishing Co.
cultures, Biotechnol. Bioeng. 8, 117-134. MELZNER,W., JAENICKE, D. (1980), Prozefianaly-
INGOLDI, Brochure I: p H Electrodes. Dr. W. In- tik, in: Ullmanns Encyklopadie der technischen
gold AG, Urdorf, Switzerland. Chemie, 4th Ed., Vol. 5 , pp. 891-944. Wein-
INGOLD11, Brochure 11: 0, Measurement Systems; heim-Deerfield Beach, Florida-Basel: Verlag
p H and 0, Measurement in Fermentors. Dr. W. Chemie.
Ingold AG, Urdorf, Switzerland. MEMMERT,K., WANDREY,C. (1987), Concept of
INGOLD111, Brochure 111: Redox Measurement. Dr. redox-regulation for continuous production of
W. Ingold AG, Urdorf, Switzerland. Bacillus exoenzymes, Proc. 4th Eur. Congr. Bio-
INGOLDIV, Brochure IV: Sterilizable CO, Probe. t e c h n o l o g y ( N ~ ~ s s ~M.,VANDERMEER,
~,O. R. R.,
Dr. W. Ingold AG, Urdorf, Switzerland. LUYBEN,K. CH. A. M., Eds.), Vol. 3, pp. 153-
JOERGENSEN, H . (1941), Studies on the Nature o j 154. Amsterdam: Elsevier Science Publisher B. V.
Bromate effect. Thesis, Copenhagen, Munks- Mou, D. G., COONEY,C. L. (1983), Growth moni-
gaard. toring and control in complex medium: A case
KEMBLOWSKI, Z., KRISTIANSEN, B., ALAYI, 0. study employing fed-batch penicillin fermenta-
(1985), On-line rheometer for fermentation liq- tion and computer-aided on-line mass balancing,
uids, Biotechnol. Lett. 7, 803-808. Biotechnol. Bioeng. 25, 257-269.
KING, R. E., ARAGONA,J., CONSTANTINIDES, A. MUSCHELKNAUTZ, E., HECKENBACH,M. (1980),
(1974), Special optimal control of batch fermen- Rheologische BetriebsmeBverfahren, in: Messen,
tor, Int. J. Control 20, 869-879. Steuern und Regeln in der Chemischen Technik
KJAERGAARD,L. (1977), The redox potential: Its (HENGSTENBERG, J., STURM,B., WINKLER,O.,
use and control in biotechnology, Adv. Biochem. Eds.) 3rd Ed., Vol. 2, pp. 431-512. Berlin-Hei-
Eng. 7, 131-150. delberg-New York: Springer Verlag.
References 25
NEUHAUS,O., LANGER,G., WERNER,U. (1983), pression, and the effect of foam on growth,
On-line Betriebsviskosimeter fur Fermentations- Process Biochem. (August) 122-123.
prozesse, in: Biotechnologie-Projekt.forderung. SCHUMPE,A. (1985), Gas solubilities in biomedia,
Projekte 1978-1984, pp. 39-53. Bonn: BMFT in: Biotechnology (REHM, H.-J., REED, G.,
(Bundesministerium fur Forschung und Techno- Eds.) Vol. 2, pp. 159-170. Weinheim-Deerfield
logie) . Beach, Florida-Basel: VCH Verlagsgesellschaft.
OEDEKOVEN, G. (1980), Standmessung, in: Messen, SHIOYA,S. (1988), Measurement of biological reac-
Steuern und Regeln in der Chemischen Technik tion rates using advanced pH control, in: Fourth
(HENGSTENBERG, J., STURM,B., WINKLER,O., Int. Congr. Computer Applications in Fermenta-
Eds.), 3rd Ed., Vol. 1, pp. 589-640. Berlin-Hei- tion Technology, IFAC, Sept. 25-29, University
delberg-New York: Springer Verlag. of Cambridge, UK (FISH, N. M., Fox, R. I.,
PAN, C. H., HEPLER,L., ERLANDER, A. P . (1972), Eds.). Chichester: Society of Chemical Industry
Control of pH and carbohydrate addition in the and Ellis Horwood Publ.
penicillin fermentation, Dev. Ind. Microbiol. 13, SIE, T.-L., SCHOGERL,K. (1983), Foam behavior of
103- 112. biological media. IX. Efficiency of antifoam
PERINGER,P., B L A C H ~ R H. E , T. (1979), Modeling agents with regard to their foam suppression ef-
and optimal control of bakers’ yeast production fect on BSA solutions, Eur. J. Appl. Microbiol.
in repeated fed-batch culture, Biotechnol. Biotechnol. 17, 221-226.
Bioeng. Symp. 9, 205-213. SPITZER, D. W. (1976), Maximization of steady-
PERLEY,C. R., SWARTZ,J. R., COONEY,C. L. state bacterial production in a chemostat with pH
(1979), Measurement of cell mass concentration and substrate control, Biotechnol, Bioeng. 18,
with a continuous viscosimeter, Biotechnol, 167- 178,
Bioeng. 21, 519-523. SQUIRES,R. W. (1972), Regulation of the penicillin
PETERSEN,0. (1980), pH-Messung in: Messen, fermentation by means of a submerged oxygen-
Steuern und Regeln in der Chemischen Technik sensitiveelectrode, Dev. Znd. Microbiol. 13,128-135.
(HENGSTENBERG, J., STURM,B., WINKLER,O., STEPHANOPOULOS, G., SAN, K.-Y. (1982), On-line
Eds.), 3rd Ed., Vol. 2, pp. 305-333. Berlin-Hei- estimation of the state of biochemical reactors,
delberg-New York: Springer Verlag. in: Chemical Reaction Engineering, Boston 1982
PRINS, A., VAN? RIET, K. (1987), Proteins and (WEI, J., GEORGAKIS,C., Eds.). ACS Symp.
surface effects in fermentation: Foam, antifoam Ser. 196, 155-164.
and mass transfer, TIBTECH, November 5, 296- TAKAMATSU,T., SHIOYA, S., YOKOYAMA,K.,
301. IHARA,D. (1981), State estimation and control
RAI, V. R., CONSTANTINIDES, A. (1973), Mathe- of a biochemical reaction process, in: Biochemi-
matical modeling and optimization of the glu- cal Engineering ZI (CONSTANTINIDES,A . ,
conic acid fermentation, AIChE Symp. Ser. 132 VIETH, W. R., VENKATASUBRAMANIAN,K.,
(69), 114-122. Eds.). Ann. N.Y. Acad. Sci. 369, 147-158.
ROSEN,W., SCHOGERL,K. (1984), Microprocessor VIESTURS,U. E., KRISTAPSONS, M. Z., LEVITANS,
controlled fed-batch cultivation of Chaetomium E. S. (1982), Foam in microbiological processes,
cellulolyticum on cellulose, Third Eur. Congr. Adv. Biochem. Eng. 21, 169-224.
Biotechnology, Munchen, 10-14 Sept., Vol. 11, VOGTLE,F., Ed. (1975), Host guest complex chem-
pp. 617-619. Weinheim: Verlag Chemie. istry I, Topics in Current Chemistry 101. Berlin-
SAN,K.-Y., STEPHANOPOULOS, G. (1984), Studies Heidelberg-New York: Springer Verlag.
on on-line bioreactor identification. IV. Utiliza- VOGTLE,F., Ed. (1981), Host guest complex chem-
tion of pH measurements for product estimation, istry 11, Topics in Current Chemistry 98. Berlin-
Biotechnol, Bioeng. 26, 1209-1218. Heidelberg-New York: Springer Verlag.
SCHINDLER,J. G., SCHINDLER,M. M. (1983), WANG, H. Y., COONEY,C. L., WANG, D. I. C.
Bioelektrochemische Membranelektroden, Ber- (1977), Computer aided baker’s yeast fermenta-
lin-New York: Walter de Gruyter. tions, Biotechnol. Bioeng. 19, 69-89.
SCHLEGEL,H. G. (1974), Allgemeine Mikrobiolo- WITTLER,R., BAUMGARTL, H., LUBBERS,D. W.,
gie, 3rd Ed. Stuttgart: Georg Thieme Verlag. SCHOGERL,K. (1986), Investigations of oxygen
S C H R ~ D E RA.
, (1980), DurchfluBmeBtechnik, in: transfer into Penicillium chrysogenum pellets by
Messen, Steuern und Regeln in der Chemischen microprobe measurements, Biotechnol. Bioeng.
Technik (HENGSTENBERG,J., STURM, B., 28, 1024-1036.
WINKLER,O., Eds.), 3rd Ed., Vol. 1, pp. 191- ZABRISKIE, D. W., HUMPHREY, A. E. (1978), Real-
364. Berlin-Heidelberg-New York: Springer Ver- time estimation of aerobic batch fermentation
lag. biomass concentration by component balancing,
SCHUGERL,K. (1986), Foam formation, foam sup- AIChE J. 24, 138-146.
2 Methods and Instruments
in Fermentation Gas Analysis
E LMAR HEI N Z LE
IRVING J. D UNN
Zurich, Switzerland
1 Introduction 30
2 Mass Balancing for Gas Analysis 31
2.1 Basic Gas Balance Equations 31
2.2 Inert Gas Balance to Calculate Flow Rates 33
2.3 Steady-State Gas Balance to Determine the Biological Reaction Rate 33
2.4 Determination of CPR with Accumulation of CO, in the Liquid Phase 34
2.5 Determination of KLa by Steady-State Gas Balancing with Well-Mixed Gas and
Liquid Phases 35
2.6 Determination of KLa by the Dynamic Method 35
2.7 Determination of Oxygen Uptake Rates by a Dynamic Method 36
2.8 Loop Reactors with External Aeration to Determine OUR 36
2.9 Methods to Measure Low Oxygen Uptake Rates 37
2.10 Oxygen Transfer in Large-Scale Bioreactors 38
3 Application of Gas Analysis Results to Elemental Balancing Methods 41
3.1 Systematics of Elemental Balancing 41
3.2 Elemental Balancing for Monitoring a Poly-P-Hydroxybutyric Acid (PHB) Producing
Culture 42
4 Error Analysis for Gas Balancing 44
4.1 Objectives of On-Line Gas Analysis and Requirements for Accuracy and Reliability 45
4.2 Definition of Measurement Requirements 45
4.3 Errors Caused by Simplification of Balancing 46
4.3.1 Simplifications Concerning Pressure, Temperature, Humidity, and Gas Flow
Rates 46
4.3.2 Errors Caused by Steady-State Assumption 47
4.4 Erroneous Estimation of Reaction Rates Caused by Measurement Errors 49
4.4.1 Errors in the Measurement of Gas Flow 49
4.4.2 Statistical Error Propagation 49
4.4.3 Errors in Oxygen Gas Analysis 50
4.4.4 Instantaneous Error Analysis for the Elemental Balancing Example PHB 51
4.4.5 Dynamic Error Analysis for Reaction Rates 54
5 Sample Pretreatment and Multiplexing 56
5.1 System without Removal of Condensable Volatiles 56
28 2 Methods and Instruments in Fermentation Gas Analysis
pG=NRT (5)
where N is moledtime and G is volume/time.
Thus, useful expressions are:
ni Pi
CGi= - =- (6)
V RT
and
V
Fig. 1. Gas transfer in a stirred tank reactor with
well mixed phases.
ni = cGi
v = __
~ ~
RT (7)
32 2 Methods and Instruments in Fermentation Gas Analysis
(9)
Also temperature correction relations which Correct application of the balance equation
allow correction to standard conditions at con- generally requires correcting for the inlet and
stant p and V , outlet conditions. This is most easily done by
calculating the moles of gas involved, in which
case the balance equation would be
(2)
(2) where CL,,iis in mol/L and H in L bar/mol.
Calculation of the molar volumes is most
easily done by applying the factor of 22.4 L,
which is the volume of one mole of gas at
or in terms of mole fractions at constant p l , standard conditions of 1 bar and 0°C. Thus,
(%)
(*) (14)
and
If the pressures and temperatures are different Oxygen Uptake and Heat
at inlet and outlet the gas balance in terms of
the mole fractions becomes, Production Rates
The concept of yield allows a quick estimate
of reactor heat loads. Thus,
ture are given in the literature (ROELS, 1983): rate. This application is very important in fer-
YQ,02= 385 to 490 kJ/mol 02. mentation technology, because it permits on-
In the following, the general equations will line monitoring of the rate of reaction.
be applied to special situations. Often specific activities are needed. These
are defined as:
C 0 2 production rate (CPR) is calculated anal- (CL,co2= Cf,co,) between gas and liquid
ogously to Eq. (20) phases provides an estimate of CL,co, accord-
ing to Eq. (3) and of CHCO;using Eq. (23)
giving
CPR=rcozvL=-NO
Y I . inert
+ CHCO;)
- A (CL,CO,
VL (3 1)
At
Since CL,co, and CHco; usually are not
measured on-line, estimates have to be derived
from gas analysis and a pH measurement.
The assumption of equilibrium conditions
Mass Balancing for Gas Analysis 35
The classical dynamic KLa method assumes The relative influence of the three proces-
that KLa and CLl are constant. Then the dif- ses depends on the values of the constants
ferential equation can be integrated analytical- rG= VG/G, l/KLa, and r E . Because of their
ly to yield: linear nature, the differential equations can be
solved analytically and it can be shown that
the area above the response curve is equal to a
(37) simple sum of time constants (DANGet al.,
1979):
Plotting the logarithmic function versus t
gives K,a as the slope. This method is accurate
at low rates of transfer, with K,a below
60 h-'.
The method of calculation is shown in Fig. 2.
As might be expected, very high KLa values
I cannot be measured accurately with relatively
1.0 slow electrodes. Thus, when - l / K L a etE, this
method is not accurate.
CE
A simple method involves neglecting the
transfer term in the gas balance and noting
-0 Time that the equations represent three lags in series
whose output is approximately the sum of the
Fig, 2. Calculation of K,a from dynamic experi- individual time constants. This is shown in
ments. C,,electrode signal; al,see Eq. (39). Fig. 3.
36 2 Methods and Instruments in Fermentation Gas Analysis
tem converts it effectively into a continuous responds to about 20% difference in satura-
stirred tank. The well-stirred tank condition of tion with a 1 minute liquid residence time.
homogeneity is fulfilled because there are only
small differences in concentration between the
reactor inlet and outlet. 2.9 Methods to Measure Low
Besides being flexible with respect to reactor
mode (batch or continuous tank), the recycle Oxygen Uptake Rates
reactor system provides the possibility of add-
ing reactants (such as 0,) and controlling the If oxygen uptake rates (OUR) are very low,
temperature and pH in the recycle line. there will be only a small difference between
The reactor-oxygenator recycle loop can be inlet and outlet oxygen concentrations in the
analyzed as a total system or broken down into gas phase. Examples are animal and plant cell
its individual components as shown in Fig. 4. culture, and sometimes wastewater treatment.
The oxygen balance taken over the total sys- The difference can be too small for error-free
tem can be simplified by neglecting the accu- gas balancing, and in such cases other methods
mulation terms and the liquid flow terms, must be found for OUR measurement. The
which will be small compared to the gas rates methods for low OUR can be separated into
and the consumption by reaction. Thus the those applicable to all reactors and those ap-
oxygen balance becomes plicable to special reactor types.
General methods include the following:
0 = GI CG,- Gz C,, + ro2 VR (41)
1. Carbon dioxide analysis to obtain OUR
The absorption tank can be described by the 2. Recycling gas phase to increase oxygen
oxygen balances for the liquid phase: utilization
3. Fed-batch oxygen control.
O=GICG1
-G2CG2-KLa(Cf-CL3)VT (43) 0 Circulation loop reactors with external
aeration (Sect. 2.8)
The oxygen balance of the liquid phase for the 0 Intermittent aeration and dynamic meas-
total system is urement of dissolved oxygen in tank
reactors (Sect. 2.7)
0 = KLa(C2- C,,)VT + ro, VR (44) Off-line batch oxygen uptake with elec-
trodes (Sect. 2.7).
where ro2 is the oxygen uptake rate of the reac-
tion. These equations, which assume ideally These methods will be discussed below.
mixed phases, are useful in designing the gas
absorber according to the required oxygen
transfer coefficient. Carbon Dioxide Balance
Balancing the oxygen around the reactor
gives The ratio of the molar evolution rate of C 0 2
to that of O2 may often be equal to unity
~CLJ + ‘0, VR
0 = LR( C L- (45) ( R Q = l ) or equal to a constant (RQ=con-
stant). Even if OUR is low, it may be possible
Thus it is seen that either Eq. (41) can be to measure the concentration of C 0 2 in the
used in a conventional gas balance form to cal- exit gas with sufficient accuracy, to allow ac-
culate OUR ( = ro2 V,) or the much more accu- curate balancing for rco2, which is proportion-
rate liquid phase balance Eq. (45) can be used. al to ro,. This method is not applicable when
In the range of 2 mg O2 per min or 3.75 mmol/h COz is used to control pH, as in cell culture.
reasonable accuracy can be expected. This cor- The solubility of C 0 2 is much higher than that
38 2 Methods and Instruments in Fermentation Gas Analysis
of 02,and additionally C 0 2 may be absorbed the local oxygen transfer rates will result in lo-
chemically as HCO; . Therefore, this method cal variations in the steady-state dissolved
is only useful at low pH ( c 5 . 5 ) . oxygen concentration. Consequently the dis-
solved oxygen concentration can vary appreci-
ably from the top to the bottom of a deep
Circulation of Gas Phase tank.
The flow conditions of the gas phase in
Recycling the gas phase in low rate systems large-scale industrial fermentors fall between
with a gas circulation pump allows greater uti- the extremes of plug flow and well-mixed
lization of oxygen. In this case only small gas states. Information on the distribution of resi-
flow rates of air or oxygen would be fed to the dence time under operating conditions is nec-
circulating gas phase, and the same rate would essary to characterize the gas phase flow. Un-
be withdrawn as off-gas. By this means the fortunately, little experimentation has been re-
concentration of O2 in the exit gas will be ported on industrial-scale equipment (JURECIC
greatly increased, thus permitting conventional et al., 1984; OOSTERHUIS, 1984; GRIOTet al.,
analysis and balancing. Alternatively the gas 1988).
phase can be enclosed in a membrane system, Hydrostatic pressure gradients in tall fer-
such as silicon tubing. mentors will cause large differences in oxygen
solubility, Ct, from the liquid surface to the
tank bottom. For example, in a 10 m tall reac-
Fed-Batch Oxygen Control tor, the oxygen solubility for a given gas com-
position is twice as great at the bottom as at
If the gas phase is partially closed or circu- the top because the total pressure is doubled.
lated, a dissolved oxygen measurement and It is shown by Henry’s law, Eq. (3). This will
control system can be used to meter the re- double the maximum OTR ( = K L a C t ) .Some
quired amount of oxygen, which would give compensation may be caused by smaller bub-
the uptake rate directly. Here also a membrane bles, due to compression, at the bottom of the
system can be used. column. If the number of bubbles remain con-
stant (no coalescence), KLa should vary with
P-0.5.
2.10 Oxygen Transfer It is possible to estimate the local dissolved
oxygen concentration during normal fermenta-
in Large-Scale Bioreactors tion conditions by equating the oxygen uptake
rate to the oxygen transfer rate at any position
Large industrial fermenters can generally be in the vessel. Thus, this balance can be applied
expected to exhibit deviations from ideal mix- to an element of gas and liquid volume at any
ing conditions. Thus, the assumptions of com- depth in the reactor.
pletely mixed gas and liquid phases may not be Assuming that the biomass concentration
valid for large bioreactors, and some of the and factors which influence Qo,are uniform
previously developed equations will not be ap- within the vessel (no local substrate or oxygen
plicable for the measurement of oxygen trans- limitation), the oxygen uptake rate may be tak-
fer and uptake. Little experimental informa- en as a constant throughout the vessel. The
tion is available on concentration inhomogen- mass transfer coefficient, KLa, will vary with
eities or gradients within large bioreactors. turbulence level, bubble size, and gas holdup.
Generally, information on the distribution of Assuming it to be a constant, independent of
residence time is not available from which a position, the balance may be solved for CL,o,
physical and mathematical model could be es- in exactly the same way as for C02, Eq. (34).
tablished. Combination with Eq. (20) for the well-mixed
Dissolved oxygen must be considered a rap- case gives
idly changing quantity because of the low solu-
bility of oxygen, and it is, therefore, necessary CLO, = ~
ro,
+-PY2,0,
to consider the possibility that differences in (K,a)o, ffo,
Mass Balancing for Gas Analysis 39
letting Eq. (54) for the plug flow gas (P), the follow-
ing is obtained:
L, (KLa)*M
(KLcI)LM.
will be only 12% lower than
The use of A C A M is a striking im-
3 Application of Gas
provement over A C W M . As can be calculated Analysis Results
from Fig. 5, the errors in KLa at y 2 = 0.15 and
0.10 are 25% for CL=2 mg/L. If the exit to Elemental Balancing
oxygen mole fraction is below y 2= 0.18 the er-
rors using A C W M will be less than 10%. As Methods
at steady state
42 2 Methods and Instruments in Fermentation Gas Analysis
where E is the elemental composition matrix sum of all sugars (saccharose, glucose, and
fructose), is defined by
C H O N
' Biomass
Substrate
Product The production of residual biomass
E= N-Source (67) (CHB,0B2NB,)equals the total biomass minus
0 2
COZ
PHB and is defined by
HzO
CHzO + CYI 0 2 + C Y ~H20 + a3 NH3 +
Three quantities are independent. Assuming 46 CHB10B2NB3 (69)
(=ro,), 45 ( = r c o 2 ) ,and 44 (nitrogen source)
are known, equations for (biomass), 4z From Eq. (69) it can be seen that a3=p3.
(substrate), 43(product), and 4, (water) can be Oxidation of sugar for ATP-production is
obtained by methods of linear algebra, as de- given by
tailed by ROELS(1983). A number of examples
in the literature used this method:
t$.E=O (72)
3.2 Elemental Balancing E is the elemental composition matrix
for Monitoring a
Poly-P-Hydroxybutyric Acid (PHB) C H O N
Producing Culture 1 1.8 0.5 0.2 Biomass
1 2 1 0 Glucose
The method of elemental balancing is dis- 1 1.5 0.5 0 PHB
cussed in more details using as example the E = 0 3 0 1 Ammonia (73)
synthesis of poly-/3-hydroxybutyric acid (PHB) 0 0 2 0
0 2
by Alcaligenes latus (HEINZLEand DETTWI- 1 0 2 co2
0
LER, 1986; HEINZLEet al., 1990). This exam- 0 2 1 0
H20
ple is later also used to analyze error propaga-
tion (Sect. 4). The elemental balances are:
Stoichiometry
Application of Gas Analysis Results to Elemental Balancing Methods 43
On-Line Calculations
4.26R + 4 & +4.51#+-4&,=0 (79)
The concentration of dissolved CO, (CL,co2)
From Eq. (77)) was estimated using Eq. (34). Differentiation
with time allowed the calculation of the liquid
phase accumulation. Combining Eqs. (25) and
(31), 6co, (=CPR/V,)can be calculated:
YO,inert
YLc0,- yo,co,-
6 R + 6 S + 6 P f 6CO,=o (81) YI,inert
6c0, =No
Using Eq. (80) and combining Eqs. (79) and
(81) yields the following expressions which can
be used directly for the calculation of meta-
bolic fluxes based on titration and gas analy-
sis: At constant pH, ammonia consumption rate
(rNH=)equals the difference between addition
rate of alkali (IoH-)and the accumulation rate
of bicarbonate:
44 2 Methods and Instruments in Fermentation Gas Analysis
I
The molar fluxes of biomass, substrate, and
product can be calculated using Eqs. (82) to CNH2 = CNH4+,0 + j &H,C dt (91)
0
(84).
It can be expected that due to integration over
a period of time any errors in the fluxes would
Integration of Fluxes for the Batch accumulate. This is analyzed below (Sect. 4).
Concentrations Fig. 8 shows an example of on-line estima-
tion of Cs and C, compared with off-line re-
sults. Both concentration values continuously
Integration of the fluxes gives the concentra- drifted away from the actual values. This may
tions of R, S, P, and NHZ at any time t: have been caused by an offset in the oxygen
partial pressure measurement, whose effect
would be accumulated by the integration of
the fluxes. As shown, a simulated offset of
A Y ~ ,=~0.0012
, would partially explain the de-
viations observed in this experiment.
The influence of errors in gas analysis on
the estimation of gas reaction rates and fluxes
of biomass, substrates, and product is dis-
cussed in more detail in the following Sect. 4.
0
glL -
6- 4 Error Analysis
*-".-..
/' for Gas Balancing
In most cases the choice of a method for gas
analysis and equipment depends on the re-
quired accuracy and reliability. These are
usually difficult to define because of the com-
plexity of error propagation. Higher quality
instruments usually cost more. The require-
ments may differ tremendously from industrial
large-scale process monitoring to research lab-
oratory gas analysis.
10- Generally two types of errors have to be
CS considered:
5-
off-line 0
(1) Errors caused by simplification of gas
0
balancing (Sect. 4.3)
0' i ii is io k0jshio (2) Measurement errors (Sect. 4.4).
Time
Fig. 8. Elemental balancing in production of PHB Systematic balancing errors can easily be
by Alcaligenes latus. Comparison of on-line (full avoided by careful analysis of the system and
line) and off-line measurement (symbols) results and proper setting up of balance equations. Meas-
possible cause of error due to incorrect calibration
urement errors depend on a number of factors,
for the measurement of oxygen partial pressure at
the reactor exit ( Y ~ , ~ ,Cp,
) . product PHB concentra- some of which may directly influence the cost
tion; C,, substrate concentration (total sugar); of equipment.
pH = 7.3; G, = 9.12 vvm. Simulation with
~ ~ -0.0012
Y ~ ,=Yz.02,true . ~ (dashed
~ ~ line).
~
Error Analysis for Gas Balancing 45
4.1 Objectives of On-Line Gas (Sect. 3). In this case errors of measure-
ment may be amplified dramatically and
Analysis and Requirements yield completely misleading results.
for Accuracy and Reliability
Three different levels of complexity using 4.2 Definition of Measurement
the results of gas analysis may be identified:
Requirements
In industrial practice it is often of para-
mount importance to guarantee and Accuracy denotes the agreement with true
prove the exact repeatability of batches. values. In many cases absolute accuracy is not
If on-line gas analysis is solely used for very important (e.g., dissolved oxygen meas-
this purpose, good precision and reliabil- urement, traces of partial pressures of gases
ity are of primary importance. The qual- and volatiles). To obtain high accuracy, care-
itative and quantitative traces of charac- ful calibration procedures using good stand-
teristic signals (e.g., signals of oxygen ards and careful analysis of the process are im-
and C 0 2 partial pressure meters, dis- portant to minimize systematic errors (e.g., to-
solved oxygen electrode signal) are used tal pressure or temperature changes in a partial
to trace the performance of a particular pressure measurement apparatus - see Sect. 8).
batch. In these cases medium composi- Good precision is achieved if signals are not
tion and metabolic pathways are often noisy and if there is no drift. Signal noise may
very complex and only poorly under- be reduced using filters or statistical methods.
stood. It is usually not necessary to Drift usually can be identified and corrected
know the true values of, e.g., oxygen by frequent recalibration. Contrary to liquid
uptake rates. phase on-line sensors, which are immersed into
In such cases, calibration has only to the sterile region of the fermenter, gas analysis
assure good reproducibility. These meas- equipment, which is usually located outside
urements are usually characteristic for a the sterile region, can be easily recalibrated
particular process and process analysis without causing sterility problems. Constant
device and, therefore, cannot be applied measurement and process environments reduce
directly to other processes. drift.
More specific information may be col- Excellent stability over extended periods in
lected if “true” concentration and reac- a harsh process environment is required for
tion rate values (e.g., OTR, CPR) and process instruments. This requires the special
RQ (respiratory quotient) are deter- design of instruments (power supply, housing,
mined. This information is of more gen- redundancy of critical instrument parts, e.g.,
eral value and may be used to compare filaments in mass spectrometers, and robust-
research results of different research ness).
groups. Calibration of instruments has In most cases the sensitivity of analytical
to be done frequently using a good qual- apparatus is sufficient for gas analysis. Prob-
ity standard. The most critical measure- lems may occur when analyzing trace gas com-
ment concerning precision and accuracy ponents. Examples of this are H2 in an anaer-
in gas analysis is oxygen partial pres- obic reactor and less volatile compounds (e.g.,
sure, because measured differences are acetoin and higher alcohols).
small. This will be discussed in detail
later.
The third level of complexity is the use
of the results of gas analysis, usually
combined with other measurements, to
indirectly estimate other concentration
values and growth and production rates
using elemental balancing techniques
46 2 Methods and Instruments in Fermentation Gas Analysis
4.3 Errors Caused by Simplification changes. This is most likely to be the case in
production reactors. In such cases reliability of
of Balancing instrumentation is of primary importance. For
research purposes specific respiration measure-
Lists of possible and potentially useful sim- ments (Qo2, Qco,) are usually required, mak-
plifications in the balance equations are given ing a balance of the mass or molar rates neces-
in Tables 1 and 2. It is important to know the sary. A schematic drawing of the equipment
possible contribution of each simplification to for gas balancing is shown in Fig. 9.
the final error in measuring reaction rates.
For CO, the situation is more complex be-
cause of its higher solubility and its reaction to 4.3.1 Simplifications Concerning
HCO; . Estimation of ro, is also indirectly in- Pressure, Temperature, Humidity,
fluenced by errors in CO, estimation.
The analysis of outlet gas concentrations is and Gas Flow Rates
for most fermentations one of the most relia-
ble on-line methods for reactor monitoring. In Starting from general balances for well-
some cases an exact quantitative balance is not mixed gas and liquid phases Eqs. (1) and (2)
needed, and it is sufficient to note relative and assuming steady state conditions (dCi/
LO %
- Reactor
Suppose there is a case in which ro, # rco, or
RQZ 1. We then know that there will be a dif-
ference in molar flow, which could be meas-
Fig. 9. Schematics of measurements in gas balanc- ured by two flowmeters with suitable T and p
ing. corrections. Alternatively, information about
the inert nitrogen or a p o n gas content can be
used, Eq. (21). Using dry gas measurements
dt=O), the influence of p , T, humidity, and and making an inert gas balance will always
gas flow rate assumptions are discussed. As- lead to correct results as is shown by column G
suming steady state, the total mass balance for in Tab. 3.
dry gas streams in terms of molar flow rates N
(mol/s) can be written as I .
0 0 $ 0 " 0 r i 0 0
I
'2 0 2 0
08q '2 ~ 8
" $ 0
q
Error Analysis for Gas Balancing 49
rors. Generally speaking, they are significant if less prone to such errors, since drift usually af-
the changes in biological activity are large fects the intensity of the internal standard in
compared to the corresponding mass transfer the same way as that of the analyzed compo-
terms. A typical case is the end of a batch cul- nent. Examples are GC, MS, and analyzers
ture, where a sudden depletion of the limiting employing dual-channel reference gas.
substrate quickly decreases OUR and CPR. A The first two types of error may be identi-
similar case is diauxic growth, where one sub- fied by recalibration or by the increasing in-
strate is depleted and the culture has to adapt consistency of balancing results. The applica-
its metabolic activities to a new substrate. An- tion of estimators with dynamic filters (e.g.,
other situation is the addition of an inhibitory Kalman filter) including a measurement model
or toxic substance to the culture, as in waste or artificial intelligence systems may support
water treatment or during addition of a toxic the identification and eventually the compen-
precursor. sation of such errors.
The dynamic behavior of complex systems An example of systematic error is given in
can best be seen by simulation. Dynamic errors Fig. 8, where the on-line estimation resulted in
and their consequences for elemental balanc- negative concentrations for the product PHB.
ing results are discussed with the example of Since it was determined that the measurement
an Akaligenes latus fermentation producing of yo, was most sensitive to errors, a miscali-
PHB (see also Sect. 2). bration for yo, may be assumed and compen-
sated as illustrated in Fig. 8.
Statistical measurement noise leads to corre-
4.4 Erroneous Estimation spondingly noisy results. The noise may be
of Reaction Rates Caused amplified and cause results which cannot be
interpreted without filtering properly. In most
by Measurement Errors cases standard electrical filters (first order) will
be sufficient to provide meaningful results.
Three sources of errors in measurement may Such filters, however, introduce a delay which
be identified: can lead to erroneous results.
- calibration offset
- instrument drift 4.4.1 Errors in the Measurement
- statistical noise.
of Gas Flow
There are different ways to avoid or mini-
mize such errors. Calibration offset errors can One kind of error which may enter into the
be minimized by using accurate standards and determination of OUR or CPR by the gas bal-
by careful calibration procedure (pressure, ance method is caused by inaccuracies in the
temperature, humidity). For high accuracy, measurement of flow rate. If Eqs. (21) and
dynamic errors during calibration must be (26) are used, the error in OUR or CPR will be
carefully avoided. It is usually difficult to proportional to the measured flow error
prove whether such errors occur or not. In (AOUR 0:ANo; ACPR a ANo). For example, a
some cases inconsistent balancing results point 10% error in the measurement of both gas
to such errors. An error analysis for the esti- flows will cause an error of equal percentage in
mation of gas reaction rates has been pub- the OUR.
lished by NEUBERTand MINKEVICH (1984).
Instrument drift can be minimized by pur-
chasing high-quality instrumentation conform- 4.4.2 Statistical Error Propagation
ing to process equipment standards. Instru-
ments should not be exposed to environmental The situation is a little more complex if both
conditions for which they are not built (tem- Go and G, are measured and have their indi-
perature, humidity, vibrations, etc.). Measure- vidual errors or if other errors are involved. In
ment techniques using internal standards are this case statistical error propagation has to be
50 2 Methods and Instruments in Fermentation Gas Analysis
estimated according to methods described in must be exercised if accurate results are re-
standard statistics textbooks (e.g., KREYSZIG, quired.
1970). For a large number of measurements HEINZLEet al. (1984) made an analysis of
with independent individual measurement er- error propagation for statistical errors in esti-
rors having Gaussian distributions, the mean mating oxygen uptake rates. Their analysis ap-
error of the estimated variable (s) of a function plies only to situations in which inert internal
z = h ( x , y ) is standards are used (e.g., GC and MS). Small
differences between inlet and outlet oxygen
concentrations are the main reason for the am-
(94) plification of measurement errors. Assuming
s = l i ( - g q p
that inlet concentrations are well known (air)
Starting with Eqs. (1) and (2), assuming and that gas flow measurements are accurate
steady state conditions and Lo=L 1=0, and
also assuming sco,02=sc1,02=0, results in
S O U R = I C O , O : S G ~ + cl,O:sG: (96)
Having similar O2 concentrations at inlet
and outlet (Co,o,= Cl,o,) and identical errors where si is the relative error for OUR,
for the measurement of both gas flow rates d=Oll,o,/Yi,inert), C=cYo,o,/Yo,inert),and ~i is
(sGo=sG,), leads to the error the relative error in measuring d. Consequent-
ly s: will also be equal to the error in the ratio
SOUR = CO,sG $ (97) of the corresponding ion currents. In Tab. 4 the
relation between s i and s; is illustrated for
With the above simplifications the mean er- a series of relative differences in oxygen
ror for $UR is proportional to the mean error concentrations at inlet and outlet { ACA, =
in G. For unequal values of C,, and SG the 2 [(Yo,0 , -Y 1,o,)~cyo,o,+Y 1,0,)l 1*
mean error in OUR is given by Eq. (96). Systematic errors in oxygen balancing may
be treated as follows: Rearranging Eq. (21)
gives
4.4.3 Errors in Oxygen Gas -OUR - Yo,o, YLO,
Analysis (99)
NinY0,inert Y0,inert Y1,inert
Time
10-
- --
= x 4
I I
0 10 20 30 40 h 50
Time
Fig. 10. Simulation for OUR (+02) and CPR (&02) and relative errors. PHI02, $02
(mol h-I); PHIC02, &o, (mol h-I); RELPHO, RELPHC, relative errors for +02 and for
4c0, (-1.
Fig. 11. Comparison of true and measured values MPHIO, MPHIC, measured values containing er- b
with error in the determination of yo, in the exit rors (mol L-4 h-')
gas. True concentration values:
Left side: relative offset calibration error for yo, R, residual biomass; S, substrate; P, product PHB
(yooffs = -0.01 = - l a ) all (c-mol L - I )
Right side: relative measurement drift for yo, (yo- Values estimated from measurement containing er-
drif = -0.0002 h-' = -0.02% h-') ror:
PHI02, PHIC02, true model values of $o, and MR, residual biomass; MS, substrate; MP, product
qbo, (mol L - ' h-I) PHB all (c-mol L -').
Error Analysis for Gas Balancing 53
vodrif=- 0.0002 h - I
yooffs: - 0.01
I I I I
10 20 30 4 0 h
Time
54 2 Methods and Instruments in Fermentation Gas Analysis
In a similar way the relative error in the de- flow values were created by the model. Based
termination of q5c02 at constant pH, accurate on these data, sugar and ammonia consump-
measurement of gas concentrations, and gas tion as well as biomass and PHB production
flows is calculated from Eq. (87): were estimated using the measured gas partial
pressures, flow rates, and p H values following
the procedure described in Sect. 3.2. The error
model for the measurements was either that of
calibration offset or of measurement drift.
*
1
(?- 1.26.10-'
) = 0.036
state assumptions for the gas phase concentra-
tions and for dissolved oxygen. It is evident
that these assumptions cause only relative er-
If q502 and q5R are measured correctly, using rors of < 2%. Resulting errors vary during the
Eq. (83) gives a relative error in estimation of process according to the process dynamics.
4.4.5 Dynamic Error Analysis Fig. 11 shows true and measured values of
q502, q5co,,as well as true values (from simula-
for Reaction Rates tion) of residual biomass (R =total biomass-
PHB), substrate (S), and product PHB (P),
The instantaneous errors in the estimation and values estimated from measurements con-
of the fluxes as explained above (Sect. 4.4.4) taining only errors in the determination of yo,
will vary in a batch process with time. As dis- (yooffs and yodrif). It can be seen that, in the
cussed in Sect. 3.2 the estimation of the con- case of calibration offset, estimation of q502
centration values requires integration over also shows basically a negative offset, whereas
time, Eqs. (88) to (91), and this integration will estimation of q5c02 is scarcely affected. In the
lead to error accumulation. The best way to case of a measurement drift (right side), the
understand this is to consider a detailed dy-
namic error propagation analysis by simula-
tion. Fig. 12. Estimation errors caused by error in deter- b
This was again done using the PHB example mination of yo, in the exit gas.
with the kinetic model of HEINZLEand LAF- Left side: relative offset calibration error for yo,
FERTY (1980) and following the data of (-0.01 IYOOFFSI0.01)
Right side: relative measurement drift for yo,
HEINZLEand DETTWILER(1986). Well-mixed
(0.00041YODRIF10.0002 h-I)
gas and liquid phases were assumed MEPHIO, Absolute estimation error for @o,
( VGl= VL * 0.1). A slight modification was to (mol L-' h-I)
couple another well-mixed gas element REPHIO, Relative estimation error for @o, (-)
( VG2= VL 0.3) representing the fermenter FINERP, Estimation error for product PHB (c-
head space. Concentration-time relations, gas mol L - ')
reaction rates, acid production rates, and gas TFIN, time (h).
Error Analysis for Gas Balancing 55
56 2 Methods and Instruments in Fermentation Gas Analysis
-
Measuring point selection 5.3 Special Valve Manifolds
for Mass Spectrometers
Filter
Valve manifolds, as described before, are
not ideal for fast and very accurate analysis of
a number of gas streams, because it is very dif-
ficult to avoid dead zones. Therefore, a num-
ber of sample valves have recently been devel-
Zem Span oped especially for fermentation gas analysis.
Reference ' .f gas gas Balzers AG (MOLLER and RETTINGHAUS,
1988) has recently developed a gas inlet system
air Condensates
shown in Fig. 15. Special solenoid valves hav-
i t ing practically no dead zones are used. The
Fig. 14. Sampling and calibration system for analy- whole valve system is located in a box within
sis of O2using a paramagnetic analyzer (PM) and of which the temperature is kept usually constant
C 0 2 using an infrared analyzer (IR). at 110°C to avoid condensation and to speed
up adsorption equilibration.
It is very useful to keep the total pressure
the residence time of the gas in the lines. This within the MS constant by controlling the flow
can be speeded up by evacuating the lines if through the capillary (Cap). This can be done
necessary. by controlling the pressure at the capillary en-
Water, especially condensate, is detrimental trance using a piezo pressure sensor (p) (Fig.
to the operation of analyzers, and humidity in- 15).
fluences the calibration. It must be removed in The sampling stream valves are under per-
a drying step. This usually involves a cooler manent flow to guarantee instant sample avail-
with dewpoint control and condensate remov- ability. Connections to the capillary inlet are
al. usually closed. Switching between gas streams
Gas flow to the analyzers is achieved either is done as follows: close V1 and V2, open V3
with overpressure control on each fermenter to evacuate sampling system, close V3, open
line or with a gas membrane pump. Since the V2, open one single sample stream valve (SV)
flow rate through the analyzer will influence or calibration gas valve (CV), open V1 to al-
the calibration, care must be taken to hold this low gas to enter the MS. The pressure at V1 is
constant.
Air is usually used for calibration of the pa-
ramagnetic analyzer, and in some instruments
it is compared directly with the gas sample.
Carbon dioxide gas mixtures are used to cali-
brate the infrared analyzer. It is useful to auto-
mate this calibration procedure.
The data from the analyzers are usually re-
duced by the computer to rates of uptake and
production, expressed in mmol/L h. As ex-
gv2
plained earlier, this requires careful attention
to the measurement of flow rates and their
conversion to molar quantities. Plotting the re-
duced data as a function of time greatly aids in
process control and interpretation. The data
can be stored for later use. Feedback control Fig. 15. Control of MS vacuum by controlling inlet
of the process can be programmed, for exam- gas flow. V1, MS inlet valve; V2, V3, switching
ple, to control feeding rates based on oxygen valves; SV1 to SVm, sample stream valves; CVl to
uptake. CVn, calibration gas valves (see text).
58 2 Methods and Instruments in Fermentation Gas Analysis
of the liquid in a wet gasometer. An electric pass (&) and one through the sensor tube
signal is created by counting revolutions. ( M I ) .This tube is designed to have laminar
Measurements obviously are directly in- flow conditions. Two coils surround the sensor
fluenced by pressure and temperature. This capillary. The coils direct a constant amount
can be compensated by using the ideal gas law. of heat through the thin walls of the sensor
Gas composition does not influence the meas- tube into the gas. The coils sense changes in
urement except when gaseous components temperature through changes in their resist-
change the surface tension of a liquid gaso- ance. During operation, heat is transported
meter. With COz, accumulation in the liquid from the upstream coil to the downstream one,
gasometer fluid has to be taken into account.
Such devices are not very useful at fermenta-
tion reactor inlets because of pressure fluctua- Immersible MFM
tions.
6.2 Rotameters
A measurement body is floated in a vertical
conical tube by an upstream gas flow. The
drag coefficient is a function of the Reynolds
P
number and hence depends on the fluid viscos-
ity. Rotameters are available in a variety of Capillary tube MFM
materials and can be used for any gas and flow Tn----
rate. The measurement is influenced by pres-
sure, temperature, and less by gas composi-
tion. The location of the floating body can be
converted into an electric signal either by elec- Coils
tromagnetic or optical transmission. The sim- IT1 I IT21
-
ple construction and low costs make rotame-
ters very popular.
Water vapor is present in any gas stream 1988; VANA, 1982; CARLEYSMITH and Fox,
leaving a bioreactor system. Its measurement 1984). Oxygen has a high magnetic susceptibil-
and control is of great importance in solid ity which changes with temperature. It is para-
state fermentation processes since active or- magnetic and is thus attracted by a magnetic
ganisms require a minimum amount of water field at temperatures below its Curie point of
activity. 80°C. Only nitric oxide (NO) and nitrogen
Instruments for Analysis of Gas Composition 61
Tab. 6. Comparison of Analysis of Volatile Com- Certain oxygen analyzers exploit this thermo-
pounds in the Liquid and Gas Phases magnetic property to create a convective flow
or “magnetic wind” within the measurement
Liquid phase analysis
Advantages Original sample
cell, which is used to cool a heated coil; this
Disadvantages Risk of infection cooling effect is detected by a Wheatstone
Separation of solid particles re- bridge. Fluctuations in pressure will disturb
quired the flow and cause errors. Only hydrogen has
Complex matrix a significantly different thermal conductivity
Imhomogeneities important be- that would require special calibration of the in-
cause of local sampling strument. In some instruments the flow of a
stream of nitrogen across the heated filament
Gas phase analysis is induced by the magnetic properties of the
Advantages Only minor sample treatment re-
test gas. This avoids contact of the test gas
quired
No additional risk of infection with the hot filament and eliminates thermal
Integral sample of the gassed reac- conductivity interference. Other instruments
tor region utilize air as a reference gas; this also has the
Disadvantages Non-equilibrium between gas and result of eliminating the influence of thermal
liquid likely conductivity. The characteristics of the indi-
Delay due to gas-liquid transfer vidual instruments are summarized in Tab. 7.
Magnetomechanical Instruments
dioxide have appreciable magnetic susceptibili-
ties which could interfere. Some instruments use the paramagnetic
properties of oxygen to create the mechanical
deflection of a rigid test body, often a nitro-
Thermomagnetic Instruments gen-filled glass dumbbell, which is suspended
from a fine fiber in a non-uniform magnetic
Heated above 80°C, oxygen becomes dia- field. The force on the body depends on the
magnetic and is repelled by a magnetic field. field intensity, the intensity gradient, and the
Addresses of manufacturers:
Hartmann & Braun AG, Grafstr. 97, D-6000 Frankfurt 90, FRG
Servomex Ltd., Crowborough, Sussex TN6 3DU, GB
Leybold AG, D-6450 Hanau 1, FRG
Maihak AG, Semperstr. 38, D-2000 Hamburg, FRG
Siemens AG, D-8000 Munchen 2, Wittelsbacher Platz 2, FRG
62 2 Methods and Instruments in Fermentation Gas Analysis
Spectral Analysis
Detectors
Following mass analysis and ion detection,
The ions can be detected either by a Faraday the spectral data must be reduced to the con-
cup and electrometer amplifier or, for faster centrations of the various components of a po-
operation and higher sensitivity, with a sec- tentially very complex mixture. This process is
ondary-electron-multiplier (SEM) and subse- known as spectral analysis.
quent electrometer amplification. The Faraday The recorded mass spectral peaks can be
cup is rugged and very reliable; it can be re- represented as a column vector of various ion
garded as an absolute detector of ions and is current signals, S, occurring at p different m / z
typically capable of measurements down to 1- values, due to q different components of a
10 ppm concentration levels. The SEM detec- mixture whose concentrations are cl-cq:
Instruments for Analysis of Gas Composition 65
Pump Valve
Cap i LLa r y
Orifice
1' to Ion-source
Fig. 19. Capillary process-MS-interface.
66 2 Methods and Instruments in Fermentation Gas Analysis
Tab. 9. Mass Spectrometers for Gas Analysis. Precision for Single Stream Analysis of Oxygen in Air (%
absolute); for Switching Time no Accuracy Value is Specific (Specified Absolute Accuracy). Calibration
Interval Recommended. M - magnetic sector, Q - quadrupole, FB - Faraday bucket, SEM - secondary
electron multiplier
References: (1) MGA 1200 - Perkin-Elmer, Norwalk, Connecticut, USA; BUCKLAND et al. (1985); SCHAE-
FER and SCHULTIS (1987). (2) PGM 407 - Balzers AG, Balzers, Liechtenstein; MULLERand RETTINGHAUS
(1988); HEINZLEet al. (1990). (3) MM8-80 - VG Gas Analysis System Ltd., Aston Way, Middlewich, Che-
shire CWlO OHT, England; WINTER(1987); Prima 600 - VG Gas Analysis Systems Ltd. (4) Micromass PC -
VG Quadrupoles, Aston Way, Middlewich, Cheshire CWlO OHT, England. ( 5 ) Questor - Extrel Corpora-
tion, 240 Alpha Drive, Box 11512, Pittsburgh, PA 15238, USA; BARTMAN (1987). (6) MSQ 1003 - Leybold
Heraeus GmbH, Bonner Strasse 498, PO-Box 510760, D-5000 Koln 51, FRG. (7) BOHATKAet al. (1987);
Atomki, Bern ter 18/C, H-4001 Debrecen, Hungary. (8) Bioquad - Spectramass Ltd., Radnor Park Indus-
trial Estate, Back Lane, Congleton, Cheshire CW12 4XR, England. (9) Unigas 325 TT - UTI Instruments
Comp., 325 N. Mathilda Av., PO-Box 519, Sunnyvale, CA 94088-3519, USA. (10) ANAGAZ 200 - Kenos
Analyse, 199, avenue du Marechal-Foch - Elisabethville, 78410 Aubergenville, France. (1 1) DSMS - Hiden
Analytical Ltd., 231 Europe Boulevard, Warrington, Cheshire WA5 5TN, England. (12) Nuclide 3-60-G -
MAAS, Inc., Nuclide Div., 1155 Zion Road, Bellefonte, PA 16823, USA. (13) MAT 271/45 - Finnigan
MAT, 355 River Oaks Parkway, San Jose, CA 95 134, USA
68 2 Methods and Instruments in Fermentation Gas Analysis
OFF
7.6 Electrochemical Analyzers
Carrier gas Electrochemical analyzers are especially
rnl 1
popular for measurement of dissolved O2
(HITCHMAN,1978) and C 0 2 (PUHARet al.,
Loop
1980). These electrodes can be used to measure
gas partial pressures, but their accuracy is
--
Sample low.
Based on similar principles, two types of
D e te c to; electrochemical analyzers are also available for
ON gas analysis (NICHOLS,1988). A fuel cell am-
Fig. 21. Column switching for the gas chromato- perometric device using cathodic reduction of
graphic analysis of air containing C 0 2 . Col 1, Poro- O2 to OH- and anodic oxidation of Pb to
pak Q; Col2, molecular sieve. Pb" is offered (Systech, Wellington Street,
Thame, Oxfordshire, England; Teledyne Ana-
lytical Instruments, 16830 Chestnut Street,
City of Industry, CA 91749, USA). This in-
TENBURG et al., 1979) and to monitor volatile strument has a wide dynamic range detecting
compounds on-line (PONS and ENGASSER, as low as 0.01 ppm and up to 30% of oxygen.
1988). The accuracy is 0.1'70 0,.
Recent developments in applying silicon mi- Another instrument using a zirconia ceramic
cromachining technology to the design of GC electrolyte is of the potentiometric type. An
have resulted in a fast analyzing instrument electrochemical potential is generated which is
(MICROSENSORTECHNOLOGY,1988). The proportional to the logarithm of the oxygen
analysis is usually completed within 30 s. This partial pressure in the gas stream. The measur-
is also true for a number of volatile com- ing cell must be operated at 600 to 800 "C. It is
pounds. The specified repeatability is f2%. especially useful for low concentrations of
oxygen. The attainable accuracies would be
sufficient for many applications ( f0.1 Yo in
the %-range).
7.5 Flame Ionization Detector
This sensitive analyzer is based on the ioni- 7.7 Semiconductor Devices
zation of the sample in a hydrogen flame. It is
most frequently used as a detector in GC and Special semiconducting devices are in-
has therefore been described in some detail in fluenced by chemical environmental variables.
the previous section on GC. The positively and Semiconductors may respond to oxygen (e.g.,
negatively charged ions migrate to their respec- Senox 1050, HBfele, Umweltverfahrenstech-
tive electrodes, thus producing a current pro- nik, Durlacher Allee, D-7500 Karlsruhe, FRG,
portional to their concentration and to the accuracy f 2 % relative), but usually the re-
number of carbon atoms in the molecule. quired accuracy cannot be obtained.
It has been used as a stand-alone instrument More interesting developments have oc-
without chromatographic separation to deter- curred for monitoring volatile organic com-
70 2 Methods and Instruments in Fermentation Gas Analysis
pounds such as ethanol. Such sensors have mation of ethanol production. The metabolic
been used for control (VORLOPet al., 1984; basis is that complete oxidation of sugar gives
AXELSSON,1988). Vogelbusch (PO Box 52, A- a respiratory quotient of nearly 1.0 according
1051, Vienna) offers complete equipment for to the following reaction:
control of molasses feed to ethanol-producing
fermentation and for vinegar production based C6HI2O6 0,+ + 6 CO, +6H20 (1 11)
on such semiconductor sensors.
Palladium-MOSFETs are sensitive to hy- Equimolar amounts of oxygen and COz are
drogen (DANIELSSONet al., 1984) and can, consumed and produced. Ethanol is produced
therefore, be used for monitoring this gas in according to
corresponding biochemical reactions. This sen-
sor could be useful in monitoring the key inter- C6HI2O6--t 2 C 2 H 5 0 H+ 2C02 (112)
mediate hydrogen in anaerobic digestion. Un-
fortunately, H2S, which usually is present in For this reaction the RQ would be infinitely
anaerobic digestion, interferes with the meas- high. Summing these two equations, we can
urement. see that the amount of ethanol produced is
Gas phase biosensors (GUILBAULTand proportional to the difference of CO, produc-
LUONG,1988) have an interesting potential in tion and oxygen consumption and that the rate
high sensitivity monitoring of trace gases. This of ethanol production correspondingly would
may be a future method for sensing fermenta- be proportional to the difference of C0,-pro-
tion components with very low vapor pres- duction rate minus 0, consumption rate
sure.
QE = Qco, - Qo, (1 13)
The respiratory quotient (RQ = CPR/
OUR = Qco,/Qo,) can be used to control sug-
8 Examples of Application ar feed to a glucose-repressed baker’s yeast
culture (WANGet al., 1979; SPRUYTENBURG et
Numerous examples in the literature de- al., 1979) to avoid accumulation of ethanol.
scribe monitoring of traces of off-gases during
fermentations. Many examples involve the es-
timation of oxygen uptake rate (OUR), carbon 8.2 Mass Balancing in a Penicillin
dioxide production rate (CPR), and usually Fed-Batch Fermentation by C 0 2
also the respiratory quotient (RQ), but in al-
most all cases neglecting any deviation from Measurements
steady state and also neglecting C 0 2 accumula- (Mou and COONEY,1983)
tion as HCO; in the liquid phase. Rather few
examples exist in which the gas analysis results In this example biomass production was
have been used for further process analysis found to be proportional to CO, production
and eventual process control. rate (CPR)
Stoichiometric relations have especially
been used for primary metabolite production,
where metabolic pathways usually are well es-
tablished.
Examples are: where Yx,co, is the yield coefficient (g cell/
mmol COz) and X,is the biomass concentra-
tion at time t.
8.1 Ethanol Production By making a carbon balance for all signifi-
cant components (substrate, precursor, COz,
If no analytic method of direct on-line meas- penicillin, and cell mass) and using an empiri-
urement for ethanol in yeast cultures is availa- cal correlation for penicillin production, the
ble, gas balancing can be used to get an esti- following equation was obtained:
References 71
(CHANG, T. M. S., Ed.). New York: Plenum YAMANE,T., SHIMIZU,S. (1984), Fed-batch tech-
Press. niques in microbial processes, Adv. Biochem.
WEAVER,J. C., ABRAMS,J. H. (1979), Use of a Eng. Biotechnol. 30, 147-194.
variable pH interface to a mass spectrometer for ZABRISKIE, D. W. (1985), Data analysis, in: Com-
the measurement of dissolved volatile com- prehensive Biotechnology, Vol. 2, pp. 175-190
pounds, Rev. Sci. Instrum. 50, 478-481. (MOO-YOUNG, M., Ed.). Oxford: Pergamon
WINTER,M. J. (1987), The application of single de- Press.
tector sector mass spectrometer systems in fer- ZADRAZIL,F., GRABBE,K. (1983), Edible mush-
mentation off-gas analysis, in: Mass Spectrome- rooms, in: Biotechnology, 1st Ed., Vol. 3, pp.
try in Biotechnological Process Analysis, pp. 17- 145-189 (REHM, H.-J., REED, G., Eds.). Wein-
38 (HEINZLE,E., REUSS,M., Eds.). New York: heim: Verlag Chemie.
Plenum Press.
3 Biosensors
B o MATTIASSON
Lund, Sweden
1 Biosensors - a Definition 77
2 The Biocomponent of the Biosensor 77
2.1 Biocatalysts 77
2.1.1 Enzymes 77
2.1.2 Cells 78
2.1.3 Organelles 79
2.2 Binding Assays 79
2.2.1 Immunochemicals 80
2.2.2 Receptors 80
3 Methods to Integrate the Biocomponent with the Transducer 81
3.1 Immobilized Layer Covering the Transducer 82
3.2 Pre-Column of Immobilized Biocomponent 83
3.3 Direct Immobilization of the Biocomponent on the Transducer 83
3.4 Reversible Immobilization by Means of Mechanical Arrangement 83
4 Transducers 83
4.1 Electrochemical Detectors 83
4.1.1 Amperometric Detectors 84
4.1.2 Potentiometric Biosensors 84
4.1.3 Fuel Cells 84
4.2 Thermometric Detectors 85
4.3 Optical Transducers in Biosensor Construction 85
5 Biocatalytically Active Biosensors 86
5.1 Enzyme-Based Sensors: Examples and Applications 86
5.2 Cell-Based Sensors: Examples and Applications 91
5.2.1 Analysis of Specific Compounds, e.g., Vitamins 91
5.2.2 Analysis of BOD 92
5.2.3 Analysis of Inhibiting Substances 92
5.2.4 Analysis of Specific Metabolites 92
76 3 Biosensors
3
Biosensor
concept.
Soecific
2.1.1 Enzymes
bi o rn olecuie Transducer
Fig. 1. Schematic presentation of a biosensor, with The most commonly used binding pair is en-
the biocomponent imparting biospecificity and a zyme-substrate. Large numbers of enzymes are
transducer registering the signal and transforming it available. So far more than 2000 enzymes have
into an observable signal. been characterized and listed in the enzyme
nomenclature handbook (IUB, 1979). There
are many more that have been studied, but so
far not isolated in pure form and character-
It should be mentioned that there are also ized. In other words, there are many enzyme-
other definitions of biosensors. The broadest substrate interactions that can be used for the
definition involves any kind of measurement monitoring of many different substrates. En-
made on a biological system. Since in that case zymes can also be used to detect and even
almost any kind of analytical device can be re- quantify inhibitors. Such molecules influence
garded as a biosensor, as long as it is used in the catalytic process. By exposing the enzyme
studying biological systems, the first, more re- to substrate under controlled conditions, the
stricted definition will be employed. influence of an inhibitor may be recognized as
an altered enzyme activity. Substances activat-
ing the enzyme may likewise be monitored in
this way. In general such interactions can be
used for quantitation down to mol/L for
substrates and competitive inhibitors. The
measurement of noncompetitive inhibitors is
more sensitive by a few orders of magnitude.
78 3 Biosensors
the reasons why cell-based sensors have not yet 2.2 Binding Assays
become more popular.
The physiological status of the cells em-
ployed may vary; Tab. 3 lists some different There is a whole range of interactions that
stages that have been used. Fully viable cells can be summarized as binding reactions. Im-
have all the characteristics of the native cell munochemical binding reactions are the best
studied and also the most successful on the
market. Receptor-based analysis has been pre-
dicted to have a bright future, but so far there
Tab. 3. Conditions of Cells Used for Constructing
Biosensors has been no real breakthrough for this kind of
assay. Lectin-carbohydrate interactions are
Only one or a few enzymes involved still another group of interactions.
Resting cells Generally speaking, these assays have the
Permeabilized cells potential for high sensitivity since the forma-
Living cells with an intact surface tion of affinity complexes between a macro-
Actively growing and dividing cells molecule and a ligand is very efficient even at
Mixed cultures of living cells, or mixtures of en- low concentrations. In most of the reported
zymes and cells
Cells co-immobilized with a specific enzyme or or-
cases the binding assay has been set up as a
ganelle competitive assay where the native ligand in
Immobilized cells or organelles with a substantial the sample to be analyzed competes with a la-
part of the metabolism intact beled ligand that is added in a fixed concentra-
tion. The result of the competition is then read
by evaluating how much of the label is
bound.
and may be very suitable when used to moni-
tor events that are coupled to cell growth and
cell division. In other cases the cell may only Tab. 4. Comparison between Enzymatic and Immu-
be regarded as a package of the appropriate nochemical Analyses
enzyme for the catalytic step to be used in the
analysis. In addition, one may exploit the Enzyme/ Antibody/
membrane properties of the intact cell in the Substrate Antigen
sense that membrane-modifying activities can
be monitored by following the metabolism of Operational range 10 -12
intact cells. (mol/L)
Specificity t t +
Availability for new
analyses - t
2.1.3 Organelles Ease of production - +
Usefulness in quan-
The same approach used for cells also ap- tifying macromole-
plies to organelles, but their higher fragility cular structures - +
and resulting tendency to lose activity has Time needed + -
hampered development.
An electrode based on rat liver microsomes
was used for the determination of thyroxin
(MEYERHOF and RECHNITZ,1979), and a glu- In more recent developments, the trend is
tamine sensor was obtained by immobilizing towards direct monitoring of the affinity inter-
the mitochondria1 fraction from porcine kid- action between the binder and the ligand. If
ney cortex cells on an ammonia gas sensor this can be done directly, easier and perhaps
(ARNOLDand RECHNITZ,1980). Microsomes faster assays can be set up.
may either be used separately or co-immobil- Binding assays can be set up according to
ized with other biocatalysts (SCHUBERT and the dipstick method, in which the binder is
SCHELLER,1988). used as a disposable chemical; otherwise, the
80 3 Biosensors
binder must be regenerated, which then puts tween different clones. Selection therefore had
some constraints on the stability of the binding also to include tests for stability, which is espe-
molecule. cially important when repeated use is plan-
In many cases binding assays are useful in ned.
concentration ranges down to lo-" mol/L, For discrete assays, when the antibody is a
and they are also useful for quantifying ma- disposable reagent, one must screen for the
cromolecules. Since most of the reported work strongest binder with the highest affinity. This
is focused on immunoassays, the discussion in may, however, not be ideal when constructing
this chapter will mainly deal with such assays. biosensors. As will be discussed later in this re-
Tab. 4 shows a comparison between the prop- view, repeated use of the antibody involves
erties of enzyme-based assays and immunoas- dissociation of immunocomplexes. When deal-
says. At the present state of development the ing with very strong complexes, harsh condi-
two types complement each other. tions must be applied and partial denaturation
of the antibody may occur.
When immobilizing antibodies it is desirable
2.2.1 Immunochemicals to avoid binding to their specific binding sites
and instead to couple via the Fc-parts. This
The basis for immunochemical binding as- can either be done by using covalent coupling
says is the specific binding of antigens to anti- to the carbohydrate chains on the Fc-fragment
bodies. When immunizing an animal by intro- (O'SHANNESSYand HOFFMAN, 1987) or by
ducing a foreign macromolecule, a protective biospecific reversible immobilization via inter-
reaction is initiated that leads to formation of action with protein A or protein G (LANCETet
antibodies. The mechanism involves activation al., 1978).
of lymphocytes to produce antibodies. Each
such activated lymphocyte produces only one
type of antibody, but since many lymphocytes 2.2.2 Receptors
respond to the immunization, the result is a
polyclonal antiserum. This means that the an- Receptor-based analysis is a great challenge,
tibodies may differ in their molecular proper- since our knowledge of receptors and their
ties, e.g., binding strength, specificity, pH-de- properties is rather poor. Only recently have
pendence, and stability. However, because of receptors become available in pure form, and
the number of different strains, it often seems it is still far from standard technology to pre-
as if the overall characteristics of the antise- pare such entities. Furthermore, since the re-
rum are stability and good binding proper- ceptors are membrane bound, the handling of
ties. the intact molecule is quite different from the
The binding constants are reported to be in treatment that is given traditional proteins.
the region of l o p 5 to lo-" mol/L. By tradi- This shows that at present we have no technol-
tion, all selections of antisera have been for ogy to successfully integrate such molecules
high-affinity antibodies, since these allow into applications such as biosensors. One easy
higher sensitivity in the subsequent binding as- and probably viable way would be to use very
say. crude preparations, consisting of more or less
The introduction of the hybridoma technol- whole cells or cell fragments, since by using
ogy for production of monoclonal antibodies such an approach the purification and stability
made it possible to obtain large amounts of in- problem of the receptor could be circum-
dividual clones of antibodies. This led to the vented.
standardization of some assays, but also
helped to unravel some points that had been
taken for granted when dealing with polyclon-
a1 antibodies. The different clones had to be
screened for good binding constants and for
the desired specificity. Furthermore, it soon
turned out that stability varied drastically be-
Methods to Integrate the Biocomponent with the Transducer 81
3 Methods to Integrate
the Biocomponent
with the Transducer
It has been said that biosensors are based on
proximity of the biocomponent and the trans- lime
ducer. In order to make this technically feasi- Fig. 2. Schematic presentation of the response of an
ble, the biocomponent must be rendered more enzyme electrode as a function of the load of en-
easily handled than an enzyme in solution. zyme.
Furthermore, if a stable signal is expected
from the biosensor, precautions must be taken
to compensate for denaturation, etc. One way
to do that is by operating at high concentra- cal activity and that the coupling is stable, i. e.,
tions, i.e., with an excess of the biocompo- the biomolecule sticks to the support and does
nent, so that the reaction becomes diffusion not leak away. In the case of immobilization
controlled. of enzymes, it has sometimes turned out that
Immobilization of biomolecules is a tech- an essential amino acid residue in the active
nology that is handy and permits operation site is very reactive and thus reacts first in the
with high concentrations. By immobilizing the coupling procedure. In this case, other chemi-
biocomponent one can apply it in situations cal approaches must be considered to avoid
and under circumstances that would otherwise utilizing these amino acid residues. An alterna-
have been practically impossible. Furthermore, tive is to try to protect the enzyme during cou-
immobilized biochemicals can in principle be pling, e.g., by immobilizing it in the presence
reused. This leads to conditions suitable for of a competitive inhibitor that will block the
continuous measurements or repetitive assays. active site (JULLIARDet a]., 1971; MATTIAS-
However, under these circumstances the oper- SON et al., 1974; BULOW and MOSBACH,
ational stability of the preparation is impor- 1982). In some cases substrates have been
tant, since a reliable signal is of utmost impor- used.
tance. Fig. 2 shows that using a large excess of Another immobilization method that has
the biomolecule in enzyme-based assays has become very popular, especially for cells and
turned out to be successful. cell organelles, is entrapment (MATTIASSON,
There are many different ways to carry out 1983). The particulate matter is mixed with a
immobilization. This has been well described solution of either a water-soluble polymer or
in recent reviews and books (MOSBACH,1976, of monomers. The water-soluble polymers are
1988; MATTIASSON, 1983) and, therefore, it then crosslinked to form a three-dimensional
shall only be treated as required for the discus- network in which the particulate matter is en-
sion later on in the chapter. trapped. In the case of monomer solutions,
The most popular way of immobilizing a polymerization is initiated and a similar three-
protein is by covalent coupling (MOSBACH, dimensional network is formed.
1976). A solid support is used with chemical Still another way of carrying out immobili-
groups that can be activated in such a way that zation has been developed especially for labile
the activated form will react with specific and/or expensive enzymes. In these cases it is
groups on the protein surface. There are many not ideal to immobilize a large excess of the
different coupling methods from which to enzyme, since from an economical point of
choose. The groups available on the support view it may be impossible or ill advised. If the
are of course important, and often that is the enzyme is very labile, a large excess of enzyme
first consideration in the choice of a coupling molecules will only marginally improve the op-
method. However, it is even more important erational stability. In such cases it is important
that the coupled protein maintains its biologi- to be able to immobilize the biocomponent
82 3 Biosensors
B t
Im mob ilized
enzyme
4.1.1 Amperometric Detectors SUZUKIet al., 1988). There is still a long way
to go before these new miniaturized sensors
Amperometric detection is based on measur- are comparable with the more traditional ones
ing the current obtained when the analyte is as far as performance, reliability, and stability
either electrochemically oxidized or reduced at are concerned. Thus, the microglucose sensor
a given solution-electrode interface. The mea- constructed along these lines was reported to
sured current is given by have a lower stability than the traditional sen-
sors. While the biocomponent had the same
i = n FdN/dl properties as usual, the transducer was less sta-
ble (KARUBEet al., 1988). The small dimen-
Proportionality between the current i and sions, however, hold such interesting possibili-
the concentration of the analyte N can be ob- ties that it is still of great interest to continue
tained, provided some prerequisites are met. this development.
A limiting factor for a wider application of
amperometric enzyme assays has been a short-
++
a1 bl age of suitable oxidases. As new enzymes are
discovered and isolated, the area of ampero-
metric enzyme sensors will broaden.
More recent development has focused on zyme technology, flow injection analysis, and
tapping the electron flow from microorgan- spectrophotometric detection. This area is re-
isms. By introduction of suitable redox media- latively well developed and some devices have
tors the efficiency of this process has been im- also reached the market. These devices were
proved. One interesting application of these mainly developed for clinical analysis and have
sensors is the cell counter, where the amplitude recently been applied for on-line analysis of
of the signal reflects the number of active cells bioprocesses (SCHUGERL,1988). The sensitivi-
in the analyzed sample. Such sensors are app- ty is determined by Lambert-Beer's law. The
licable both to bacterial and mammalian cells light path used is often 1 cm or shorter. A
(MATSUNAGAet al., 1980; MIYABAYASHI et higher sensitivity would be attainable if a long-
al., 1987). er light path were used. There are several ad-
Biological fuel cells attract great interest be- vantages in using spectrophotometric detec-
cause of the potential that one may be able to tion. A huge amount of reference literature
generate electricity directly from the biological from conventional biochemistry is available.
process. Besides monitoring absorbance/transmit-
tance and fluorescence, it is also of interest to
register light scattering, since in that case the
4.2 Thermometric Detectors amount of particulate matter may be analyzed.
This is valuable when quantifying cells and
Most biocatalytic reactions produce/con- also flocs created by, e.g., immunochemical
sume some heat, and thus in theory most reac- binding.
tions could also be monitored by means of a The instrumentation required for the above
thermometrical detector. Because of the lack analyses has often been a conventional labora-
of selectivity of the transducer it has been im- tory unit equipped with a flow cell. However,
portant to compensate for various nonspecific there has been a need for making the optical
reactions. This has mainly been carried out by reading outside the instrument and close to the
the use of split-flow devices with one sample bioreaction. This has been achieved by the in-
and one reference transducer. troduction of fiber optics.
The most commonly used transducer is the An optical sensor configuration is schemati-
thermistor, a semiconductor with a resistance cally shown in Fig. 6A-C. The basic construc-
that is temperature-dependent. With a Wheat- tion of an optical fiber sensor is shown in Fig.
stone bridge it is possible to monitor the 7 where the two mechanisms of light exchange
change in resistance as a biocatalytic process are both illustrated. The first involves an illu-
proceeds. Thermistors measure temperature mination cone from the end of the fiber. The
changes down to l o p 2"C. This has been con- alternative is an evanescent wave from the
verted into sensitivities in the enzyme thermis- naked portion of the fiber. In the general con-
tor concept of mol/L, and for the en- figuration the fiber is covered with an opaque
zyme thermal probe it is less sensitive by a few covering, but when evanescent wave measure-
orders of magnitude (WEAVERet al., 1976; ments are used, the wall of the fiber must be
FULTONet al., 1980). Thermocouples have directly exposed to the sample solution.
also been used, as have Peltier elements (PEN- Today many different applications are
NINGTON,1976). based on this technology. The p H , for exam-
ple, is monitored using an optical fiber with a
p H indicator immobilized at its tip (PETERSON
4.3 Optical Transducers in et al., 1980). When the pH changes, the ab-
sorption spectrum of the indicator changes,
Biosensor Construction and thus the p H at the tip of the fiber can be
read.
Measurement of light absorption/transmis- Optical sensor configurations for monitor-
sion is a standard technique in a biochemical ing p C 0 2 , p 0 2 or dissolved oxygen, metal
laboratory. Application of small flow cells has ions, etc. have also been described (SEITZ,
made it possible to combine immobilized en- 1987).
86 3 Biosensors
A Opaque covering
Illumination zone
region
Fig. 7. An optical fiber illustrating the two different
principles used for constructing optical biosensors.
At the tip of the sensor is an illumination zone. If an
I immobilized reagent
- phase
. is placed in this zone,
t
-
any changes, e.g., by binding t o the zone may be
detected. In the evanescent region the fiber is naked
c after removal of the cladding layers; here evanescent
wave measurements may be performed. In that case
a solution similar to that in Fig. 6C is achieved.
When determining the degree of freshness of Recently, an alternative way to avoid the
meat, a measure of the glucose content is in- negative effects of hydrogen peroxide resulting
versely proportional to the number of bacteria from the oxidase activity has been presented.
present (J. HIGGINS,Cranfield Institute, UK, The enzyme performs the oxidation of the sub-
personal communication). strate with the concomitant reduction of FAD
1 11?-~
Much of the biosensor development has to FADH2. The latter is regenerated via media-
been conducted by clinical chemists. In order tors replacing oxygen. The mediators are then
to standardize and simplify the handling of the oxidized by the electrode.
analyses, application of immobilized enzymes
in analysis became attractive. The interest fo- glucose GODIFAD 2Mt
cused on analysis of discrete samples (DA-
NIELSON et al., 1977), whereas in intensive
electrode
care and in fermentation control the desire has
been for continuous monitoring (ENFORS,
1981; CLARKet al., 1988), or at least frequent glucono- GOD/FADH2 2e-
intermittent analyses (HOLSTet al., 1988; HA- lactone
KANSON, 1989).
The enzyme used for setting up these assays Ferricyanide may be used as mediator, but
has in most cases been glucose oxidase (E.C. the most popular one belongs to the ferrocene/
1.1.3.4). This enzyme, produced by Aspergil- ferrocinium group (CASS et al., 1984).
lus niger, is a glycoprotein with FAD as pros- One step further would be to directly
thetic group. It does not depend upon any sol- branch off the electrons from FADHz without
uble cofactor, and this has facilitated its devel- use of mediators. Efforts in this direction have
opment. A drawback is that hydrogen per- involved direct coupling of FAD to the electro-
oxide is formed during the reaction: de surface, thereby facilitating the reoxidation
of FADH2 (WINGARD1984; ALBERYet al.,
glucose + 0, + gluconolactone + H 2 0 2 1985).
An alternative arrangement is to covalently
Hydrogen peroxide is harmful to many en- attach mediator molecules to the enzyme,
zymes. During the initial experiments with glu- thereby making it easier to branch off the elec-
cose electrodes these harmful effects were not trons from the interior of the protein over to
observed, and the enzyme was regarded as ex- the electrode surface (DEGANIand HELLER,
tremly stable. It turned out later that the inter- 1987, 1988; BARTLETTet al., 1987).
mittent exposure of the enzyme preparation to This development further leads to the use of
the hydrogen peroxide did not cause much conducting polymers to tap the electron flux to
harm, whereas when monitoring continuously or from the biocomponent of the biosensor.
and thus exposing the enzyme continuously to By this way of linking the reaction before for-
hydrogen peroxide, a rapid decrease in activity mation of hydrogen peroxide, two things may
was observed. The harmful effects are ascribed be achieved: avoidance of hydrogen peroxide
to the oxidative activity of the peroxide as well formation and elimination of oxygen depend-
as to the radicals formed. In order to stabilize ence, thereby making it possible to operate
the enzymes, co-immobilization has been ap- over extended ranges of concentration.
plied. These stability problems were first en- When monitoring the glucose oxidase cata-
countered when immobilized glucose oxidase lyzed process, several options are at hand. The
was used in enzyme technological process ap- most common option is to use a polarographic
plications (MESSING, 1974; BOUIN et al., oxygen sensor to register the oxygen tension.
1976a, b). Experience from that field was ap- When glucose is present, this will be reflected
plied in the biosensor work. By coupling catal- in a decreased oxygen tension. An alternative
ase with glucose oxidase, instantaneous de- is to use a different polarization current and
gradation of the peroxide was achieved. When register the hydrogen peroxide formed. This
superoxide dismutase was also included, any makes it easier to avoid interference by other
radicals formed were eliminated. substances in the sample, e.g., by ascorbic
88 3 Biosensors
acid. This sounds like a minor problem in fer- LAND and ENFORS, 1984). Oxygen was pro-
mentation control, but in clinical analysis and duced at the tip of an oxygen electrode by elec-
also in medical applications it is a definite trolysis of water. The oxygen tension was set
problem. A third alternative in monitoring the at a constant value, and thus the amount of
glucose oxidase catalyzed process is detection water that needed to be hydrolyzed was pro-
of the protons formed upon hydrolysis of the portional to the amount of glucose present.
resulting gluconolactone. The gluconic acid The energy needed to perform this hydrolysis
thus formed is dissociated and it lowers the was used to read the glucose concentration.
pH. However, it should be stated that pH-sen- This approach provided an oxygen electrode
sitive electrodes as transducers in biosensor that was linear up to at least 20 mmol L-'.
work have not been very successful. The buf- Other enzymes are available for construct-
fering capacity rarely remains constant for ing an enzyme based glucose assay. Hexokin-
long periods of time, and thus the sensitivity in ase catalyzes the phosphorylation of glucose at
detecting the acid produced varies with time the expense of ATP, and the glucose-6-phos-
(NILSSONet al., 1973; ENFORSand NILSSON, phate formed may in a subsequent step be
1979; RUSLINGet al., 1976). To make this oxidized by glucose-6-phosphate dehydrogen-
kind of sensor useful, eleborate standardiza- ase with a concomitant reduction of NADP +.
tion and calibration has to be carried out. This can be monitored by a fluorometer or
Another detection principle that has been spectrophotometer. Electrochemical alterna-
applied to monitor glucose oxidase activity is tives are also available (SCHELTER-GRAFet
thermal registration by thermistors (DANIELS- al., 1984). However, in this case there is a re-
SON et al., 1977). Here, the heat of reaction is quirement for a free cofactor, and therefore
registered. The glucose-containing sample is such a method is only suitable for analyzing
pumped through a bed of immobilized glucose discrete samples under well-controlled labora-
oxidase, and at the outlet of the bed a thermis- tory conditions.
tor registers the temperature of the effluent. A more recent alternative is based on a new
When glucose is present in the sample, enzyme group of enzymes, the PQQ-enzymes utilizing
catalysis takes place during transit through the pyrrolo quinoline quinone as cofactor. It has
column, and the reaction heat is transported been found that many enzymes that carry out
with the effluent. The heat is proportional to oxidation reactions do not operate with any of
the amount of glucose converted. Since the the traditional biochemical cofactors, but in-
system operates with reproducible volumes, a stead use PQQ (MATSUSHITAet al., 1982;
clear relation exists between concentration of NEIJSSELet al., 1983). This cofactor is sponta-
glucose and heat generated. In this system as neously regenerated in situ and such systems
well as in the others, oxygen concentration is therefore require no external addition of a co-
limiting. The solubility of oxygen in water is factor (D'COSTAet al., 1986).
only 0.25 mmol/L at 25 "C. This means that
one can only expect a linear correlation be-
tween the concentration of glucose and the re- Glucose Analyzer for Blood Monitoring or
sponse up to approx. 0.5 mmol/L. However, Fermentation Control
when catalase is present some oxygen is pro-
duced as the hydrogen peroxide formed is de- In addition to the mutual arrangement of
graded. In this manner an expansion of the the immobilized enzyme and the transducer in
concentration range is achieved. certain positions, there is a need for a rational
By operating with thick enzyme-membranes way of treating the sample. It does not matter
it has been possible to shift the dynamic range how well a sensor functions if the sample han-
for glucose electrodes toward higher concen- dling technology is poor. Below, the sampling
trations. unit, the sample treatment system, and the
There are different ways to improve oxygen analytical part are described in detail. These
supply to such sensing systems (ADLERCREUTZ three units belong together when setting up
and MATTIASSON,1982). One elegant system successful monitoring systems (Fig. 8) (HOLST
was reported by ENFORSand coworkers (CLE- et al., 1988). The preparation of immobilized
Biocatalytically Active Biosensors 89
Fig. 8. Flow scheme for the glucose monitor used for fermentation control. (From
HOLSTet al., 1988, with permission)
enzyme is designed to last in the clinical situa- plication it is not the sensitivity that raises
tion for five days, and for the industrial appli- problems; it is merely the sampling and the
cation to last for one fermentation. Even if it sample handling.
would be possible to use one and the same en-
zyme unit for two or more fermentations, it is
better to replace it after each fermentation. Thermistor Application
For glucose monitoring in clinical samples a
different operational range is needed than for Thermometric registration of the enzymatic
fermentations. The level of operation may well reaction is the most general of the different
be controlled by designing the dialysis step ways to follow enzymatic reactions discussed
properly, either by shifting to a more suitable here. Since it may also be used in optically
membrane, or by varying the flows on either very dense solutions, it offers certain potential
side of the membrane. Blood glucose levels are
normally around 5 mmol/L; therefore, the
system is optimized for that concentration I I , I ,
range. Since the unit is based on a polaro- 30 30
graphic oxygen sensor, any oxidative process mrnol l-f mmal t-’
may be monitored. By applying lactate oxi-
dase, a useful lactate sensor for continuous t20
monitoring was achieved. However, due to
stability problems of the enzyme, a sampling
frequency has to be used that protects the en- 110 i
c
m
are fully acceptable. Fig. 9 shows the results of Fig. 9. Tracings of glucose and lactic acid using two
one such cultivation where both glucose and analytical devices as shown in Fig. 8, one equipped
lactate were monitored using two separate with immobilized glucose oxidase and the other with
monitors (HAKANSONet al., 1990). In this ap- immobilized lactate oxidase.
90 3 Biosensors
ditional analyses. Therefore, after each analy- to use a dense population of immobilized cells
sis the cell preparation has to be replaced. In and to obtain a quick answer (DISSINGet al.,
order to make comparisons between different 1984).
experiments, it is important to mount the
membrane on the transducer in a very repro-
ducible way. This was achieved by using an 5.2.3 Analysis
electromagnetic holder. Several analyses of vi-
tamins have been described (MATSUNAGAet of Inhibiting Substances
al., 1978).
The introduction of biosensors has speeded The effects of inhibitory substances may be
up bioassays to such an extent that they may analyzed by exposure of cells to a surplus of
become a realistic alternative to binding as- substrate and a suspected inhibiting factor.
says. Since whole cell metabolism is involved, one
can monitor both specific toxic compounds in-
teracting with one specific site in the cell and
5.2.2 Analysis of BOD more generally acting compounds.
cules, this is often the case. Depending on the basis for immunoassays that are widely used in
complexity of the medium to be analyzed, it is clinical chemistry and are now expanding into
important to make sure that there are no com- veterinary medicine, agricultural analysis, en-
peting or alternative reactions going on. To vironmental analysis, and food analysis. The
achieve this, partial denaturation of the cells assays available on the market are mainly
has sometimes to be carried out. The analyzing based on establishing equilibrium in the bind-
technique and the problems encountered are ing reaction between the two reactants. This
very much the same as for enzyme-based bio- leads to long incubation times and thus labo-
sensors. rious handling. When designing biosensors for
the same reactant pairs, two main types of im-
munosensors can be identified: the dip-stick
type and the reversible type.
The other reactant pairs described in Tab. 5
6 Binding Assays: can all be utilized for biosensor construction.
So far very few sensors have been described.
Examples and Applications Receptor-based sensors will be discussed in
this paragraph. The other potential biosensors,
Biochemical recognition assays involving no e.g., lectin sensors, may very well become im-
catalytic reaction are often grouped as binding portant in the future. Today, however, very
assays. A broad spectrum of reactant pairs can few publications are available (BORREBAECK
be utilized to set up such assays. Tab. 5 lists and MATTIASSON, 1980).
some of the reactant pairs and their analytical
applications. Enzymes are not included in the
table even if one, at least in theory, can use an 6.1 Immunosensors
enzyme for binding without catalysis. This
may be carried out in pH-regions where the en-
zyme is capable of binding, but where catalysis 6.1.1 Dip-Stick Immunosensor
does not occur. The most commonly studied
reactant pair is antigedantibody. It forms the By tradition, antisera have been selected ac-
cording to the binding constant. The stronger
the binding the better. This has formed the ba-
Tab. 5. Reactant Pairs with Biospecific Binding and sis for sensitive binding assays. With strong
Examples of Their Use in Analysis binding, reversibility in the binding has been
sacrificed. This means that immunoassays set
Entity 1 Entity 2 Application up according to traditional principles are sensi-
tive, but that it is difficult to split the complex
Antibody Antigen Immunoassay and reuse the immobilized entity. When de-
Antibody Hapten Immunoassay signing biosensors in this way, a disposable
Avidin Biotin Immuno/histo- sensor is the result, i.e., an immuno-dip-stick.
chemistry
Protein A Fc on IgG Immunoassay
There may be reasons for using such devices,
Protein G Fc on IgG Immunoassay especially if they can be designed to avoid wet
Lectin Carbohydrate Carbohydrate assay chemistry. However, it is doubtful whether
Glycoprotein they can be regarded as biosensors in the strict
Glycolipid sense.
DNA DNA-probe Specific DNA A variation on this theme is to dissociate the
analysis bound compound and to reuse the immobil-
RNA RNA-probe Specific RNA ized entity. This is the basis for flow injection
analysis ELISA (MATTIASSON et al., 1977; MATTIAS-
Receptor Ligand Receptor assay
SON and LARSSON,1987; MATTIASSON et al.,
Ligand assay
Enzyme Inhibitor Enzymehnhibitor 1989). Here, the antibody is immobilized to a
Enzyme Activator Enzyme/activator support with very low nonspecific adsorption,
e.g., Sepharose. It is then placed in a small
94 3 Biosensors
column in a continuous flow of buffer. The The shorter the time of exposure, the faster the
sample to be analyzed is mixed with labelled analysis. However, to achieve speed one has to
antigen prior to injection into the flow system. sacrifice sensitivity. In most cases it is possible
The sample will pass the column of immobil- to carry out analyses in the region of lo-* to
ized antibodies as a short pulse. During pas- l o e 9 mol/L. If higher sensitivity is needed
sage there is competition for binding between longer times of exposure can be used. The ex-
labelled and native antigen. In a subsequent perimental procedure involves very few manu-
step a pulse of substrate is administered to the al steps, and, therefore, it is possible to sub-
system, and the amount of enzyme-labelled stantially reduce experimental error in compar-
antigen is read. From this it is possible to de- ison with what is normal for ELISA.
duce the amount of native antigen present in Flow injection ELISA has been performed
the sample. The assay cycle is terminated by with several different sensors such as polaro-
splitting the immunocomplex, for instance, at graphic oxygen sensors (MATTIASSONand
a low p H (glycine p H 2.2), and after recondi- NILSSON, 1977), spectrophotometers (MAT-
tioning the system is ready for another cycle. TIASSON and BORREBAECK, 1978), and flow
Upon repeated use one can expect inactivation calorimeters (MATTIASSONet al., 1977).
of the antibodies and, therefore, a decreased More recently a sandwich type of assay us-
capacity of the immunosorbent. However, on ing the same experimental setup has been pub-
passage through the column, the native and lished (DE ALWISand WILSON,1985).
the labelled antigen will compete under identi- As indicated in Sect. 3, several arrange-
cal conditions even if the absolute number of ments of the biocomponent in relation to the
transducer are known. By covering the tip of
the electrodes with a layer of the immunoac-
tive component, true immunoelectrodes may
be achieved. In most cases they are based on
competitive binding assays as described for the
flow-ELISA-procedure. When such an elec-
trode is placed in a flow system, a flow-
a: ELISA-system is obtained (MATTIASSONet
al., 1977; MATTIASSON and LARSSON,1987).
20 1 The alternative is to dip the immunoelectrode
0 200 h 200 into the solution (AIZAWAet al., 1979; DOYLE
Time et al., 1984).
Fig. 12. Response stability of an antibody-Sepha- There may be many further alternative con-
rose CL 4B-column. Curve 1: peak height when figurations (e.g., immuno-FET) but the basic
passing a 100% sample, i.e., a sample with only en- chemistry and performance is the same.
zyme-labelled antigen. Curve 2: peak height in per-
cent of a preceding pure aggregate pulse obtained
for reference samples containing a constant amount
of native antigen, in this case human serum albu- 6.1.2 Reversible Immunoassays
min. (From BORREBAECK et al., 1978, with permis-
sion)
If immunoassays are to be used in fermenta-
tion control or continuous monitoring in medi-
cal biotechnology, it facilitates the handling
binding sites decreases. Therefore, the relative substantially if a direct binding assay is set up.
amount of the bound labelled antigen will re- In such a system direct interaction between the
main constant, even if the absolute quantity of native antigen in the sample and the immobil-
both labelled and native antigen is reduced, as ized antibody results in a readable signal. Sev-
can be seen in Fig. 12 (BORREBAECK et al., eral potential solutions have been presented to
1978). meet these requirements (Tab. 6 ) . Some of
This concept of fast competitive binding op- these are based on expensive equipment and
erates under conditions far from equilibrium. others are very sensitive to external influences.
Binding Assays: Examples and Applications 95
--
m
Pheromones and insect antennae demon-
c strate clearly what can be achieved when con-
c
aJ
c structing biosensors based on receptor-ligand
a
0
interactions. To set up such an assay one needs
to be able either to isolate the receptor or to
include the whole cell or organ where the re-
ceptor is located.
One preliminary sensor was constructed by
using whole crab antennae. The sensor showed
very low stability and may be regarded only as
an academic demonstration that it is possible
to use receptors for constructing biosensors.
20 30 t The construction of a receptor-based assay
lo Time for nerve gases was described in an interesting
Fig. 14. Time course for production of IgM from a presentation (TAYLORet al., 1988). The sensi-
hybridoma cell cultivation. Filled symbols show the tivity was high as expected, but the stability
results from a RIA, and open symbols are the po- was far better than what had been achieved be-
tentiometric readings from the streaming potential. fore. It still remains a task for chemists to dis-
close the mechanism of immobilization before
one can decide whether this is a general tech-
sensitivity varies with the charge density of the nique for stabilizing receptors.
compound bound to the support, but in gener-
al terms it is possible to achieve signals sepa-
rated from background noise at concentration 6.3 Analytical Performance
levels above lo-’ mol/L. It is still too early to
state what the potential of this technique will of Affinity Biosensors
be. At present it is possible to monitor the con-
centration of monoclonal antibodies during From the examples discussed above it is ob-
cultivation (Fig. 14). vious that immunoassays can be designed to be
very sensitive, fast, and even operative in a
continuous manner. At the present state of de-
6.1.2.2 Reflectometry velopment these characteristics cannot all be
obtained using one set of operating conditions.
When plane-polarized light is reflected from One must, therefore, make compromises and
a solid surface there is a minimum reflectance optimize the assay according to set require-
at a certain angle of incidence, the pseudo- ments. However, there is still a very small data
Brewster angle. When organic material is ad- base available from biotechnological processes
sorbed onto the surface an increase in reflec- regarding variation of concentrations of ma-
tance is observed. The more material is ad- cromolecules with time.
sorbed, the larger the change. High sensitivity The concentration range can easily be de-
is obtained in immunoassays because of the fined by carrying out conventional off-line
large difference in refractive index between sil- analyses. In most cases if not all, the concen-
icon and organic material. trations of the target molecule will be far
7 Biosensors in Extreme Environments 97
above the detection limits of equilibrium bind- processes. Is this science fiction, or is there
ing assays. This opens the possibility for ap- a chance that the dream will become a real-
plying faster, non-equilibrium binding assays ity?
or looking for continuous measurements. It is, Bioorganic synthesis, i. e., the technique of
however, still far too early to formulate a gen- using enzymes in organic solvents for organic
eral strategy when designing and utilizing im- synthesis, has undergone a development that
munobased biosensors. few even dreamt of ten years ago. Today it is
fully possible to operate with enzymes in pure
organic solvents, e.g., in chloroform, toluene,
or hexane (TRAMPERet al., 1985; LAANEet
al., 1987). There is a requirement for an ex-
7 Biosensors tremely small amount of water for hydration
of essential regions on the protein molecule. In
in Extreme Environments several studies it has been shown that the en-
zymes under these very non-physiological con-
ditions are more stable than in aqueous solu-
Efforts have mainly been directed towards tion or immobilized in contact with an aque-
applying biosensors in aqueous environments ous environment (KLIBANOV,1986; RESLOW
under conditions a biochemist would call gen- et al., 1987).
tle. However, there are many other needs for Very few examples of successful utilization
bioanalysis that raise strong challenges to bio- of the enzyme-biosensor concept with organic
sensor technology. solvents have as yet been reported. In two of
these, oxidases (mainly cholesterol and phenol
oxidase) have been studied, since both the sub-
7.1 Biosensors in Biological Liquids strate and oxygen have higher solubilities in
the organic phase than in water (DANIELSSON
A challenge encountered by every biosensor et al., 1989; HALL et al., 1988; FLYGARE et
in fermentation broth or body fluids is fouling al., 1988; KAZANDIJAN et al., 1986). We have
by protein and, eventually, by cells. Such sec- tried to monitor, e.g., ester synthesis in or-
ondary layers may change the response charac- ganic solvents using a potentiometric elec-
teristics of the biosensor and may lead to total- trode. It was clearly shown that the signals ob-
ly erroneous signals. tained correlated very well with the substrate
Furthermore, for in vivo measurements on concentrations. When the enzyme was omitted
patients such protein-depositing activities may or any other factor was changed, no enzymatic
activate the coagulation system. It is therefore activity was noted and no potentiometric sig-
of utmost importance for future medical appli- nal was registered (MIYABAYASHI et al., 1989).
cations as well as for use in bioreactors that The performance of the enzyme electrode cor-
biosensors be made with biocompatible sur- related very well with earlier observations on
faces. the dependence on the water activity in the me-
dium (RESLOWet al., 1987).
From the studies carried out so far it is ab-
7.2 Biosensors in Organic Solvents solutely clear that one can operate with en-
zyme-based sensors in organic solvents and
In organic chemistry it would be desirable generate signals that reflect the process in the
in many cases to follow a process more accu- medium. However, to apply such sensors to
rately. Today, samples are removed from the traditional organic processes may require
reaction vessel and analyzed outside, either us- much time and effort. Temperature has been
ing gas chromatography or HPLC. However, carefully controlled and many of the aggres-
if it were possible to apply the biosensor con- sive reagents often used in organic synthetic
cept under these conditions, one could exploit work are avoided. The basic strategy may be
the unique selectivity and regioselectivity of to draw off a small stream and use that for
enzymes for monitoring various synthetic monitoring. Only in the longer perspective
98 3 Biosensors
might it be realistic to insert the sensors direct- multifunctional sensors are to be developed.
ly into the reactors. Much work is already under way in this direc-
tion. Using such small sensors it will be possi-
ble to include sensors for several essential com-
7.3 Biosensors Operated in Air pounds in the same sensor so that one expo-
sure of the sensing device can provide informa-
It is a great challenge to analyze air using tion on many different essential components.
biosensors. One early report in this field was To reach this goal, the integration of
the nerve gas sensor developed for the US microelectronics and biosensor work needs to
army (GOODSONand JACOBS,1976). How- advance further. Current reports in this field
ever, in this application the air was passed are very promising, which is why it seems real-
through a scrubber and any nerve gas present istic to forecast that this development will take
was thus transferred into aqueous solution place in the near future.
(BAUMANet al., 1965). Since it has been Protein stability is one of the bottlenecks
shown that one can operate with enzymes with when constructing stable biosensors. By apply-
a very thin hydration layer in organic solvents, ing protein engineering it is possible, e.g., to
it seems reasonable to assume that the same eliminate sensitive SH groups or to establish
conditions should hold for measurements in S-S bonds (PERRYand WETZEL, 1984). It
air. One area of application is the analysis of may thus be possible to tailor the enzymes to
various volatile compounds such as nerve the applications. Another alternative is to ap-
gases, small organic alcohols, and aldehydes. ply more stable enzymes from thermophilic or-
ganisms when constructing the sensor. The in-
troduction of doped electrodes and the emerg-
ing awareness of how to integrate electrically
conducting polymers show promise for the fu-
8 Future Developments ture.
Some recent developments on applications
Biosensor research has led to the develop- of biosensors in extreme environments are
ment of a broad spectrum of biosensors, at mentioned in Section 7. These developments
least as judged by the academic literature. will continue, and further developments, espe-
However, commercialization has not been as cially in the area of bioorganic sensors, can be
successful as expected. Though many types of expected.
biosensors have been described, there are few
that work reliably under realistic conditions
outside the well-controlled laboratory environ-
ment. This lack of commercial success has led
to the situation that only the true enthusiasts 9 References
use biosensors for process control. The more
common applications will have to wait until re- ADLERCREUTZ, P., MATTIASSON, B. (1982), Oxy-
liable and robust instruments are available. gen supply to immobilized biocatalysts. A model
Therefore, in the future much work will be study, Acta Chem. Scand. B36, 651-653.
done to make biosensors more adapted to the AIZAWA,M., MORIOKA,A., SUZUKI,S., NAGAMU-
field so that at least some of this enormous po- RA, Y. (1979), Enzyme immunosensor. 111. Am-
tential can be tapped. perometric determination of human chorionic
Implantable biosensors have great potential gonadotropin by membrane bound antibody,
importance in medicine and in fermentation Anal. Biochem. 94, 22-28.
ALBERY,W. J., BARLETT,P. N., CRASTON,D. H.
control. By tailoring the surface properties of (1985), Amperometric enzyme electrodes. Part
the sensors so that no growth or adsorption 11. Conducting organic salts as electrode materi-
takes place, it will be possible to apply such als for the oxidation of glucose oxidase, J. Elec-
sensors in many new environments. troanal. Chem. 194, 223-235.
Miniaturization of sensors may be attrac- ALWIS, DE, W. U.,WILSON,G. S. (1985), Rapid
tive, especially for implantable sensors, and if sub-picomole electrochemical enzyme immuno-
References 99
sensor for immunoglobulins, Anal. Chem. 57, CHEN,B., KUAN,S., GUILBAULT, G. G. (1980), A
2754-2756. creatinine specific enzyme electrode, Anal. Lett.
APPELQVIST,R., MARKO-VARGA,G., GORTON, 13, 1607-1624.
L., TORSTENSSON, A., JOHANSSON,G. (1985), CLARK,L. C . , JR., LYONS,C. (1962), Ann. N. Y.
Enzymatic determination of glucose in a flow Acad. Sci. 102, 29.
system by catalytic oxidation of the nicotinamide CLARK,L. C., NOYES,L. K., SPOKANE, R. B., Su-
coenzyme at a modified electrode, Anal. Chim. DAN, R., MILLER,M. L. (1988), Long-term im-
Acta 169, 237-247. plantation of voltammetric oxidase/peroxide glu-
APPELQVIST,R . , MARKO-VARGA,G., GORTON, cose sensor in the rat peritoneum, Methods Enzy-
L., JOHANSSON, G. (1986), Amperometric deter- mol. 137, 68-89.
mination of starch, glycogen and amylopectin in CLELAND, N., ENFORS,S.-0. (1984), Monitoring of
a FIA system using enzyme reactors and a modi- glucose consumption in an Escherichia coli culti-
fied electrode as detecting device, Proc. 2nd Znt. vation with an enzyme electrode, Anal. Chim.
Meeting Chemical Sensors, Bordeaux 1986, G- Acta 163, 281-285.
25, 603. COLLINS,S., JANATA,J. (1982), A critical evalua-
ARNOLD,M. A., RECHNITZ,G. A. (1980), Com- tion of the mechanism of potential response of
parison of bacterial, mitochondrial, tissue and antigen polymer membranes to the corresponding
enzyme biocatalysts for glutamine selective mem- antisera, Anal. Chim. Acta 136, 93-99.
brane electrodes, Anal. Chem. 52, 1170-1 174. CORCORAN, C. A., RECHNITZ,G. A. (1985), Cell
BARLETT,P. N., WHITAKER, R. G., GREEN,M. J., based biosensors, Trends Biotechnol. 3, 92-96.
FREW,J. (1987), Covalent binding of electron re- CUYPERS,P. A,, HERMENS,W. T., HEMKER,H.
lays to glucose oxidase, J. Chem. SOC. Chem. C. (1978), Elipsometry as a tool to study protein
Commun., 1603-1604. films at liquid-solid interfaces, Anal. Biochem.
BAUMAN,G., GOODSON,L., GUILBAULT, G. G., 84, 56-57.
KRAMER, D. (1969, Preparation of immobilized DANIELSSON, B., GADD, K., MATTIASSON,B.,
cholinesterase for use in analytical chemistry, MOSBACH,K. (1977), Enzyme thermistor deter-
Anal. Chem. 37, 1378-1381. mination of glucose in serum using immobilized
BORREBAECK, C., MATTIASSON, B (1980), Lectin- glucose oxidase, Clin. Chim. Acta 81, 163-175.
carbohydrate interactions studied by a competi- DANIELSSON, B., FLYGARE,L., VELEV,T. (1989),
tive enzyme inhibition assay, Anal. Biochem. Biothermal analysis performed in organic sol-
107, 446-450. vents, Anal. Lett. 22 (6), 1417-1428.
BORREBAECK, C., B~RJESSON, J., MATTIASSON, B. DANIELSSON, B., MATTIASSON, B., KARLSSON, R.,
(1978), Thermometric enzyme linked immunosor- WINQVIST,F. (1979), Use of enzyme thermistor
bent assay in continuous flow system: optimiza- in continuous measurements and enzyme reactor
tion and evaluation using human serum albumin control, Biotechnol. Bioeng. 21, 1749-1766.
as a model system, Clin. Chim. Acta 86, 267-278. DANIELSSON, B., MATTIASSON, B., MOSBACH,K.
BOUIN,J. C., ATALLAH,M. T., HULTIN,0. H. (1981a), Enzyme thermistor devices and their
(1976a), The glucose oxidase-catalase system, analytical applications, Appl. Biochem. Bioeng.
Methods Enzymol. 44, 478-488. 3, 97-143.
BOUIN,J. C., ATALLAH,M. T., HULTIN,0. H. DANIELSSON, B., RIEKE, E., MATTIASSON,B.,
(1976b), Parameters in the construction of an im- WINQUIST,F., MOSBACH,K. (1981b), Determi-
mobilized dual enzyme catalyst, Biotechnol. nation by the enzyme thermistor of cellobiose
Bioeng. 18, 179-187. formed on degradation of cellulose, Appl. Bio-
BROOKS,S. L., ASHBY,R. E., TURNER, A. P. F., chem. Biotechnol, 6, 207-222.
CALDER,M. R., CLARKE,D. J. (1987), Develop- D’COSTA,E. J., HIGGINS,I. J., TURNER, A. P. F.
ment of an on-line glucose sensor for fermenta- (1986), Quinoprotein glucose dehydrogenase and
tion monitoring. Biosensors 3, 45-56. its application in an amperometric glucose sen-
BULOW,L., MOSBACH,K. (1982), Ligation of re- sor, Biosensors 2, 71-87.
striction endonuclease-generated DNA fragments DEGANI,Y., HELLER,A. (1987), Direct electrical
using immobilized T4 DNA ligase, Biochem. Bio- communication between chemically modified en-
phys. Res. Commun. 107, 458-465. zymes and metal electrodes. 1. Electron transfer
CASS,A. E. G., DAVIS,G., FRANCIS, G. D., HILL, from glucose oxidase to metal electrodes via elec-
H. A. O., ASTON,W. J., HIGGINS,I, J., PLOT- tron relays, bound covalently to the enzyme, J.
KIN, E. V., SCOTT,L. D. L., TURNER, A. F. P. Phys. Chem. 91, 1285-1289.
(1984), Ferrocene mediated enzyme electrode for DEGANI,Y., HELLER,A. (1988), Direct electrical
amperometric determination of glucose, Anal. communication between chemically modified en-
Chem. 56, 667-671. zymes and metal electrodes. 2. Methods for
100 3 Biosensors
bonding electron-transfer relays to glucose oxi- enzyme electrode for the substrate urea, J. A m .
dase and D-amino-acid oxidase, J. A m . Chem. Chem. SOC.92, 2533-2538.
SOC.110, 2615-2620. GUILBAULT, G. G., SHU,F. R. (1971), An electrode
DISSING,U., LING, T. G. I., MATTIASSON,B. for the determination of glutamine, Anal. Chim.
(1984), Monitoring of methanogenic processes Acta 56, 333-338.
with an immobilized mixed culture in combina- HALL, G. F., BEST, D. J., TURNER,A . P. F.
tion with a gas-flow meter, Anal. Chim. Acta (1988), Amperometric enzyme electrode for the
163, 127-1337 determination of phenols in chloroform, Enzyme
DOMININGUEZ.E., MARKO-VARGA,G., HAHN- Microb. Technol. 10, 543-546.
HAGERDAL,B., GORTON,L. (1988), A flow in- HIKUMA,M., SUZUKI,H., YASUDA,T., KARUBE,
jection analysis system for the amperometric de- I., SUZUKI,S. (1979), Amperometric estimation
termination of xylose and xylulose using coim- of BOD by using living immobilized yeast, Eur.
mobilized enzymes and a modified electrode, J. Appl. Microbiol. Biotechnol. 8, 289-297.
Anal. Chim. Acta 213, 139-150. HOLST, O., HAKANSON,H., MIYABAYASHI, A.,
DOYLE,M. J., HALSALL,H. B., HEINEMAN, W. R. MATTIASSON, B. (1988), Monitoring of glucose
(1984), Enzyme-linked immunoadsorbent assay in fermentation processes using a commercial
with electrochemical detection of a,-acid glyco- glucose analyser, Appl. Microbiol. Biotechnol.
protein, Anal. Chem. 56, 2318-2322. 28, 32-36.
ENFORS,S. 0. (1981), Oxygen stabilized enzyme HORISBERGER, M. (1980), An application of elip-
electrode for glucose analysis in fermentation sometry. Assessment of polysaccharide and gly-
broths, Enzyme Microb. Technol. 3, 29-32. coprotein interactions with lectin at a liquid/solid
ENFORS,S. O., NILSSON,H. (1979), Design and re- interface, Biochim. Biophys. Acta 632, 298-309.
sponse characteristics of an enzyme electrode for HUBBARD,A. L., GOHN, Z. A. (1976), in: Bio-
measurement of penicillin in fermentation broth, chemical Analysis of Membrane (MADDE,A. H.,
Enzyme Microb. Technol 1, 260-264. Ed.), pp. 427-501. New York: Wiley.
FATT, I. (1976), Polarographic Oxygen Sensor. Its HAKANSON,H. (1988), Portable continuous blood
Theory of Operation and its Application in Bio- glucose analyzer, Methods Enzymol. 137, 319-
logy, Medicine and Technology. Cleveland, 326.
Ohio: CRC Press. HAKANSON,H., HOLST, O., WEHTJE,E., MAT-
FLYGARE,L., DANIELSSON,B., MOSBACH,K. TIASSON, B. (1990), Monitoring of glucose and
(1988), Abstr. B 29, 8th International Biotechno- lactic acid in cultivations of lactic acid bacteria
logy Symposium, Paris, July 1988. using an automated biosensor-based analyzer, in
FOULDS,C. N., LOWE, C. R. (1986), Enzyme en- preparation.
trapment in electrically conducting polymers. Im- IUB (1979), Enzyme Nomenclature 1978. New
mobilization of glucose oxidase in polypyrrole York: Academic Press.
and its application in amperometric glucose sen- IWAMOTO,G. K., VAN WAGENEN,R. A., AN-
sors, J. Chem. SOC.Faraday Trans. Z. 82, 1259- DRADE,J. D. (1982), Insulin adsorption: intrinsic
1264. tyrosine interfacial fluorescence, J. Colloid Znter-
FOULDS,C. N., LOWE, C. R. (1988), Immobiliza- face Sci. 86, 581-585.
tion of glucose oxidase in ferrocene-modified JOHANSSON, G., EDSTROM,K., OGREN,L. (1976),
pyrrole polymers, Anal. Chem. 60, 2473-2478. An enzyme reactor electrode for determination of
FULTON,S. P., COONEY,C. L., WEAVER,J. C. amino acids, Anal. Chim. Acta 85, 55-60.
(1980), Thermal enzyme probe with differential JULLIARD, J. H., GOLDINOT,G., GAUTHERON, D.
temperature measurements in a laminar flow- C. (1971), Some modifications of the kinetic
through cell, Anal. Chem. 52, 505-508. properties of bovine liver glutamate dehydrogen-
GLAD, C., SJODIN, K., MATTIASSON,B. (1986), ase (NAD(P)) covalently bound to a solid matrix
Streaming potential - a general affinity sensor, of collagen, FEBS Lett. 14, 185-188.
Biosensors 2, 89-100. KARUBE,I., MATSUNAGA, T., MITSUDA,S . , Suzu-
GOODSON,L. H., JACOBS,W. B. (1976), Monitor- KI, S. (1977a), Microbial electrode BOD sensor,
ing of air and water for enzyme inhibitors, Meth- Biotechnol. Bioeng. 19, 1535-1547.
ods Enzymol. 44, 647-658. KARUBE,I., MATSUNAGA, T., SUZUKI,S. (1977b),
GREEN,N. M. (1963), Avidin. I. The use of I4C bio- A new microbial electrode for BOD estimation,
tin for kinetic studies and for assay, Biochem. J. J. Solid-Phase Biochem. 2, 97-104.
89, 585-591. KARUBE,I., MITSUDA,S., MATSUNAGA, T., Suzu-
GUILBAULT, G. G. (1984), Analytical Uses of Zm- KI, S., KADA,T. (1977c), A rapid method for the
mobilized Enzymes. New York: Marcel Dekker. estimation of BOD using immobilized microbial
GUILBAULT, G. G., MONTALVO,J. G. (1970), An cells, J. Ferment. Technol. 55, 243-248.
References 101
MESSING, R. A. (1974), Simultaneously immobil- protein toward thermal inactivation, Science 226,
ized glucose oxidase and catalase in controlled- 555-557.
pore titania, Biotechnol. Bioeng. 16, 897-908. PETERSON,J. I., GOLDSTEIN,S. R., FITZGERALD,
MEYERHOF,M. E., RECHNITZ,G. A. (1979), Mi- R. V . , BUCKHOLD, D. K. (1980), Fiber optic pH-
crosomal thyroxine measurements with iodide se- probe for physiological use, Anal. Chem. 52,
lective membrane electrode, Anal. Lett. 12, 864-869.
1336-1346. P I L L O T OR.,
~ , MASCINI,M., CASELLA,I. G., FES-
MIYABAYASHI, A., MATTIASSON,B. (1990), A dual TA, M. R., BOTTARI,E. (1987), Lactose determi-
streaming potential device used as an affinity sen- nation in raw milk with a two-enzyme based elec-
sor for monitoring hybridoma cell cultivations, trochemical sensor, Anal. Lett. 20, 1803-1814.
Anal. Biochem. 184, 165-171. RESLOW,M., ADLERCREUTZ, P., MATTIASSON, B.
MIYABAYASHI, A., DANIELSSON, B., MATTIASSON, (1987), Organic solvents for bioorganic synthesis.
B. (1987), Development of a flow-cell system I. Optimization of parameters for a chymotryp-
with dual fuel-cell electrodes for continuous sin catalyzed process, Appl. Microbiol. Biotech-
monitoring of microbial populations, Biotechnol. nol. 26, 1-8.
Tech. 1, 219-224. ROEDERER,J. E., BASTIAANS, G. J . (1983), Micro-
MIYABAYASHI, A., RESLOW,M., ADLERCREUTZ, gravity immunoassay with piezoelectric crystals,
P., MATTIASSON,B. (1989), A potentiometric Anal. Chem. 55, 2333-2336.
enzyme electrode for monitoring in organic sol- RUSLING,J. F., LUTTRELL,G. H., CULLEN,L. F.,
vents, Anal. Chim. Acta 219, 27-36. PAPARIELLO, G. J. (1976), Immobilized enzyme-
MOSBACH,K . (Ed.) (1976), Methods Enzymol. 44. based flowing stream analyzer for measurement
New York: Academic Press. of penicillin in fermentation broths, Anal. Chem.
MOSBACH,K. (Ed.) (1988), Methods Enzymol. 135- 48, 1211-1215.
137. New York: Academic Press. SCHELLER,F., SIEGBAHN,N., DANIELSSON,B.,
NEISSEL, 0. M., TEMPEST,D. W., POSTMA,P. MOSBACH,K. (1985), High sensitivity enzyme
W., DUINE,J. A., JZN, J. F. (1983), Glucose me- thermistor determination of L-lactate by sub-
tabolism by K -limited Klebsiella aerogenes: evi-
+ strate recycling, Anal. Chem. 57, 1740-1743.
dence for the involvement of a quinoprotein glu- SCHELTER-GRAF, A., SCHMIDT,H.-L., HUCK,H.
cose dehydrogenase, FEMS Microbiol. Lett. 20, (1984), Determination of substrates of dehydro-
35-39. genases in biological material in flow-injection
NGO, T. T. (Ed.) (1987), Electrochemical Sensors in systems with electrostatic NADH-regeneration,
Immunological Analysis. New York: Plenum Anal. Chim. Acta 163, 299-303.
Press. SCHUBERT,F., SCHELLER,F. W. (1988), Organelle
NILSSON, H., AKERLUND,A,-C., MOSBACH, K. electrodes, Methods Enzymol. 137, 152-160.
(1973), Determination of glucose, urea and peni- SCHULTZ,J. S., SIMS, G. (1979), Affinity sensors
cillin using an enzyme-pH-electrode, Biochim. for individual metabolites, Biotechnol, Bioeng.
Biophys. Acta 320, 529-534. Symp. 9, 65-71.
OLSSON,B., STALBOM,B., JOHANSSON, G. (1986), SCHULTZ,J. S., MANSOURI,S., GOLDSTEIN,I. J .
Determination of sucrose in the presence of glu- (1982), Affinity sensor: a new technique for de-
cose in a flow-injection system with immobilized veloping implantable sensors for glucose and oth-
multi-enzyme reactors, Anal. Chim. Acta 179, er metabolites, Diabetes Care 5, 245-253.
203-208. S C H ~ G E R K.
L , (1988), On-line analysis and control
O'SHANNESSY, D. J., HOFFMAN,W. L. (1987), Im- of production of antibiotics, Anal. Chim. Acta
mobilization of glycoproteins to hydrazide con- 213, 1-9.
taining supports, Hoppe-Seyler's 2. Biol. Chem. SEITZ,W. R. (1987), Optical sensors based on im-
368, 767. mobilized reagents, in: Biosensors. Fundamentals
PANDEY, P. C. (1988), A new conducting polymer- and Applications, pp. 599-616 (TURNER,A. F.
coated glucose sensor, J. Chem. SOC. Faraday P., KARUBE,I., WILSON,G. S., Eds.). Oxford:
Trans. 184, 2259-2265. Oxford University Press.
PARKER,C. W. (1976), Radioimmunoassay of Bio- SHINOHARA,H., CHIBA,T., AIZAWA,M. (1988),
logically Active Compounds. Englewood Cliffs, Enzyme microsensor for glucose with an electro-
New Jersey: Prentice Hall. chemically synthesized enzyme polyaniline film,
PENNINGTON, S. N. (1976), A small volume micro- Sens. Actuators 13, 79-86.
calorimeter for analytical determinations, Anal. SUZUKI,H., TAMIYA,E., KARUBE,I. (1988), Fabri-
Biochem. 72, 230-237. cation of an oxygen electrode using semiconduc-
PERRY,J. L., WETZEL,R. (1984), Disulfide bond tor technology, Anal. Chem. 66, 1078-1080.
engineering into T4 lysozyme. Stabilization of the TAYLOR,R. F., MARENCHIC,I. G., COOK, E. J .
References 103
ANDREAS
LUBBERT
Hannover, Federal Republic of Germany
1 Introduction 109
2 Characteristic Properties 110
2.1 Mass Transfer Across Gas/Liquid Interfacial Areas 110
2.1.1 Mechanism 110
2.1.2 Characteristics of Mass Transfer 111
2.2 Bulk Mixing Properties of Bioreactors 114
2.2.1 Relevant Dispersion Mechanisms 114
2.2.1.1 Convective Large-Scale Flows 114
2.2.1.2 Statistical Turbulence 114
2.2.1.3 Molecular Diffusion 115
2.2.1.4 Viscous Shear 115
2.2.1.5 Dispersion by Rising Bubbles 116
2.2.1.6 Interaction of the Different Mechanisms 116
2.2.2 Characteristics of Mixing in Bioreactors 116
2.2.2.1 Deterministic Circulatory Flow Modes 117
2.2.2.2 Random Dispersive Motions 118
2.2.2.3 Balancing Random and Deterministic Motions 118
2.2.2.4 Characteristics of Heat Transfer in Bioreactors 119
2.3 Conclusions about Practice of Characterization 119
3 Global Measuring Techniques 119
3.1 Integral Gas Holdup 119
3.1.1 Volume Expansion Method 119
3.1.2 Pressure Differences 120
3.2 Integral Specific Interfacial Areas 120
3.3 Residence Time Distributions and Related Measurements 121
3.3.1 Tracer Techniques 121
3.3.1.1 General Methodology 121
3.3.1.2 Gas Residence Times 121
3.3.1.3 Liquid Residence Times 123
3.3.1.4 Circulation and Mixing Times 123
108 4 Characterization of Bioreactors
all cases, appropriate measuring techniques are tion. It should be mentioned, however, that
indispensible to base these activities on reliable the different aspects discussed cannot be put
data. Therefore, the main interest of this chap- into a universal order of importance, because
ter must be directed toward methods which the order of importance depends upon the par-
can actually be used to obtain data from ac- ticular application.
tually existing biological systems.
-
2.0
I
A
I \
gas phase. Moreover, the amount of the gas
circulated once around the loop can be calcu-
lated from the relationship of the integrals of
Gas load 401Nm3/hl the resolved peaks. The circulatory flow of
bubbles in loop reactors leads to a feedback of
bubbles into the gas input and bubble forma-
-
05 tion region, where they may coalesce with
00
38s
I I I I I
Timelsl AXIAI-RADIAL h
4
h
concentration and, hence, reducing the driving \NGENTIAL
force for mass transfer.
This is a drawback not only of loop reac-
tors. Even in stirred tank reactors a considera-
ble part of the gas is recirculated back into the
stirrer region, where it is caught by the venti-
lated cavities formed behind the stirrer blades
and, hence, in all probability coalesces. It is
immediately clear that the amount of gas recir-
culated drastically depends on the bubble size
distribution and, hence, also on the coales-
cence/redispersion behavior in the bulk of the
dispersion. Smaller bubbles follow the liquid
phase motion much better and contribute to
backmixing much more. In this way, small
bubbles, which do not tend to coalesce in the
bulk phase, paradoxically contribute more to
backmixing than the coalescing ones.
Experimental evidence on such bubble recir-
culation can be obtained from the bubble ve-
locity field within a common stirred tank reac-
tor, as shown in Fig. 2. This figure clearly de-
monstrates that the bubble motion is compli-
cated and is completely different from the flow
structure of the continuous liquid flow. As op-
posed to the gross liquid flow structure, which
is essentially independent of the rheological
properties of the media, Fig. 3 shows that a
change in the viscous properties of the contin-
uous liquid phase leads to completely different
0
I
bubble paths. Since such bubble motion maps 300
cannot be calculated, they must be obtained Radius(mm1
experimentally. Such detailed data are also Fig. 3. Velocity profile of the bubbles within a stir-
necessary to put forward the theoretical inves- red tank reactor operated with a Rushton turbine
tigation of bioreactor hydrodynamics (TRA- (BRORINGet al., 1990). As compared to Fig. 2 the
GARDH,1988). measurement was performed in a model fermenta-
This recirculation leads to another essential tion medium. One percent CMC in water was cho-
hydrodynamic effect. It is known from various sen. The bubble flow profiles are then totally differ-
experiments that the power absorbed by a dis- ent from the results shown in Fig. 2.
persion at constant stirrer speed decreases with
the gas flow rate. The main reason for this
power loss is thought to be the formation of (MIDDLETON, 1985). This is another argument
previously mentioned ventilated gas cavities for measuring and controlling bubble sizes and
behind the stirrer blades, which lower the hy- velocities.
drodynamic resistance (drag) of the impeller. Thus, the coalescence/redispersion process
The gas flow through these cavities determines of gas bubbles is a general problem of highest
their size. This gas flow cannot be calculated, importance, not only for mass transfer, but
since one does not know the amount of gas also for most other aspects of multiphase flow
which is recirculated into the stirrer. It is as- hydrodynamics. Coalescence of bubbles is
sumed that about the same amount of recircu- greatly reduced in solutions compared to pure
lated gas as of freshly fed gas is recirculated solvents. Thus, many important parameters
114 4 Characterization of Bioreactors
determining the fluid dynamics and the mass 2.2.1 Relevant Dispersion
transport, e.g., bubble size distribution, inter-
facial area, and holdup, are changing during
Mechanisms
the fermentation. Immediate consequences are
grossly different residence times and, especial- The problem to be solved is analogous to
ly in airlift reactors, different liquid circula- supplying the population of a country with a
tion times. Even a minor amount of a surfac- highly perishable food, e.g., fruit that must be
tant component (ppm amount) can drastically imported. Several transport systems are re-
change the coalescence and redispersion be- quired for an efficient distribution. First of all,
havior of the contactor (BUCKLAND et al., a fast transport of larger amounts of this food
1988). This will result in extreme changes of covering all states of the nation is necessary.
k,a. As a consequence, such systems cannot be This may be done by big trucks covering long
modelled from first principles. It is necessary distances on interstate highways across the
to rely on experimental determination of the whole country. More locally, regional whole-
mentioned quantities. For the same reasons, sale dealers transport the goods on regional
measurements must be performed in exactly roads in order to distribute them within states
the same material system as the original, pref- to the stores in different cities and villages.
erably under the same fluid dynamical condi- These stores must present the food to as many
tions. This fact has not been taken due notice families as possible, who buy it and distribute
of in the past. It is not possible to predict k,a it directly to the smallest units, the family
for other systems with confidence (MIDDLE- members, where it is consumed.
TON, 1985). The coalescence behavior also in-
fluences the power uptake of the liquid from a
stirrer, because this influences the formation 2.2.1.1 Convective Large-Scale
and the stability of the cavities behind the stir-
rer blades. The relevant mechanisms and the Flows
possibilities of influencing them are not com-
pletely understood at this date. Geometrically, any material that is initially
fed into a bioreactor at one place appears as a
separate volume inside the reactor. The aim of
the flow inside the reactor is to break this vol-
2.2 Bulk Mixing Properties ume element into parts and to distribute them
of Bioreactors homogeneously over the whole reactor. Analo-
gous to the trucks driving on the highways,
Another essential task of bioreactors is the large-scale convectional flows are guided by
distribution of substrates or other components appropriate devices in order to obtain a fast
which are introduced at one place in the reac- but coarse distribution of matter all over the
tor or, more generally, the reduction of the bioreactor. The first aim of bioreactor con-
concentration and temperature gradients in a struction is thus to guide these deterministic
sufficiently fast and efficient way. The success flows so that they reach every corner of the
of the different mixing mechanisms, which bioreactor within a short time.
take part in this homogenization, depends on
the spatial resolution of interest. For biochem-
ical reactions, homogeneities, down to the mo- 2.2.1.2 Statistical Turbulence
lecular scale, are required.
Wholesale business must correspond to the
induction of several additional random disper-
sive flows. In practice, one tries to apply tur-
bulence, which is known to be the most effi-
cient regional dispersion mechanism. The ad-
vantage of turbulence in aiding mixing can
best be understood in terms of the statistical
Characteristic Properties 1 15
l=(v3/&)1’4 (5)
can provide a rough idea as to how far down Apart from single-phase tube flow, howev-
in size turbulence can disperse liquid fluid ele- er, it is not always simple to find appropriate
ments (microscale size I , scale of the smallest characteristic dimensions d and velocities u to
eddies) in a fluid of viscosity v, if an energy calculate a meaningful Reynolds number. But
dissipation density E is assumed. This estima- it can be seen immediately from this simple
tion is very rough, since, as is well-known, the characteristic number that the different mech-
local E can vary widely across the reactor. It anisms are not independent of each other. Fast
can be more than 100 times larger around the global transport supports local mixing mecha-
impeller than at places far away from the stir- nisms, since it directly increases the Reynolds
rer. number by the mean velocity u. It can also be
The Kolmogoroff theory, which has been seen that larger reactors, driven at the same
developed for single-phase flows of large-scale mean velocity u, are more susceptible to turbu-
flow systems, cannot be applied quantitatively lence, since the characteristic dimensions be-
in chemical reactors (MAHOUASTet al., 1989). come larger.
Hence, it is not surprising that it is difficult to
apply in multiphase reactors (LUBBERT and
LARSON, 1990). Qualitatively, however, it is 2.2.1.4 Viscous Shear
perceptible that more energy in random mo-
tion will lead to a smaller scale of the segre- In highly viscous systems, the Reynolds
gated fluid elements, whether the flow is tur- number, necessary for turbulence, may not be
bulent or not. obtainable. Then, larger transport cross-sec-
116 4 Characterization of Bioreactors
tions for molecular diffusion must be obtained compete with ordered, forced convective flows
with streaking fluid elements by applying ap- in transporting matter over distances of the or-
propriate shear stresses. This can be done, der of the reactor dimension. On the other end
e.g., by some special agitation devices which of the spectrum, at the microscale, molecular
deform the fluid elements so that their interfa- diffusion, which is neglible at large scales, is
cial areas become larger until the streaks even- the only effective mechanism.
tually break into small parts leading to even Since, ultimately, molecular diffusion is in-
larger transport cross-sections for molecular dispensible, one must concentrate on maximiz-
diffusion (GODFREY,1985; EDWARDS,1985). ing the overall rate of diffusion in every possi-
Optimally, the striation thickness becomes so ble way. In a somewhat oversimplified formu-
slight that particles can penetrate by diffusion lation, one can state that all other dispersion
within a short time. Concentration differences mechanisms must be directed to support mo-
can be kept relatively high by folding such lecular diffusion. At any rate, there is a syner-
streaks, thereby building layers of different gism in the interplay of the different mecha-
concentrations. nisms, which must be exploited by bioreactor
engineering.
In sum, however, it is not necessarily the
2.2.1.5 Dispersion diffusion process which is limiting. As NAU-
MAN (1989) showed, bulk circulatory flow may
by Rising Bubbles be the rate-limiting transport process in batch
reactors. This seems to be confirmed by nu-
Since we are dealing with multiphase flows merous experimental results obtained from
in bioreactors, there are additional mixing larger bioreactors, e.g., from the results of
mechanisms which result from the interplay of HANSFOLDand HUMPHREY(1966) to those of
the different phases in the reactor. The most BUCKLANDet al. (1988). Thus, detailed data
interesting effects are due to the displacements on mixing and circulation times in real fermen-
necessary for a bubble to rise in a liquid phase tation systems are a prerequisite to gaining
and those due to transport within the wakes of more insight into the rate-limiting processes.
particles (WEBERand BHAGA, 1982; RIETE- This is also necessary to choose the optimal ag-
MA, 1982; WASOWSKI and BLASS, 1989; LOB- itator system for a specific application.
BERT and LARSON,1990). This effect is signif-
icant if the rising or settling velocities of the
particles relative to the continuous phase are 2.2.2 Characteristics
sufficiently high. Mixing occurs, since a rising
bubble pushes away the liquid ahead in a n ir- of Mixing in Bioreactors
reversible way so that the fluid elements be-
hind it are rearranged. Furthermore, liquid is In order to bring the different dispersion
carried along inside the bubble wakes. This mechanisms effectively into play, the proper
leads to a continuous exchange of matter with flows should be induced in bioreactors by
the surrounding bulk phase. means of an appropriate input of mechanical
energy. In stirred tank reactors this is done by
using an optimal impeller system, in airlift
2.2.1.6 Interaction reactors this is obtained by introducing the gas
phase appropriately. It has been known for a
of the Different Mechanisms long time that in industrially important aero-
bic fermentations, critical performance param-
Generally, as in the mentioned analogy, all eters, e.g., the oxygen uptake rate by the mi-
fluid dispersion mechanisms are active simul- croorganisms, are highly dependent on the
taneously. Their relative efficiencies, however, physical parameters, both operational and
differ widely at different scales. While turbu- geometrical (WANG and FEWKES, 1977).
lence is far the most effective means to homo- FIELDSand SLATER(1984) showed that the lo-
genize fluids at intermediate scales, it cannot cal liquid mixing behavior in airlift loop reac-
Characteristic Properties 117
tors affected the respiration of the microor- velocity field. However, measurements for its
ganisms. comprehensive construction would be virtually
impossible. Therefore, a model assumption is
necessary. This is examined by measuring pro-
2.2.2.1 Deterministic Circulatory files of the flow along critical cross-sections or
by point measurements. Detailed information
Flow Modes on the velocity structure is needed to detect
dead volumes which are insufficiently supplied
In order to achieve fast global transport, by the flow. In highly viscous media, it is, e.g.,
some kind of forced convective flow is induced possible to detect the flow volumes which are
in all bioreactors. In most cases, the resulting not agitated by simply looking for the places
large-scale circulatory motion is characteristic where there is no mean bubble motion. This
for a given bioreactor. Form and size of the can be done by ultrasound Doppler measure-
flow structure depends (apart from the rheolo- ment. An example of measuring curves ob-
gy of the fluid) mainly on boundary condi- tained in a xanthan bioreactor is shown in
tions, i.e., on the geometry of the vessel which Fig. 4. This permits a determination of wheth-
confines the flow, and on the properties of the er the bubbles are moving or only fluctuating
power input. The liquid in a given cylindrical around a fixed position. More globally, the
vessel, for example, is capable of flowing in flow can also be characterized by the size of
several different global flow modes (LUBBERT, the circulation loops and by the mean time
1987). Different modes are supported or sup- particles need to circulate once around the
pressed by the way in which they are supplied reactor. Since this is a random variable, a
with the required power. These modes, charac- more complete characterization is given by a
teristic- or eigen-motions, form the intimate circulation time distribution. Circulation velo-
basis for the similarity approach, which can be cities influence the cooling capacity of the
used as a first approach in design and scale-up bioreactors (OUYOUNG et al., 1989) and are,
of bioreactors. furthermore, necessary to estimate the power
A complete description of this circulatory which must be applied to prevent sedimenta-
motion would require knowledge of the whole tion in respective applications.
liquid systems and their particular physico- and its proper detection in the output flow. It
chemical properties, e.g., the coalescence be- is not easy to mark the fluid elements at the
havior of the two-phase system. The latter is desired rate without changing their properties.
essential to all fluid-dynamical aspects of bio- Traced fluid elements must be assumed not to
reactors. This limitation led to various alterna- differ from all other elements of the fluid com-
tive physical methods, which - in contrast - ponent. Otherwise one cannot use them to esti-
are local measuring techniques. SRIDHARand mate the flow behavior of this component.
POTTER (1978) compared the chemical and the
physical light transmission method of CAL-
DERBANK (1958) and found discrepancies. 3.3.1.2 Gas Residence Times
They found that the chemical method consis-
tently yielded higher values of the interfacial A gaseous component which is steadily fed
areas in stirred vessels and supposed that this into the bioreactor, e.g., the air in aerobic fer-
was due to the much larger mass transfer coef- mentations, can easily be labelled by addition
of small amounts of a tracer gas, injected by
ficients within the impeller area. This essential-
ly says that at a constant specific interfacial pulses through small steel tubes directly into
area the results of the chemical method are de- the gas distributor. The tracer pulses can sim-
pendent on the hydrodynamics in the multi- ply be controlled by fast valves. As tracers, no-
phase flow. ble gases, e.g., helium, are most often used in
practice. Small amounts can be detected in the
gas outlet by mass spectrometers. A considera-
3.3 Residence Time Distributions ble experimental problem is the sampling of
and Related Measurements the tracer leaving the dispersion, since the sur-
face is strongly fluctuating and covered by
foam layers. There are two possible ways of
3.3.1 Tracer Techniques solving the problem. One is to use membrane
sampling devices (HEINZLEet al., 1983), and
3.3.1.1 General Methodology the other is to mechanically destroy the foam
entering the detection system by means of a
Residence time distribution (RTD) measure- small foam destroyer, as shown in Fig. 5. The
ments are techniques taken over from chemical mechanical technique (LUBBERTet al., 1987a)
engineering (DANCKWERTS,1953; NAOR and can be used when one cannot make sure that
SHINNAR,1963; NAUMAN and BUFFHAM, the membrane will be of constant transparency
1983), where an extensive literature exists on for the probe gas (e.g., in technical-grade sys-
the different aspects of the measuring tech- tems) over longer periods of time.
niques. All techniques are based on a stimu- In order to enhance the signal-to-noise ratio
lus/response approach. A tracer is used to of the results, it is advantageous to use signal
mark fluid elements in the input stream of the forms s(t)which differ from the conventional-
reactor and, viewed as the system’s response to ly used Dirac delta functions in order to mark
this marking, its concentration is observed at the gas flow dispersed into the reactor. Gains
the outlet of the reactor as a function of time. in the signalhoise ratio of up to two orders of
Although there may be situations in multi- magnitude can be obtained using pseudosto-
phase systems in which the interpretation of chastically distributed pulse trains of tracers.
residence time distributions measured by tracer The system’s response to such pseudorandom
techniques is difficult (SINNAR and RUM- signals then also becomes a random signal.
SCHITZKI,1989), the methods can provide Therefore, the weighting function, in this case
much insight into complex multiphase flows in the time-of-flow distribution, must be calcu-
biotechnology (BRYANT, 1977; LUBBERT et lated from s ( t ) and the system’s response a ( t )
al., 1987a; SWAINE and DAUGULIS, 1988, measured at the reactor exit by means of the
1989). cross-correlation function R,,(r) of both sig-
The main problems of these simple measur- nals (Fig. 6). &(r) is related to the weighting
ing techniques are due to the kind of marker function W(t) by the convolution integral
122 4 Characterization of Bioreactors
Foam separator
-1 000000000000
000000000000 -Foam destroyer
-Foam
-Gas sampling
Tracer input
Fig. 5 . Sampling technique built for gas residence time distribution measurements in pilot scale
fermenters which are operated with technical substrates (FROHLICH,1986). It consists of a bell-
shaped sampling funnel, a mechanical foam destroying unit, and a mass spectrometer connected
by a transmission line.
Rdt)=
m
s
--OD
RSAt - W ( 0dt (9) A L L Cross - correlation
functions are often measured at two locations, circulation time in batch fermentations has
as pointed out by EINSELE(1976) and SITTIG been estimated.
and RAMSPECK(1979). SYKES(1965) was the first to report use of
If one uses a single Dirac pulse of the tracer this method in investigating the circulation
as the signal, then the tracer concentration at paths traced by fluid particles inside biochemi-
the detector will be a damped oscillation (Fig. cal reactors. He used strips of plastics, which
7). The period q,of this oscillation is the exact were followed visually, and the number of cir-
circulation time. The damping of the tracer culations seen was written down. Such visual
concentration is due to liquid mixing, which experiments, which are restricted to clear liq-
smears out the marked fluid elements in space uids, have been performed by several investi-
and reduces the tracer intensity at the detection gators. For opaque media, more complicated
point when the cloud passes. From the enve- flow followers have been developed. In table-
lope of this damped oscillation one can, thus, tennis-ball-like bodies, small radio frequency
obtain the time constant of the decay of the transmitters were installed. The transmitted ra-
tracer cloud, which is one way of characteriz- diation was then detected by antennas which,
ing the time necessary for mixing (EINSELE, appropriately mounted, could detect the pas-
1976). In most cases, however, a simpler defi- sage of such bodies (BRYANT, 1969, 1977;
nition of the characteristic mixing time is used. MANN et al., 1981; OOSTERHUIS, 1984).
This is the time after which the decaying con- BRYANT(1977) reported use of radio pills, 20
centration fluctuations stay within predefined mm in diameter, each consisting of a plastic
limits around the final concentration value cf. sphere with a transmitter (10 MHz) and a bat-
Normally, one uses the interval [cf-5%, tery for its power supply. The even more min-
Cf+5%]. iaturized pills of SCHMIDTand BLENKE(1983)
The choice of an appropriate tracer is a dif- and BLENKE (1988), 13 mm in diameter,
ficult task and depends upon the particular worked for several months. In their experi-
process. As tracers one can use different dyes ments in airlift tower loop reactors, they used
(EINSELE,1976; BEYELERet al., 1981; EIN- two antennas. By recording the time difference
SELE et al., 1978; LAINE and KUOPPAMAKL, distribution between the passage times of the
1979; SCHEPER, 1985; SCHUGERL et al., probe body through both control volumes,
1987), alcohols (SWAINE and DAUGULIS, they additionally calculated a mean velocity of
1988, 1989), or radioactive tracers (PROKOPet the dispersion.
al., 1969; SEHERand SCHUHMACHER, 1979). There are several other ways to construct
flow followers. SCHMIDTand BLENKE(1983)
reported on pills which only contained an elec-
3.3.2 Flow Follower Techniques tric resonance circuit. Their advantage is that
no battery is needed, which is the most volumi-
There are several practical difficulties with nous component of radio pills. The antenna
tracers in measuring liquid circulation time within such a flow follower absorbs power, ra-
distributions in bioreactors. One is due to trac- diated from the transmitting antenna at the
er accumulation inside the bioreactor, if one reactor wall, every time the pill crosses
repeats the addition of tracer pulses to obtain through a control volume. This method has
more reliable results. Since, furthermore, it been implemented in tower jet loop reactors.
has not been simple to find proper liquid trac- Another possibility which may not be as at-
ers to mark fluid elements, some investigators tractive for bioreactors is the use of radioac-
have replaced tracers by macroscopic, inert, tive pills in conjunction with radiation detec-
solid particles adjusted in density to the disper- tors.
sion so that they are of neutral buoyancy. MUKAKATA et al. (1976, 1980) developed a
Thus, it can be assumed that they follow the permanent magnetic flow follower to study the
global liquid flow circulating through the reac- circulation time distribution in stirred tank
tor, at least in the time average. From the se- bioreactors. They twisted a wire around the
quence of the times at which these flow follow- reactor several times to build an inductance
ers pass a defined control region, the liquid coil in the plane defined by the impeller, as
Global Measuring Techniques 125
on the special biological system involved and ance when a strain is placed on them. In this
on the rheological conditions. Thus, the results application they measure the power delivered
must be assessed accordingly. to the shaft at the impeller end. The problem
that occurs in transmitting the measuring sig-
nal from the moving wheel outside the reactor
3.4 Power Input is frequently solved by telemetry readouts,
which are available commercially. Strain
Since nearly all fluid-dynamical parameters gauges must be calibrated. This can be done
in chemical or biochemical reactors depend on statically using weights and a lever arm. The
the mechanical power put into the medium, accuracy of a power measurement is of the or-
most of the transport mechanisms acting in der of a few percent (KUBOIet al., 1983).
bioreactors are controlled by it. Hence, its ac- Unfortunately, the strain gauges cannot be
tual value is of primary importance in charac- sterilized, therefore, measurements done dur-
terizing reactors. ing real fermentation runs in bioreactors can-
In stirred tank bioreactors, the power is in- not be found in the literature. As EINSELEand
troduced primarily by the agitation system, al- FIECHTER (1974) proposed, a strain gauge can
though in high, slender aerobic bioreactors a be mounted on the stirrer wheel between the
considerable part is due to the air compressed gear and the shaft seal. In that way, the seal is
into the liquid. A simple measurement of the a mechanical resistance, and it must be consid-
electrical power, drawn by the agitation sys- ered accordingly.
tem, is too inaccurate. It is very difficult to es- A method, which, however, is only applica-
timate the efficiency of the motor and especial- ble to smaller reactors, is the dynamometer
ly the frictional power losses due to the gear technique proposed by RUSHTONet al. (1950).
and agitator shaft seal. The best method of de- Here it is assumed that the torque transferred
termining the effective power input by a stirrer into the liquid can be measured by means of
is to measure the torque put onto the impellers the torque experienced by the reactor vessel. A
by the stirrer wheel. motor driven at variable speeds at a constant
Torque T, which the impellers transfer into torque is mounted on a revolving platform
the liquid, leads to a mean strain shear rate t that can be turned coaxially to the stirrer axis.
at the wall of the reactor (KAI and SHENGYAO, The frictional loss of the table is reduced as
1989): much as possible by means of a roller-bearing.
At the periphery of the table, the force neces-
sary to compensate for the torque introduced
by the action of the stirrer via the dispersion is
where k l is a constant of proportionality and V measured. Very small vessels can also be
the reactor volume. placed on an air-bearing to reduce friction
With stirrer speed N , general mechanics losses (CALDERBANK, 1967). These techniques
lead to the power input P have been used in vessels stirred with a Rush-
ton turbine, a radial conveying device. Stirrers
P = 271N T (Nm/s) which promote axial flow components may
not give correct results with such a measuring
In chemical engineering, one often uses the setup.
shaft strain gauge technique to measure torque One may ask, why not measure the electri-
T. This requires the insertion of torsion ele- cal power applied to the motor of the agitation
ments into the wheel of the agitator. The system? This will yield only very approximate
measurement is then done by means of a cali- results, since such a measurement is affected
brated strain gauge mounted onto the shaft of by large bias errors due to the efficiency of the
the torsion element. These elements are cov- motor and the power losses by the gear and the
ered by protective coatings to prevent the agitator shaft seals. Since these sources of
penetration of the fluid into the gauge. Such power loss are neither easy to determine nor
strain gauges are essentially electric metal layer constant, a proper correction cannot be made
resistances (ca. 1 kSZ) which change their resist- with sufficient accuracy.
Global Measuring Techniques 127
In airlift reactors, the power for the agita- culated into the stirrer region. The latter de-
tion is introduced pneumatically with the gas pends on the coalescence/redispersion proper-
phase. It can be measured in a much easier ties of the dispersion, which cannot be pre-
way. Since the gas input is usually assumed to dicted with sufficient accuracy. Only very re-
take place by isothermal expansion of the gas, cently have such patterns been measured ex-
it can be calculated by perimentally (BRORINGet al., 1990).
These techniques provide quantitative local most significant for the characterization of the
data on the geometrical form of the flow. performance of bioreactors. These are mainly
Hence, they may also be regarded as local the properties of the continuous liquid and the
measuring techniques. For the dynamic prop- dispersed gas phase. The developments in this
erties of the flow, they usually build a basis for field have been stimulated by investigations in
more detailed investigations using the local many other fields (WILDet al., 1986), e.g., by
measuring techniques discussed in the next sec- measuring techniques for multiphase flows in
tions, since local measurements can only be modern power-generation systems (JONES,
performed at a limited number of selected 1983) or blood-flow measuring techniques in
points that seem to be interesting with regard medicine (WEBSTER,1978; PAYNE,1985).
to the whole flow structure. Sometimes it may be interesting to calculate
global quantities from local measuring values,
e.g., the mean specific interfacial area. This is
not the main goal of local measuring tech-
niques. To obtain values of quantities spatially
4 Local Measuring averaged over the whole reactor, many local
values must be measured and integrated. These
Techniques may be very difficult to obtain in practice.
however, is interrupted if the point electrode is could only reach frequencies of up to 10 kHz
covered by an (insulating) gas bubble. Then, because of electronic difficulties. YASUNISHI
by means of an electronic signal-conditioning et al. (1986), however, reported 95 kHz. The
unit, one can obtain a binary signal, which difficulties arise mainly with small point
takes the value 1, if the point electrode is with- probes.
in a bubble and 0 if it is not. A necessary condition for the performance
of these probes is that the liquid film must dis-
1. Single-point probes appear at a sufficiently high rate from the
probe surface after the detection point has en-
Local gas holdup measurement tered a bubble to interrupt the current between
the two electrodes. This limits the viscosity of
If the current is recorded as a function of the liquid phase of the dispersions that can be
time, then, at a sufficiently long measuring in- investigated. Another severe limitation relates
terval T , the time-integral of the dimensionless to the bubble sizes that can be detected. These
signal can be interpreted as the integral time must be large as compared with the character-
TB the probe spent within bubbles. Divided by istic probe diameter, otherwise the probe will
T, this is equal to the probability of meeting a not pass across the bubble. The smaller the
bubble at the probe tip, i.e., to the local gas bubbles, the more they tend to follow the mo-
holdup there. tion of the continuous liquid phase around the
probe. Keeping the probe dimensions small is
E = lim TB/T necessary to reduce the interference between
t- m
probe and flow; normally this can only be ac-
In modern implementations, these times are complished at the price of reduced mechanical
determined in simple digital microprocessor- stability.
controlled devices. On account of the binary The measuring system must be calibrated
signal obtained from the analog electronics, because of the numerous effects which in-
one can omit an analog-to-digital converter. fluence the signals of the probes. Such a cali-
The data can simply be sampled using one bit bration in real flow conditions is not simple,
of a parallel digital input port of a microproc- since a measuring standard does not exist.
essor card. The necessary sampling rate of Thus, the probes are calibrated in slug flows
about 100 kHz can then be reached with sim- within small tubes, where one can measure and
ple 8-bit microprocessors. control the gas flow and the mean bubble vol-
Although the principle of this method is ex- ume by sampling a predefined number of bub-
tremely simple, there are some important tech- bles.
nical problems with the signal-conditioning Bubble probes proved to be applicable to
electronics. First of all, one cannot assume the fermentation broths, provided they are not too
current to be a smooth, stationary signal if the viscous (CZECH,1986; F R ~ H L I C H1987;
, LAR-
probe is operated within real liquids. On ac- SON et al., 1990). They are a simple means of
count of changes in the conductivity of techni- measuring profiles of the local gas holdups
cal broths and electrochemical effects at the which develop in all larger fermenters.
point electrode, one observes significant drift
in the current. The long-term conductivity 2. Two-point probes
drift must be compensated by appropriate con-
trollers and by using alternating voltages at Bubble velocity distributions
sufficiently high frequencies to reduce electro-
chemical effects. Generally, a few kHz would Bubble probes can be constructed based
be enough to avoid these effects; but if one upon the same principles as discussed above.
takes into account that the signal must be sam- They can be used to measure bubble velocity
pled at a rate of at least tens of kHz, one nor- distributions. Such bubble probes must con-
mally needs very different frequencies for the tain at least two point-like bubble detectors,
voltages which draw the current through the which are posed at adjacent points along the
liquid phase. LEWIS and DAVIDSON(1983) mean bubble path. These two detectors will
130 4 Characterization of Bioreactors
m
Metal wires ,
I
time delay t2, since the bubble surface needs
some time to proceed from one point to the
other. Since the distance dp between the two Voltage
detection points can be measured quite accu-
rately by means of a measuring microscope, Current
the velocity component v, of the bubble sur-
face along the direction defined by the two i
points can be calculated by
V, = d p / t 2 (15)
/Point electrodes
Since the electrodes must be at small dis-
tances of less than a millimeter, problems may Fig. 11. Principle of a four-point conductivity
arise with cross-talk and capacity effects, probe. The point electrodes are the end points of
thin wires placed parallel into a glass rod which end
which require special electronic measures in on the cone ground at the end of the rod. The sig-
order to obtain sufficiently high alternating nals obtained from twp of the point sensors for a
currents between the point electrodes and the time interval within which a bubble was hit by the
common counter-electrode at frequencies up probe are shown on the right-hand side of the fig-
to 100 kHz (LARSONet al., 1990). ure. From the times indicated, the bubble velocity
An interesting alternative was used by ZUN and chord length can be estimated as explained in
and SAJE (1982), who obtained the mean time the text.
delay between the two signals by calculating
the cross-correlation function of both signals.
The advantage is that one can eliminate the ef- BUCHHOLZand coworkers (BUCHHOLZet
forts to adjust the trigger levels which deter- al., 1983; STEINEMANN and BUCHHOLZ,1984,
mine the location of the edges of the condi- 1985; STEINEMANN, 1985) optimized the tech-
tioned signals. However, as can be seen later, nique by constructing extremely small multi-
with this correlation method one loses the in- point conductivity probes. The authors used
formation about bubble diameters that is also probe diameters down to about 0.5 mm, and
contained in the bubble signals. they claimed that bubbles down to two milli-
meters (STEINEMANN and BUCHHOLZ,1984)
3. Multi-point probes were detectable. Such extremely thin detectors
are, however, very susceptible to mechanical
If one constructs a 4-point bubble detector damage, which restricts their use to scientific
(Fig. 1l), in which the detection points are ar- work.
Local Measuring Techniques 131
ing the optical fiber probe is a measurement of ployed two such sensors to measure bubble
the light intensity reflected at the specially pre- velocities and sizes. CALDERBANK and PEREI-
pared tips of the fiber. RA (1977) constructed a multiple-fiber arrange-
In a fiber-optical probe, a light beam is ment similar to the conductivity probe of BUR-
guided through a thin glass fiber (Fig. 12). At GESS and CALDERBANK (1975). ABUAFet al.
the tip of the fiber, where there is no coating, (1978) improved the technique to feed back the
most light is transmitted into the liquid pro- reflected light. These authors connected two
vided the probe is immersed in a fluid with a thin 0.125 mm fibers to one sensor. DELHAYE
larger refraction index than the glass. Only a (1981) took even thinner fibers, 0.04 mm
small part of the light is reflected back into the mono-mode ones. LASA et al. (1984) used a U-
fiber. This reflected intensity is measured. shaped probe in a three-phase fluidized bed.
However, if the fiber tip is within a bubble, VAN DER LANS (1985) applied a probe consist-
which usually has a lower refractive index than ing of five-step index quartz fibers. FRIJLINK
the fiber, total reflection appears at the inter- (1987) improved the shape of the probe tip of
face, and consequently much more light ar- the step index fiber after a careful investiga-
rives at the detector. Although an optical tech- tion of the influence of the shape.
nique, it is not necessary that the liquid phase
be optically transparent. The time resolution
of the detectors is usually high. 4.1.1.3 Isokinetic Sampling Probe
There are some advantages to the optical
technique as compared to conductivity probes. Method
Thin liquid films on the sensitive fiber tip only
slightly decrease the signal level of the optical The isokinetic sampling probe is a bubble
sensor. Thus, one of the main disadvantages probe which differs from the point probes dis-
of the conductivity probes, e.g., the time con- cussed so far. A small stream of the dispersion
stant of the liquid film flowing off the point is sucked out of the reactor at a constant flow
electrode, is essentially removed. Another ad- velocity through a glass capillary (Fig. 13). In
vantage, common to all fiber-optical probes, is the capillary, the bubbles are elongated to
that the signal is not susceptible to electromag- form slugs filling the whole cross-section. At
netic cross-talk effects or other effects of elec-
tric or magnetic fields that influence electric
signal lines. The requirement for conductive Dispersion Detection unit
liquids obviously disappears. This, however, is Reactor wall
not relevant to biotechnological applications.
FRIJLINK(1987) investigated several probes
Pump
thoroughly. Some probe constructions suffer
from difficulties with contamination of the
probe surfaces, especially probes with sharp
edges due to grinding. Moreover, sharp pol-
ished tips are expensive and more fragile than
round shapes. At elevated viscosities, interfer-
ence effects between probe and bubbles caused
by bubble deformation were observed by FRIJ-
LINK (1987).
MILLERand MITCHIE(1969, 1970) were the
first to publish the idea of fiber-optical probes Fig. 13. Principle of the isokinetic sampling probe.
A small representative sample flow of the dispersion
in the chemical engineering literature. Instead is sucked through a capillary by means of a pump.
of fibers, they used single glass rods of several Bubbles are stretched to elongated slugs within the
millimeters in diameter in their pioneering ex- capillary, filling its whole cross-section. The length
periments. DANELand DELHAYE(1971) were of these slugs, which is proportional to the bubble
the first to use U-shaped optical fibers to con- volume, is measured by means of two adjacent light
struct bubble probes. GALAUP (1975) em- barriers.
Local Measuring Techniques 133
istic length of the optical path, and d , the scattering effects by LANDAUet al. (1977a,b),
Sauter diameter, defined by AL TAWEELet al. (1984), and URUAand DEL-
CERRO(1987).
Optical measuring techniques usually re-
quire sufficient light transmission through the
With the assumption of the presence of investigated part of the medium. In most prac-
merely spherical bubbles, geometrical relation- tical cases of biotechnological broths, this con-
ships between the gas holdup, the specific in- dition is not fulfilled.
terfacial area, and the Sauter bubble diameter In former years, many investigators tried to
lead to measure bubble-size distributions or specific
interfacial areas by means of photographic ob-
a=6c/dS (20) servations. These measurements cannot lead to
representative results, since the bubble-size
This, introduced into the above equation forf, and density distributions change significantly
leads to across the radius of a fermenter. Photographi-
cally, however, one only observes bubbles in a
f= exp ( - aL/6) (21) layer adjacent to the reactor wall. Thus, the
observations are not necessarily representative
which is nearly the result of CALDERBANK. of the overall values in the reactor. They can
MCLAUGHLINand RUSHTON(1973) veri- lead to contradictory conclusions, as OYE-
fied their results in light scattering in polydis- VAAR et al. (1988) have recently stated.
perse liquid-liquid dispersions by determining
the specific interfacial area. The authors stated
that the results were also applicable to other 4.1.2.2 Ultrasound Techniques
dispersions, because no requirements were
made on physical parameters other than the 1, Ultrasound Doppler technique
geometrical ones. The method has been im-
proved with regard to the influence of multiple The ultrasound Doppler technique is a meth-
od for measuring bubble velocity distributions.
The measuring device is very robust and does
not suffer from unknown interactions between
the probes and the flow of the dispersion with-
in the measuring volume. As opposed to bub-
ble probes, this technique can also be applied
to turbulent flows.
As schematically shown in Fig. 14, ultra-
sound pulses are transmitted into the disper-
sion. Bubbles act as excellent reflectors, hence,
part of the ultrasound power is reflected into a
detector. In the case of moving bubbles, the
detected signal is shifted in frequency accord-
ing to Doppler’s principle. This frequency
Fig. 14. Principle of the ultrasound Doppler tech- shift is proportional to the bubble’s velocity
nique to measure bubble velocities. An ultrasound component in the direction along half angles
wave is transmitted from the probe into the disper- between the input and the reflected beam.
sion and reflected at all bubble surfaces within the With electronic circuits it is possible to con-
region seen by the ultrasound detector. If the re- struct a signal, which fluctuates at the Doppler
flecting bubble is moving with the velocity uB, the
reflected ultrasound is shifted in frequency by an shift frequencies, from the signal employed to
amount fD. This shift is proportional to the bubble drive the transmitter and the detected reflec-
velocity component along half angles between input tion signal. A spectral analysis of that signal
and output beam. Furthermore, it depends on the then results in the distribution of the bubble
sound velocity c in the liquid phase. velocities along that direction in space.
Local Measuring Techniques 135
Measuring volume ger 1989). They are steam-sterilizable and
\ \
mounted on top of a steel tube, and they can
be installed in reactors of any size.
Typically, an ultrasound transmission fre-
quency of 4 MHz is used in bioreactors
I 1
(KORTEand LUBBERT,1985). The band width
- 1 -dL- of the Doppler shift signal is then less than 5
Ultrasound pulsed Ooppler kHz in most cases. Such signals can easily be
single probe technique digitized. Using modern microelectronic com-
Fig. 15. Single-probe ultrasound reflection tech- ponents, e.g., signal processors, the digital sig-
nique. An ultrasound pulse is transmitted from the nal analysis can be done in real-time. Unfortu-
probe into the dispersion. The pulse is reflected by nately, a simple calculation of the power spec-
bubbles within its path. A part of the reflected ultra- tral density function is not sufficient for an
sound power is then detected by the same ultra- analysis of the data as long as some bubbles
sound probe. By means of an electronic gating tech- approach the ultrasound probe while others
nique, only those reflections are recorded, which ar- disappear. If such bubbles have the same abso-
rive at the probe within a predefined time window.
Together with the known sound velocity a defined lute value for their velocity component along
measuring volume can be obtained. the axis, defined by the probe alignment, they
can not be distinguished, because, physically,
Doppler frequencies are always positive.
Approaching and departing bubbles can,
It is not necessary to use a separate detector however, be distinguished if one also takes ad-
because the same probes can be used as trans- vantage of the phase information in the signal.
mitters and receivers as well. If one employs This can be exploited by means of a quadra-
ultrasound pulses instead of a continuous ul- ture detection technique, as described by
trasound wave, it suffices to use only a single KORTE (1986). LUBBERTet al. (1987b) sam-
probe, which is alternatively switched t o be the pled and digitized Doppler signals at a rate of
transmitter or the receiver (Fig. 15). Using 20 kHz and showed how to calculate the veloc-
electronic gating techniques, the measuring ity distribution of the bubbles on-line. T o cope
volume can then be shifted within several cen- with the large data rate, they performed the
timeters apart from the probe. In that case, the analysis by means of special electronic devices
bubble velocities along the direction of the ul- based on signal processors (Texas Instruments
trasound beam are probed. TMS320). Such processors allow for calculat-
Ultrasound probes have been constructed ing the bubble velocity distribution simulta-
that can be used in fermentations (SOLLIN- neously with the data sampling procedure.
30'0i
Fig. 16. Typical bubble velocity distribu-
I
tion as measured with the ultrasound
7 pulsed Doppler technique in a stirred
=mm
c
20 0 tank bioreactor of 1 m3 operating vol-
P ume during a penicillin fermentation.
m The turbulent flow within this reactor
normally shows negative as well as posi-
tive velocity values, corresponding to
bubbles approaching and disappearing
from the probe. The noiseless curve is
the fit of a normal distribution to the re-
-100.0 - 50.0 0.0 50.0 100.0 sult, which shows that the velocity distri-
Bubble velocity lcm Is1 bution is slightly non-symmetric.
136 4 Characterization of Bioreactors
Thus, the resulting velocity distribution is di- mean reflecting surface, which, in turn, is
rectly available at the end of the data acquisi- characterized by their mean projected area A .
tion period. Since q is independently proportional to both
A typical result obtained in a stirred tank physical quantities, it is proportional to their
reactor during a penicillin production is shown product, which turns out to be the projected
in Fig. 16. Due to turbulent motions, one ob- area a, per unit volume.
serves positive as well as negative velocity val- For particles of random shapes that do not
ues along most flow directions in such flows. have concave surfaces. it was found that
The total measuring time required to obtain a
result of the quality shown in Fig. 16 is about q = a/4
10 minutes. The moments of the distributions
measured at different points in the reactor can where a is the specific interfacial area. The va-
be used to characterize the global bubble mo- lidity of this relationship was tested with light
tion. The mean values along the coordinate scattering in air-in-water dispersions, where
axis lead to contour maps like that shown in the scatterers were bubbles.
Fig. 2. The variances give some information As STRAVS and VON STOCKAR(1985)
about the randomness of the bubble motion, pointed out, this relationship can also be used
i.e., on the local mixing behavior. in ultrasound transmission experiments, if a is
constant along the radiation path and the fre-
2. Ultrasound transmission methods quency is far away from the resonance fre-
quencies of the bubbles, which are capable of
Since the work of SAUTER(1928), the trans- compression or elastic form oscillations.
mission of radiation through dispersions has Hence, if one measures the ultrasound absorp-
been a familiar technique to determine proper- tion Z/Z, along a predefined path length L , one
ties of a dispersed phase. Unfortunately, it is can determine the specific interfacial area a.
not possible to use the elegant light techniques, One avoids the effects of perpetual wave in-
since nearly all fermentation media are optical- terferences, if one uses an ultrasound pulse in-
ly opaque. Thus, some radiation must be chos- stead of continuous waves. By means of prop-
en which can be transmitted through cultiva- er gating, one reduces the influence of multiple
tion broths. One possibility is ultrasound. scattering. A closer look at this topic shows
If one transmits an ultrasound wave that the scattering of ultrasound waves by par-
through a dispersion containing ultrasound re- ticles is a function of the bubble radius and the
flectors such as gas bubbles, one observes a re- frequency of the wave.
duction in the transmitted power, as compared STRAVSand VON STOCKAR applied this sim-
to the transmission through the continuous ple transmission technique successfully to Can-
medium alone. This attenuation can be de- dida utilis cultivations. RIEBELand LOFFLER
scribed by the general physical equation of ab- (1989) reported on initial investigations ex-
sorption of radiation, where the intensity Z de- ploiting the wavelength dependency of ultra-
creases exponentially with the path length L : sound scattering. They used a broadband ul-
trasonic radiation source and measured the
Z/Zo = exp ( - q L ) or q = -In (Z/Zo)/L (22) transmission at different frequencies simulta-
neously. If the ultrasonic extinction cross-sec-
The absorption coefficient q is dependent tions were known (!), they showed how to set
on the wavelength of the transmitted radia- up a linear equation system that can be solved
tion. If the scattering particles appear at the to obtain particle-size distributions and par-
ultrasound beam as a projection area per unit ticle concentrations. With this ultrasound
volume, a,, then q is equal to a,, as CALDER- spectrometric technique, they demonstrated in
BANK (1958, 1967) showed for light scattering, model media that solid particles ranging from
in which the wavelength is small compared to 20 to 1000 Fm in diameter can be analyzed, if
the scattering centers. This is intuitively clear, frequencies in the range from 1.7-81 MHz are
since q should only depend on the number used.
density n of the reflecting bubbles and their
Local Measuring Techniques 137
exp ( -L a/4) (24) one can assume that practically the same a is
found at both places. BRORING(1990) was
L being the length of the ultrasound path, able to enhance the signal-to-noise ratio of the
which can be adjusted electronically by shift- a measurements in bioreactors with this tech-
ing the time gate within which the reflected sig- nique.
nal is recorded.
Since both effects are statistically indepen-
dent, the overall signal power P is proportion- 4.1.2.3 Further Radiation
al to the product of both:
Techniques
P(a)=const.aexp( -La/4) (25)
Since ultrasound transmission through bio-
Fig. 17 shows a plot of this function. The logical culture media is also limited to a maxi-
constant can be obtained experimentally by mum of ten centimeters, it leads one to think
simply reducing the gas flow through the reac- about radiation of other wavelengths or of
tor until the detected ultrasound power reaches other qualities. Electromagnetic radiation, for
its maximum value. The coordinates of this example, does not suffer from the transmis-
maximum are sufficient to determine the cali- sion limitation, if a much lower wavelength is
bration constants of both axes of the coordi- chosen than that of light. In the literature one
nate system. Thus, the relation P(a) can then can find examples in which light is replaced by
be used to determine the interfacial area from gamma radiation. YOUNG et al. (1987) used a
the measured ultrasound power, provided one translating gamma densitometer to measure
can make sure that the specific interfacial area the cross-sectionally averaged gas holdup pro-
a is definitely larger or smaller than amax.This file along the riser of an airlift reactor. The
decision is easy to make in normal cultivation void fraction can be calculated directly from
practice. the attenuation of the gamma beam. This tech-
The difficulty of making special calibration nique has the advantage that it can be used
measurements can be avoided by measuring non-intrusively, i.e., one can irradiate through
the reflected power at two different but known the reactor walls. It is also possible to use X-
positions along the probe axis. Then one can ray radiation. The technique is well established
determine the specific interfacial area from the in nuclear reactor thermal hydraulics, where it
two observed ultrasound intensity values. If is very important to measure void fractions
these two measuring points are close together, (GINOUX,1978). BEINHAUER (1971) and PIKE
138 4 Characterization of Bioreactors
et al. (1965) measured gas holdups in two- There is, however, another problem that
phase flows by X-ray attenuation. As a matter limits the application of laser-Doppler ane-
of principle, however, measuring techniques mometry (LDA) in multiphase flows. Since
using ionizing radiation do not find much ac- usually the measuring volume is small com-
ceptance in biotechnology. pared to the dispersed particles, e.g., bubbles
or solid particles, the signal, measured at one
particular point within the multiphase flow, is
4.2 Liquid Flow Properties often interrupted by signal segments which
contain invalid information. These correspond
to those time intervals during which the meas-
4.2.1 General Anemometers uring point falls inside a bubble or a particle.
At these places, the liquid velocity is unde-
Many different instruments have been devel- fined. To extract an exact mean velocity from
oped in hydrodynamics. As mentioned before, such a signal would be a formidable task, since
most of them, especially the elegant optical de- it would require detection and elimination of
vices, cannot be used in bioreactors, since the the invalid signal segments. Thus, the method
broths are optically opaque. Even if the con- is practically restricted to multiphase flows,
tinuous liquid phase is transparent, the great either with extremely small dispersed particles
density caused by the large number of small (CARTELLIERand ACHARD, 1985) or with
bubbles prevents the transmission of visible only a few large bubbles, the effect of which
light through the whole reactor. There may be on the signal can easily be recognized and
some very special cases where photographic eliminated.
methods, laser-Doppler, phase-Doppler, and
the many others can be applied, but in normal
cases of practical interest they cannot be used. 4.2.1.2 Hot Film Anemometers
Thus, optical techniques are not discussed in
much detail here. and Other General Techniques
In some bioreactors, e.g., in wastewater
4.2.1.1 Laser Doppler Technique treatment reactors at low gas loads, the me-
dium is so similar to water that one can use hot
Laser Doppler techniques are the most ele- film anemometry (HFA) to investigate the lo-
gant optical measuring methods to investigate cal flow field within the reactors. The HFA
single-phase flows. Sophisticated measure- technique has been applied to investigate flows
ments have been made to investigate the liquid in bubble columns as well as stirred tanks (Lu
flow inside chemical reactors by several au- and J u , 1987). Modern versions of anemomet-
thors (LAUFHUTTE and MERSMANN,1985; er probes are developed to make the technique
BAUCKHAGEet al., 1987; WEETMAN and direction sensitive. Dual split-film probes were
OLDSHUE,1988; COSTESand COUDERC,1988; applied by FRANZ(1983) and triple-split-film
W u and PATTERSON,,1989). Especially, local probes by MENZELet al. (1989). The argu-
investigations of velocity profiles and Rey- ments regarding signal interruptions, however,
nolds stresses in the surroundings of impellers raised in the discussion of point-like measuring
can give much insight into the pumping char- volumes in multiphase flows in connection
acteristics of mixing impellers under ungassed with laser-Doppler anemometry, also apply to
conditions. HFA. Therefore, HFA is not recommended in
One way to reduce the limitations in light normal biotechnological applications.
transmission is to reduce the light beam paths Other methods used in single-phase flows,
in the flow to only a few centimeters by means e.g., Pitot tubes (TANAKAet al., 1989) or fan
of optical fibers (KYUMAet al., 1981). If the velocimeters (HBNTZSCHGmbH, 1986), did
ungassed liquid is optically transparent, there not give reliable results in multiphase flows
is a chance that one can then measure liquid (HILLS, 1983).
velocities.
Local Measuring Techniques 139
4501 n
I . void fractions flowing at higher local veloci-
ties. Even in biotechnological cultivations at
higher viscosities, which are inaccessible by
other methods, measurements are possible.
Fig. 19 should serve as an example of a time-
Mycelium pmduction
of-flow curve measured during a rheologically
extremely complicated production of cephalo-
300pL..y-;-;:nz,-;
250’
,
0.5 1.0
..
1.5
...__._.:.
2.0
.
2.5
sporin in a tower loop reactor. This cultivation
was operated with an additional solid phase in
the form of peanut meal at a volumetric solid-
phase holdup of about 20%. It could be shown
that this technique can also be used in larger
airlift reactors to measure flow profiles at
Time-of-flow (s) higher gas loads. Fig. 20 shows an example ob-
Fig. 19. Time-of-flow result obtained within a labo- tained at 40 N m3/h, corresponding to 25 cm/s
ratory-scale airlift loop reactor during a production superficial gas velocity. Even three times that
of cephalosporin. Solid peanut meal at a volume gas load does not lead to problems.
fraction of about 30% was used as substrate. The Time-of-flow distributions contain informa-
probe distance was 13 cm. tion not only about the mean liquid velocity,
but also about the mixing behavior of the
flow. T o understand this, consider the fluid el-
2001 ements, which are marked by one pulse of the
Gas load:40iN m?hI heater, as a particle cloud. This cloud will be-
come larger with time through the influence of
the different mixing mechanisms. Thus, the
time dependence of the cloud width directly re-
flects the active mechanisms. LUBBERT and
LARSON(1990) showed that the standard de-
f“ 0 viations s of the time-of-flow curves can be re-
-50
lated to the mixing mechanisms. They ana-
-150 -125 -100 -75 -50 -25 0 lyzed the measured data by means of a nonli-
Radial distance (mml near least square fit of an appropriate model.
Fig. 20. Half profile of the axial component of the In the risers of airlift tower loop reactors, they
mean liquid velocity along the radius of the riser of found that most mixing was done by the wakes
a pilot-scale airlift tower loop reactor (4 m3). The of the rising bubbles.
data were obtained with the heat pulse technique at The pseudorandom time-of-flow technique
higher gas throughputs corresponding to 25 cm/s involving heat pulses to tag the fluid elements
superficial gas velocity. has also been used to measure very low volu-
metric flow rates of simple single-phase flows
in narrow tubes, e.g., in oil feed lines (WITTE
temperature signal intensities at probe dis- and BAIER, 1985), or even non-invasively in
tances of more than 10 cm. biotechnological and medical applications
The fluid elements can be heat-labelled by (KOHLRUSCH,1988; NEUGEBAUER, 1989).
several simple techniques. The simplest one is Heat as a possible tracer in local flow exper-
to use a heated wire (LOBBERTand LARSON, iments has been discussed several times in the
1986). A faster method is to use high frequen- literature. KRIZANand PILHOFER (1981) de-
cy currents directly through the electrically scribed a correlation method, in which they
conductive biosuspension (LARSONand LUB- correlated two temperature signals measured
BERT, 1990). As demonstrated, the method at two adjacent points along a trajectory in the
can be implemented very inexpensively. flow. Originally, an attempt was made to use
The heat pulse technique is especially suited the correlations between natural temperature
for measurements in dispersions with high fluctuations within the dispersion. However,
Conclusions and Future Trends 141
measurable effects turned out to require the gle-phase fluid dynamics. New methods have
use of a third probe which dispersed heat to to be developed to provide insight into the
the flow continuously. complex multiphase flows in real cultivation
Methods, which are based on a correlation broths. The techniques must continue to be de-
between fluctuating signals, e.g., velocity fluc- veloped in the future, since we have by no
tuations obtained from two measuring points means as yet attained a satisfactory state-of-
within an extended flow, cannot generally be the-art method.
applied in the same way as those known from Better measuring techniques are a prerequi-
tube flows. The distortions, leading to fluctua- site to obtaining data necessary to improve our
tions, may be influenced by a wave front, knowledge of bioreactors. Global measuring
which propagates in a direction different from techniques, such as residence and circulation
the line defined by the two probes or which time distribution measurements, well-known in
may have a propagation velocity different chemical engineering, can yield much valuable
from the fluid flow. In both cases, the correla- information about the global flow structure.
tion peak will not give correct information By new implementations, e.g., using pseudo-
about the time-of-flow of real fluid elements. stochastical instead of single-pulse or step exci-
Therefore, the pseudostochastical heat pulse tation, the decisive signal-to-noise ratio can be
technique is to be preferred. increased so much that more details can be ex-
tracted from the measuring data.
Many techniques developed so far utilize
4.2.2.3 Ultrasound Techniques probes in direct contact with the fluid ele-
ments. In most cases, this leads to a n interac-
POPOVIC and ROBINSON(1988, 1989) also tion between the two, which may change the
used an ultrasound Doppler device, but they sensors and/or the fluid flow. Although one
tried to determine the circulating liquid veloci- tries to minimize the interactions, it is better to
ty in an external loop airlift reactor. In such an avoid direct contact. This means that, as far as
application ‘no slip assumptions’ must be possible, methods must be developed which
made, i.e., the bubbles must be assumed to are based on radiation effects. Unfortunately,
move at the same velocity as the liquid phase. elegant optical methods, which have been de-
Such an assumption may be justified at the veloped in single-phase or low-density fluid
end of the downcomer of an airlift reactor, dynamics, cannot be used in most bioreactors.
since there are predominantly small bubbles Thus, new instruments using different wave-
there, which travel practically at the liquid ve- lengths or other types of radiation, e.g., ultra-
locity. The authors do not report applications sound, must be developed. As shown in this
of this method during real cultivations, but chapter, there are some promising first steps in
their measurements in cultivation broths of this direction.
Chaetomium cellulolyticum transferred into a Even from another point of view, the state
loop reactor, especially for test measurements, of development of multiphase flow measuring
give enough assurance that the technique can techniques leaves much to be desired. Measur-
also be used during cultivation. ing devices, if they exist for a nontrivial physi-
cal quantity in question, cannot be applied like
a simple thermometer. In most cases, they
must be adapted laboriously to the rheological
system under investigation.
5 Conclusions Furthermore, the measurement techniques
d o not provide data with spectroscopic accura-
and Future Trends cy. Whenever possible one should build in re-
dundancy, using more than one method, to
Fluid-dynamical characterization of bioreac- measure a quantity accurately and to ensure
tors during their normal operation cannot be that the devices are working correctly.
done with sufficient accuracy using the well- Dealing with many sensors simultaneously
known measuring techniques developed in sin- using model-aided measuring methods or so-
142 4 Characterization of Bioreactors
NAUMANN, E. B., BUFFHAM,B. A. (1983), Mixing transfer studies of external-loop airlifts and a
in Continuous Flow Systems, New York: John bubble column, AZChE J. 35, 393-405.
Wiley. PROKOP,A . , ERICKSON,L. E., FERNANDEZ, J.,
NEUGEBAUER, T. (1989), Komponenten zur Mikro- HUMPHREY,A. E. (1969), Design and physical
wellenerwarmung von Fliissigkeiten, Diploma characteristics of a multistage, continuous tower
Thesis, Universitat Hannover, FRG. fermentor, Biotechnol. Bioeng. 11, 945-966.
NIENOW,A . W., WISDOM,D. J., MIDDLETON,J. RAPER,J. A., DIXON,D. C., FELL, C. J. D., BUR-
C. (1978), The effect of scale and geometry on GESS, J. M. (1978), Limitations of the Burgess-
flooding, recirculation and power in gassed stir- Calderbank probe technique for characterization
red vessels, paper F1, 1-16, Proc. 2nd Eur. Conf. of gas liquid dispersions on sieve trays, Chem.
Mixing, Cranfield: BHRA. Eng. Sci. 33, 1405.
OLDSHUE,J. Y. (1989), Fluid mixing in 1989, RIEBEL,U., LOFFLER,F. (1989), On-line measure-
Chem. Eng. Progr., May 1989, 33-42. ment of particle size distribution and particle
OOSTERHUIS, N. M. G. (1984), Scale-up of bioreac- concentration in suspensions by ultrasonic spec-
tors - A scale-down approach, Doctoral Disserta- trometry, Chem. Eng. Technol. 12, 433-438.
tion, Technical University, Delft, The Nether- RIET, K., VAN’T (1975), Turbine agitator hydrody-
lands. namics and dispersion performance, PhD Thesis,
OUYOUNG, P . K., CHISTI,M. Y., MOO-YOUNG,M. Technical University, Delft, The Netherlands.
(1989), Heat transfer in airlift reactors, Chem. RIETEMA,K. (1982), Science and technology of dis-
Eng. Res. Des. 61, 451-456. persed two-phase systems, Chem. Eng. Sci. 31,
OYEVAAR, M., WESTERTERP, R. (1989), The use of 1125-1 150.
the chemical method for the determination of in- RIKMANIS,M., VANAGS,J., VIESTURS,U. (1987),
terfacial areas in gas-liquid contactors, Chem. Determination of the characteristic parameters of
Eng. Sci. 44, 2691-2701. mixing of media in bioreactors (in Russian), Bio-
OYEVAAR,M., ZIJL, A., WESTERTERP,R. (1988), technologija 3, 70-78.
Interfacial areas and gas hold-ups at elevated ROBERTSON, B., ULBRECHT,J. J. (1987), Measure-
pressures in a mechanically agitated gas-liquid ment of shear rate on an agitator in a fermenta-
reactor, Chem. Eng. Technol. 11, 1-10. tion broth, pp. 31-35, in: Biotechnological Proc-
PAULI,J., MENZEL,T., ONKEN,U. (1989), Direc- esses: Scale-up and Mixing (Ho, C. S . , OLD-
tional specific shear-rate measurements in gas-li- SHUE,J. Y., Eds.), New York: American Insti-
quid two-phase flow, Chem. Eng. Technol. 12, tute of Chemical Engineers.
374-378. RUSHTON,J. H., COSTRICH,E. W., EVERETT,H.
PAYNE,P. A. (1985), Medical and industrial appli- J. (1950), Power characteristics of mixing impell-
cations of high resolution ultrasound, J. Phys. ers, Chem. Eng. Progr. 46, 395 and 461.
E.18, 465. SAUTER,J. (1928), Forsch. Arb. Geb. Ing. 2/8,
PHILIP,J., PROCTOR,J. M., NIRANJAN, K., DAV- 312.
IDSON,J. F. (1990), Gas-holdup and liquid circu- SCHEPER,T. (1985), Messung zellinterner und zell-
lation in internal loop reactors containing highly externer Parameter zur Fermentationskontrolle,
viscous Newtonian and non-Newtonian liquids, Doctoral Dissertation, Universitat Hannover,
Chem. Eng. Sci. 45, 651-664. FRG.
RIKE,R. W., WILKING,B., WARD, H . C. (1965), SCHEPER,T., SCHUGERL,K. (1986), Bioreaktor
Measurement of void fraction in two-phase Charakterisierung mittels in-situ Fluorimetrie,
flow by X-ray attenuation, AIChE J. 11, 794- Chem. Ing. Tech. 58, 433.
800. SCHMIDT,J. (1990), Mixing mechanisms in multi-
PILHOFER,T., MILLER,H. D. (1972), Photoelektri- phase flows with a continuous liquid phase, On-
sche Sondenmethode zur Bestimmung der Bla- going Dissertation, Universitat Hannover, FRG.
sengroflenverteilung mitteldisperser Tropfen in SCHMIDT,K. G., BLENKE,H. (1983), MeBmethode
einem nichtmischbaren fliissigen Zweistoffsy- zur Bestimmung der Fliissigkeitsgeschwindigkeit
stem, Chem. Zng. Tech. 44, 295. von heterogenen Gas-Liquid-Systemen in Schlau-
PILHOFER,T., JEKAT,H., MILLER,H. D. (1974), fenreaktoren, vt Verfahrenstechnik 17, 593-597.
Chem. Ing. Tech. 46, 913. SCHNECKENBURGER, H., REUTER,B. W., SCHO-
POPOVIC,M., ROBINSON,C. W. (1988), External- BERTH, s. M. (1985), Fluorescence techniques in
circulation-loop airlift bioreactors: study of the biotechnology, Trends BiotechnoL 3, 257.
liquid circulating velocity in highly viscous non- SCHUGERL,K., LUBBERT,A., SCHEPER,T. (1987),
Newtonian liquids, Biotechnol. Bioeng. 32, 301- On-line Prozenanalyse in Bioreaktoren, Chem.
312. Ing. Tech. 59, 701-714.
POPOVIC, M., ROBINSON,C. W. (1989), Mass SCHUMPE,A., GRUND,G. (1986), The gas disen-
References 147
gagement technique for studying gas holdup Gas-/Flussigstromungen, Chem. Techn. 14, 146-
structure in bubble columns, Can. J. Chem. Eng. 153.
64, 891-896. STRAVS,A. A., STOCKAR,U., VON (1985), Meas-
SEHER,A., SCHUHMACHER, V. (1979), Determina- urements of interfacial areas in gas-liquid disper-
tion of residence times of liquid and gas phases in sions by ultrasonic pulse transmission, Chem.
large bubble columns with the aid of radioactive Eng. Sci. 40, 1169-1175.
tracers, Ger. Chem. Eng. 2, 117-122. SUN, Y., NOZAWA,T., FURUSAKI, s. (1988), Gas
SHAH, Y. T., KELKAR,B. G., GODBOLE,S. P., holdup and volumetric oxygen transfer coeffi-
DECKWER, W. D. (1982), Design parameters esti- cient in a three-phase fluidized bed bioreactor, J.
mations for bubble column reactors, AZChE J. Chem. Eng. Jpn. 21, 20-24.
28, 353-379. SWAINE,D. E., DAUGULIS, A. J. (1988), Review of
SHINNAR, R., RUMSCHITZKI, D. (1989), Tracer ex- liquid mixing in packed bed biological reactors,
periments and RTD’s in heterogeneous reactor Biotechnol. Progr. 4, 134-148.
analysis and design, AZChE J. 35, 1651-1658. SWAINE,D. E., DAUGULIS, A. J. (1989), Liquid re-
SITTIG,W., RAMSPECK,W. (1979), Methode zur sidence time distributions in immobilized cell bio-
Mischzeitbestimmung durch pH-Messung in Gas- reactors, Biotechnol. Bioeng. 33, 604-612.
Flussig-Reaktoren, Hoechst Forschungsbericht SWEERE,A. P. J., MESTERS,J. R., JANSE,L.,
Nr. 16. LUYBEN,K. C. A. M., KOSSEN, N. W. F.
SMITH,J. M., WARMOESKERKEN, M. M. C. G., (1988 a), Experimental simulation of oxygen pro-
ZEEF, E. (1987), Flow conditions in vessels dis- files and their influence on baker’s yeast produc-
persing gases in liquids, pp. 107-115, in: Biotech- tion: I. One fermentor system, Biotechnol. Bio-
nology Progress, New York: AIChE. eng. 31, 567-578.
SOBOLIK,V., MITSCHKA,P., MENZEL,T. (1986), SWEERE,A. P. J., JANSE,L., LUYBEN,K. C. A.
CSSR-Patent PV727886. M., KOSSEN,N. W. F. (1988b), Experimental
SOLLINGER,S . (1989), Entwicklung einer Ultra- simulation of oxygen profiles and their influence
schall-Doppler-Sonde, Diploma Thesis, Universi- on baker’s yeast production: 11. Two fermentor
tat Hannover, FRG. system, Biotechnol. Bioeng. 31, 579-586.
SPINDLER,K., LORENZ,G., HAHNE,E. (1987a), SWEERE,A. P. J., LUYBEN, K. C. A. M., KOSSEN,
Faseroptischer Sensor zur Messung des ortlichen N. W. F. (1987), Regime analysis and scale-
Gasgehalts in Flussigkeiten, Technisches Messen down: tools to investigate the performance of bio-
2, 50-55. reactors, Enzyme Microbiol. Technol. 9, 386-398.
SPINDLER, K., LORENZ,G., HAHNE,E. (1987b), SYKES,J. (1965), cited by BRYANT(1966).
Messungen des lokalen Gasvolumenanteils mit TANAKA,M., SENDAI,T., HOSOGAI,K. (1989),
Hilfe einer faseroptischen Sonde, Chem. Zng. Flowing characteristics in a circular loop reactor,
Tech. 59, 73-737. Chem. Eng. Res. Des. 61, 423-427.
SPINDLER, K., LORENZ,G., ERHARD,A., HAHNE, TAWEEL,A. M., AL, LANDAU,J. (1977), Turbulen-
E. (1988), Faseroptischer Sensor zur lokalen Gas- ce modulation in two-phase jets, Znt. J. Multi-
gehaltmessung in Zweiphasenstrdmungen, VDZ- phase Flow 3, 341-351.
Berichte 611, 237-240. TAWEEL,A. M., AL, DIVARKARLA, R., COMMA,
SRIDHAR,T., POTTER,0. E. (1978), Interfacial H. G. (1984), Measurement of large gas-liquid in-
area measurements in gas-liquid agitated vessels, terfacial areas, Can J. Chem. Eng. 62, 73-77.
Chem. Eng. Sci. 33, 1347-1353. TAWEEL,A. M., AL, DIVARKARLA, R., COMMA,
SRIRAM,K., MANN,R. (1977), Dynamic gas disen- H. G. (1987), Erratum: Measurement of large
gagement: A new technique for assessing the be- gas-liquid interfacial areas, Can. J. Chem. Eng.
haviour of bubble columns, Chem. Eng. Sci. 32, 65, 570.
571-580. TODTENHAUPT, E. K. (197 l), Blasengrofienvertei-
STEINEMANN, J. (1985), Charakterisierung des Gas- lung in technischen Begasungsapparaten, Chem.
phasenverhaltens in Blasensaulen mittels einer Zng. Tech. 43, 336.
neuentwickelten elektrischen Mehrpunkt-LeitfB- TRAGARDH, C. (1988), A hydrodynamic model for
higkeitssonde, Doctoral Dissertation, Universitat the simulation of an aerated agitated fed-batch
Dortmund, FRG. fermentor, pp. 117-134, in: Proc. 2nd Znt. Conf.
STEINEMANN, J., BUCHHOLZ,R. (1984), Applica- Bioreactor Fluid Dynamics (KING,R., Ed.), Lon-
tion of an electrical conductivity microprobe for don: Elsevier Applied Science Publisher.
the characterization of bubble behavior in gas- URUA,I. J., DELCERRO,M. C. G. (1987), Meas-
liquid bubble flow, Part. Charact. 1, 102-107. urement of large gas-liquid interfacial areas by
STEINEMANN, J., BUCHHOLZ, R. (1985), LeitfBhig- the light transmission method, Can. J. Chem.
keits-Mikrosonden zur Charakterisierung von Eng. 65, 565-569.
148 4 Characterization of Bioreactors
VERMA, A. K. (1989), Heat transfer in bubble co- DE KRAA,J. (1963), Interfacial areas in agitated
lumns, Chem. Eng. J. 42, 205-208. gas-liquid contactors, Chem. Eng. Sci. 18, 157-176.
WALTER,J. F., BLANCH,H. W. (1986), Bubble WICHTERLE, K., KADLEC,M., ZAK, L., MITSCH-
break-up in gas-liquid bioreactors: Breakup in KA, p. (1984), Chem. Eng. Commun. 26, 25-32.
turbulent flows, Chem. Eng. J. 32, B7-Bl7. WILD, G., ANDRE, J. C.5 MIDOUX, N. (1986),
WANG,D. I. c . , FEWKES,R. c. J. (i977), Effect of Mehr L i c k photophysikalische Meherfahren
operating and geometric parameters on the be- ZUr Untersuchung von Gas/Flussig-Reaktoren,
havior of non-Newtonian mycdial antibiotic fer- OWR Ink?. Tech. 58, 142-143.
mentations, Dev. Ind. Micrpbiql. 18, 39-57. WITTE,W., BAIER,P. W. (1983), Korrelative Volu-
wARMOESKERKEN, M, M, C, G,,sMITH, J , M.
menstrom-Messung von MineralBlen mit pseudo-
(1988), Impeller loading in multi-turbine vessels, zufalligen WarmeimPulsen, Chem. I W . Tech. 55,
179-197, in: Proc. 2nd Int. Conf. Bioreactor 795.
Fluid Dynamics (KING, R., Ed.), London: Else- wu, He, PATTERSON, G. K. (1989), Laser-Doppler
vier Applied Science Publisher. measurements of turbulent-flow parameters in a
stirred mixer, Chem. Eng. sci. 44, 2207-2221.
WASOWSKI,T., BLASS, E. (1989), Neue Aspekte
zum EinfluR des Blasenwakes auf die Hydrody- YASUNISH1f *., FUKUMAp M., MURoYAMA,K.
namik und den Stoffaustausch in einer Blasen- (19861, Measurement of behavior of gas bubbles
saule, Chem. Eng. Sci. 61, 519-530. and gas holdup in a slurry bubble column by a
WEBER, M. E., BHAGA, D. (1982), Fluid drift dual electroresistivity probe method, J. Chem.
caused by a rising bubble, Chem. Eng. Sci. 31, Eng. Jpn. 19, 444-449.
113-126. YOUNG, M. A., CARBONELL, R. G., OLLIS,D. F.
(1987), Non-Newtonian broths in airlift bioreac-
WEBSTER,J. G. (19781, Measurement of flow and tors, pp. 45-49 in: Biotechnology processes:
volume of blood, Chap. 8 in: MedicalInstrumen- scale-up and Mixing (Ho, c. s., J. y.,
tation (WEBSTER,J. G., Ed.)? Houston: Houg- Eds.), New york: American Institute of Chemi-
ton Mifflin Co. cal Engineers.
WEETMANj R. J * , OLDSHUE3J. y. (1988), Power, Yu, Y. H., KIM, S. D. (1988), Bubble characteristics
flow and shear characteristics of mixing impell- in the radial direction of three-phase fluidized
ers, Proc. 6th Eur. Conf. Mixing, Pavia, Italy. beds, AIChE J. 34, 2069-2072.
WEILAND, p., BRENTRUP, L., ONKEN, u. (19801, ZEHNER,P. (1986), Int. Chem. Eng. 26, 29.
Measurement of bubble size distribution in fer- ZUN, I., SAJE,F. (1982), Statistical characteristics
mentation media using a photo-electric probe, of bubble flow, P. 112, in: proc. 3rd Austrian-
Ger. Chem. Eng. 3, 296-302. Italian-Jugoslavian Chem. Eng. Conf., Vol. 2,
WESTERTERP,K. R., VAN DIERENDONCK, L. L., Graz.
5 On-Line Analysis of Broth
KARL SCHUGERL
Hannover, Federal Republic of Germany
1 Introduction 150
2 Need for On-Line Process Analysis and Control 150
2.1 Carbon Catabolite Regulation 150
2.2 Nitrogen Metabolite Regulation 151
2.3 Phosphate Regulation 151
2.4 Enzyme Induction and Regulation by Other Components 151
2.5 Realization of the Concentration Control of Key Components 151
3 On-Line Analysis in Non-Sterile Areas 152
3.1 Aseptic Sampling 152
3.1.1 Sampling Systems Integrated into a Recirculation Loop 152
3.1.1.1 Systems with a Flat Membrane 152
3.1.1.2 Systems with a Tubular Membrane 153
3.1.2 In situ Sampling Systems 154
3.1.2.1 Systems with a Discus-Shaped Body and a Flat Membrane 154
3.1.2.2 Systems with a Rod-Shaped Body and a Tubular Membrane 155
3.2 Analysis Techniques 156
3.2.1 Continuous Air-Segmented Analyzers 156
3.2.2 On-Line Flow-Injection Analyzers, FIA 159
3.2.3 On-Line High Performance Liquid Chromatography, HPLC 165
3.2.4 Fast Protein Liquid Chromatography, FPLC 167
3.2.5 Ion Chromatography, IC 167
3.2.6 Chemically Specific Detectors 169
4 Prerequisites for in situ Analysis 170
5 Future Developments 171
5.1 General Use of Spectroscopic Methods 171
5.2 Adaptation of Immunoassays for On-Line Monitoring of Proteins 172
5.3 Supercritical Fluid Chromatography 172
5.4 Improvement of the Reliability of the Applied Techniques by Means of Expert Systems 173
6 References 173
Biotechnology Second, Completely Revised Edition
Edited by H.-J. Rehm and G.Reed in cooperation with
A. Puhler and P. Stadler
copyright@WILEY-VCH Verlag GmbH, D-69469 Weinheim (Federal Republic of Germany). 2001
biotic synthesis. Oxygen starvation after the ranging from 0.3 to 300 mmol/L generally
onset of cephalosporin C production led to a support extensive cell growth, but concentra-
pronounced increase in penicillin N formation, tions of 10 mmol/L and above suppress the
which indicates that the ring expansion enzyme biosynthesis of many antibiotics (MARTfN and
is repressed by oxygen limitation (QUEENER et DEMAIN,1980; MARTfN, 1977). Therefore, by
al., 1984; SCHEIDEGGER et al., 1988). on-line measurement and control of the inor-
Several other carbon substrates (lactose, ganic phosphate concentration in the broth, a
galactose, mannan) do not influence the bio- high growth rate in the growth phase and a
synthesis of antibiotics, e.g., penicillin, actino- high productivity during the production phase
mycin, and streptomycin. The C-substrate can be attained.
consumption rate must be kept within a well-
defined range by on-line measurement and
control of the substrate concentration in the 2.4 Enzyme Induction and
broth during the production phase of second-
ary metabolites because of the carbon catabol- Regulation by Other Components
ite regulation.
Several enzymes are formed by induction
(BOING, 1982), e.g., amylase formation is in-
2.2 Nitrogen Metabolite Regulation duced by starch, invertase by sucrose, /3-galac-
tosidase by lactose, etc. Some of these inducers
Nitrogen metabolite regulation is well-docu- are used as substrates. Therefore, substrate
mented in bacteria, yeasts, and molds (MAR- fed-batch is practiced to keep the enzyme for-
TIN and DEMAIN,1980; MARTfN and LIRAS, mation rate high. By measuring the substrate
1981). Ammonia (or some other readily used concentration and keeping it in the optimal
nitrogen source) represses enzymes involved in range, a high enzyme productivity can be at-
the use of other nitrogen sources. Penicillin tained. The product of the enzymatic reaction
(HERSBACHet al., 1984) and cephalosporin often represses enzyme formation. By keeping
(MARTIN and DEMAIN,1980; QUEENERet al., the product concentration below the critical
1984) production is reduced by the inhibition level, the productivity can be improved.
of the activity of glutamate synthetase by in- The synthesis of some secondary metabo-
creasing the ammonium concentration in the lites can be enhanced by a precursor (e.g., the
broth. This holds true also for the synthesis of production of penicillin G by phenylacetic
other antibiotics (AHARONOWITZ, 1980). acid) or by a stimulator (cephalosporin C pro-
The ammonium (or any other quickly con- duction by methionine or sulfate, depending
sumable nitrogen substrate) concentration on the type of the strain) (MATSUMURA et al.,
must be measured and controlled in the broth 1978, 1980; SMITH,1985).
in order to achieve a high productivity because Productivity can be increased considerably
of this nitrogen metabolite regulation. by keeping the concentration of the precursor,
inducer, or stimulator in the optimal range.
a)
-
Fig. 1. Biopem. Aseptic on-line
b)
sampling device with disc-shaped
ultrafiltration membrane unit inte-
Projection X
grated in a loop, a) sampling mod-
B ule, b) schematic drawing with ad-
justable stirrer assembly.
A similar system has been developed by EN- ume of the module was 2.56 mL. However, the
GASSER (GHOULet al., 1986). The filter me- response time also depends on the volume of
dium consists of a continuous band of 10 cm the loop and on the broth circulation rate. In
wide ultrafiltration membrane, which is trans- the special case of a laboratory fermentor and
ported after each filtration process to avoid re- a short loop, the response time approached 5
duction of the permeation rate by membrane min or more at a broth flow rate in the tube of
fouling. 1 m/s.
The permeation rate decreased from about 1
mL/min to 0.6 mL/min with nylon and Dura-
3.1.1.2 Systems with a Tubular pore membranes and from 0.6 to 0.2 mL/min
Membrane with polysulfon 100000 membranes within sev-
eral days, provided broth of a high solid con-
A rod-shaped, stainless-steel module, 5.5 tent (peanut flour) was sampled. With broth of
mm in inner diameter, and a 200 mm long, tu- a low viscosity (e.g., broth of Zymomonas mo-
bular polypropylene microfiltration membrane bilk), the permeation rate remained constant
were used for on-line sampling of fermenta- (about 1 mL/min). At a high broth circulation
tion broths in a cross-flow mode (LORENZet rate - necessary to avoid fouling - marprene
al., 1987; SCHMIDTet al., 1986). The dead vol- tubing must be used in the peristaltic pump to
154 5 On-Line Analysis of Broth
O-b 5brnm-
O-rings 7
/ I
Membrane
Fig. 3. Aseptic in situ on-line sampling device with discus-shaped ultrafiltration membrane
unit (NIEHOFFet al., 1986).
On-Line Analysis in Non-Sterile Areas 155
example, during the sampling of Saccha- ble for long periods of time (1000 h) (NIEHOFF
romyces cerevisiae broth through an ex- et al., 1986).
ternal loop, ethanol was produced, The disadvantage of these types of modules
which falsified the analysis of the broth is that they can only be mounted inside the
composition. This is the reason why in reactor. Therefore, they are unsuitable for
situ sampling systems have been devel- large commercial reactors. This drawback is
oped. avoided by the use of rod-shaped modules
which can be mounted outside the reactor.
Different types of discus-shaped sampling
systems have been developed. A microfiltra-
tion device was placed near the stirrer to avoid 3.1.2.2 Systems with a Rod-Shaped
fouling (LORENZ et al., 1987). Better perform-
ance was attained with a discus-shaped ultra- Body and a Tubular Membrane
filtration membrane module (Fig. 3) (LORENZ
et al., 1987; SCHMIDT et al., 1986) and the dis- WAARVIK(1985) developed a rod-shaped
cus-shaped module with a porous plate mem- module with a microporous Accurel polypro-
brane support (Fig. 4), (SCHUGERL,1988). pylene membrane of 0.1 pm mean pore diam-
These modules with a 100000 Dalton ~ 0 1 ~ eter. - The membrane tubing had an inner diam-
sulfon-membrane have the following prbpei- eter of 5.5 mm, an outer diameter of 8.6 mm,
ties: and a surface roughness of 125 pm. It was
possible to maintain a filtration rate of 1.6
discus- disc- mL/min during the growth of Escherichia coli
shaped shaped (approx. 12 h) and 0.8 mL/min during the
module module production of penicillin V by Penicilliurn chry-
(Fig. 3) (Fig. 4) sogenum (approx. 250 h). Also, a 25 cm long
free filtration
surface area (cm2) 13.5 41.5 Ceraflo layered ceramic tubing with an outer
dead volume (mL) 1.35 2.5 diameter of about 5 mm and inner diameter of
membrane diameter (mm) 62 62 about 3 mm, a nominal surface pore size of
response time (min) 3.5 9 0.001 pm (ultrafiltration membrane) to 0.1 pm
(microfiltration membrane) and a surface
Both were used for the control of antibiotic roughness of 0.12 to 1.25 pm, was installed
production (SCHUGERL, 1988). They were and used for sampling the penicillin broth. It
used with microfiltration as well as with ultra- was possible to maintain a filtration rate of 0.5
filtration membranes. It was possible to steam- mL/min during the production.
sterilize them several times, and they were sta- A prototype of a nylon dialysis sampling
probe was described by GIBSONand WOOD-
WARD (1988) without information about its
Porous plate performance.
11 I
The ABC Company (ABC, 1988) offers a
Membrane
\-rings
rod-shaped, sterilizable stainless steel module
with polypropylene microfiltration tubing, de-
veloped at the University of Hannover (Fig. 5)
(GRAF, 1989; WENTZ,1989). This module has
the following properties:
Sulfate
a no cross-sensitivity exists with other amino acids and medium components except for histidine (37%) and
cephalosporin C (5%) (BAYERet al., 1986)
with blank after enzymatic cleavage of P-lactam ring
Tab. 2. Relative Molar Absorbances of Different possible to give generally valid data about the
Reducing Sugars and Sugar Derivatives with p- reliability of the analysis. Selectivity, detection
HBAH (p-hydroxybenzoic acid hydrazide) limit, and accuracy can only be given for a
(SCHMIDTet al., 1985) definite analyte in a particular broth and the
analyte concentration range using a specific di-
Substance Rel. Ab- Rel. Mol.
Conc. sorbance Absorbance lution and reaction.
(0.4 g L -') (070) (VO)
When using glucose oxidase (GOD) for glu-
cose analysis, H202 is formed, which reacts
Glucose 100 100 with an oxygen acceptor and forms a colored
Mannose 95 95 product. The sensitivity and detection limit of
Galactose 78 78 glucose depends on the chromogene selected.
Fructose 104 104 With four different chromogenes (sulfonated
Xylose 73 61 2,4-dichlorophenol and 4-aminophenazone; 3-
Ribose 14 62
77
methyl-2-benzothiazolinone (MBTH) and form-
Lactose * H 2 0 39
Maltose H 2 0 51 213 aldehyde azine of MBTH; 3,3',5,5'-tetra-
Cellobiose 60 114 methylbenzidine; 3-dimethylamino benzoic
Sucrose 0 0 acid (DMAB) and MBTH), the sensitivities
Raffinose * 5 HzO 1 3 vary in Saccharomyces cerevisiae cultivation
Trehalose . 2 H z 0 0 0 broths by a factor of 12 (0.0033, 0.0021,
Glucosamine HCl 58 69 0.030, and 0.042 (mg/mL) (BROWN et al.,
Glucuronic acid 1987).
(gamma lactone) 67 65 Calibration curves are shown in Fig. 7 for
N-acetylgalactosamine 26 32
NH:, phosphate, sulfate, urea, DOC, glu-
N-acetylglucosamine 26 32
cose, galactose, and lactose in penicillin broth
(NIEHOFF, 1983). The concentration ranges
covered by the analyzer system are depicted.
Selectivity, detection limit, and accuracy of The maximum error is less than 2%.
this technique depend on several factors, such Large amounts of chemicals are needed for
as the analyte and its concentration range (di- long-term cultivations, because the analysis is
lution), composition of the broth, and the carried out continuously. Since the detection is
reaction used for detection. Therefore, it is not often based on equilibrium reactions, these
On-Line Analysis in Non-Sterile Areas 159
lm
’-
;
50N
iC- 3 109 PPm
t
E
0.5-
v_ 100 300
C-
5 0 700 m g l t
cE0.5
t’
E
E 0.5
0.5
c-
1000 3000
cooc -5000 mglt
Tab. 3. Advantages and Disadvantages of Contin- HPLC systems for industrial process control
uous Air-Segmented Automatic Analyzers because of the above disadvantages.
Advantages
0 Continuous signal (easy to use for control)
0 Simple automation 3.2.2 On-Line Flow-Injection
0 Relatively inexpensive ($10000/channel) Analyzers, FIA
Disadvantages
0 Protein precipitation and cell growth in tubing, Flow-injection analysis (FIA) has become
coils, and detector very popular recently (RUZICKAand HANSEN,
0 Large amounts of chemicals are needed 1988). This technique is also based on the clas-
0 Long response time (the determinations are often sical wet chemical analysis.
based on the equilibrium of the reactions) The principle of the flow injection analysis
is simple, being based on the injection of a
definite volume of a liquid sample solution
into a moving, non-segmented continuous car-
analyzers usually have long response times. On rier stream of a suitable liquid (Fig. 8) (HAN-
the other hand, process control can be handled SEN, 1985). The injected sample forms a zone
easily by employing a continuous signal. which begins to disperse .and reacts with the
In Tab. 3 the advantages and disadvantages carrier stream as it is transported toward a de-
of the continuous air-segmented automatic tector, in which the concentration of the reac-
analyzers are compiled. tion product and/or substrate is measured. A
This technique will gradually be replaced by typical recorder output is shown in Fig. 8. The
on-line flow-injection analyzer and on-line peak height (H)or the surface below this peak
160 5 On-Line Analysis of Broth
Controlled disoersion
S
- 2 -30 s
W
..............
11 C
k
C
c c
Ileproducible timing
Fig. 9. Dispersed sample zone of original concentra-
tion Co-injected at position S and the corresponding
recorder output. To each concentration C corre-
sponds a specific dispersion value. Each specific dis-
persion value can be related to a fixed delay time t ,
which is at its minimum at ,C , and increases to
large values along the continuous gradient (HAN-
SEN, 1985).
p, .......
r;
PRINTER DETECTOR MANIFOLD INJECT VALVE PUMP rh SAMPLE CHANGER
EVA-Manifold EVA-Pump-
’ EVA- Select or
Accessories
0 EVA-lite 0 Merging o 8 Positions
0 Printer Stream Tee 0 External Loop 0 46 Channels o Sample Ports
o EVA-trade 0 Air Trap Injcct ion 0 Calibralion
o Recorder o Dispersion 0 Clulch Pork
0 EVA-zyme Coil 0 Time Based
0 Reduction Injection o Soft Start IVA-SamDler
0 EVA-duct Column o Sample
0 Host o Diffusion 0 Reversed Piercing
0 EVA-dapter Cell Flow 0 Wash Station
Fig. 10. Typical flow injection analyzer system with sample changer: EVA (Eppendorf Variables
Analysersystem).
(EPPENDORF GERATEBAU, Netheler & Hinz, Barkhausenweg 1, D-2000 Hamburg).
On-Line Analysis in Non-Sterile Areas 161
Tab. 4. Some Analytes Measured Off-line by FIA which are Important for Biomedical Technique and Bio-
technology
Tab. 4. Continued
Fig. 11. Typical flow injection analyzer system with on-line sampling unit: EVA (Eppendorf
Variables Analysersystem).
(EPPENDORF GERATEBAU, Netheler & Hinz, Barkhausenweg 1, D-2000 Hamburg).
On-Line Analysis in Non-Sterile Areas 163
1
Phenoxyacetic acid Tetrabutyl ammonium UV-detector [11, PI, PI
K-Penicillin V hydrogen sulfate/
p-Hydroxypenicillin V methanol on
Penicilloic acid nucleosil CIS5 pm
Penicilloic acid (reversed phase)
Methionine Tetrabutylammonium UV-detector
Penicillin N hydrogen sulfate on
2-Hydroxy-4-methyl- nucleosil CIS 10 pm
mercaptobutyric acid
Deacetylcephalo-
sporin C
Diacetoxycephalo-
sporin C
Cephalosporin C
Ethanol Differential 171
refractometer
Ethanol Differential [81
Glucose refractometer
Glycerol
Erythromycin UV-detector [81
D- and L-Amino acids D- and L-Amino oxidases, [9]
immobilized on Pt-
electrode, amperometric
detection of H202
Amino acids, Phosphate/methanol/ 4-Fluoro-7-nitrobenzo- [111
especially THF on nucleosil 2-oxa-l,3-diazole
tryptophan ODS (NBD-F) adducts,
fluorometer
Lactose Bio-Rad UV-detector [lo1
Glucose HPX-87-H
Galactose
Lactic acid
Acetic acid
Ethanol
Gluconic acid
The separation of amino acids and peptides published about the use of this technique in an
from protein hydrolysates and their quantita- on-line mode (Tab. 7).
tive determination by HPLC is a fairly diffi- In liquid chromatography very high purity
cult problem (HEARNet al., 1983). Recently a and care is required, since the smallest dust
combination of HPLC and mass spectrometer particles can stick to the capillary wall or in
has been used for this purpose (LEE and HE- the pores of the metal filter and cause clog-
NION, 1989). ging. The eluents are prepared with twice-dis-
In contrast to the large number of publica- tilled water, cleaned through a microfiltration
tions on HPLC, very few papers have been (e.g., 0.45 pm) membrane and degassed in an
On-Line Analysis in Non-Sterile Areas 161
ultrasound bath. To avoid gas absorption they fermentation broths. The main reasons are the
are gassed with helium during the analysis. difficulties connected with analytical tech-
The samples have to be freed of compounds niques (reliability), the technological problems
(e.g., proteins) which can precipitate in the (sampling), and the working conditions (GRES-
analyzer or adsorb on the column material. SIN, 1988). For the wider acceptance of this
Therefore, the samples are usually treated with technique, its selectivity, sensitivity, and relia-
methanol (1 : 1) and deproteinated by ultrafil- bility must be improved.
tration, if low molecular weight components GRESSIN(1988) reported on such a system
are to be analyzed. For more details see the used for monitoring of the protein concentra-
original papers. tion in fermentation broth, which was devel-
Usually a single calibration of the compo- oped at Rhone Poulenc SantC and based on an
nents for all of the samples from different pro- FPLC manufactured by Applied Automation.
duction phases is adequate. A comparison of Reversed liquid chromatography was used in
the off-line and on-line analysis results is rec- an isocratic mode on Spherosil P3,-600 C18 as
ommended. stationary phase at a moderate pressure (c100
If the separation of the peaks is satisfacto- bar), together with a UV detector with a fixed
ry, and the concentration of the components is wavelength, and a pre-column for increasing
in the intermediate concentration range, the the lifetime of the analytical column. The ana-
error of the analysis is usually less than 1-2070. lyzer was coupled to a data-acquisition system.
Usually 20-30 minutes are necessary for a Adaptive feeding of the substrate was possible
complete analysis of a multicomponent sys- by means of this system.
tem. Therefore, the analysis frequency is low. The analysis is more difficult if intracellular
In Tab. 8 the advantages and disadvantages protein monitoring is desired. Such a case was
of the on-line HPLC analysis are shown. described by GUSTAFSSON et al. (1986). The
On account of its high flexibility and excel- cells (Escherichiu colij were separated from the
lent performance, it is expected that on-line broth by centrifugation at 4900 g for 10 min to
HPLC will gain in importance in the future. eliminate the extracellular proteases. After
sonification of the cells, the cell debris was
separated again by centrifugation and the su-
Tab. 8. Advantages and Disadvantages of On-Line pernatant filtered through a microporous (0.45
HPLC Analysis wm) membrane. The filtrate was injected into
the FPLC anion exchanger column (Mono Q
Advantages HR 5 / 5 , Pharmacia AB), which was equipped
0 Analyzes several components at the same time
0 Easy automation
with a gradient programmer. Again, high-puri-
0 Relatively inexpensive (about $7500/component) ty water and analytical grade chemicals are
needed, and buffer solutions must be degassed
Disadvantages and filtered by microfiltration (e.g., 0.22 pm)
0 Very sensitive to impurities
0 Discrete data (difficult to use for control)
membranes, as in the HPLC analysis. With p-
0 Only 2-3 analyses per hour (low scanning rates
galactosidase as the model analyte, the elution
and large dead times) time was reduced to 9 minutes.
It is expected that the use of FPLC for the
monitoring of protein formation will gain in
importance.
3.2.4 Fast Protein Liquid
Chromatography , FPLC 3.2.5 Ion Chromatography, IC
Protein liquid chromatography is an estab- For the separation of ions of strong acids
lished technique in biochemical laboratories and bases (e.g., C1-, NO;, N a + , K + ) , ion
(RICHEY,1983; HORVATH,1980; BUSSOLO, chromatography has been developed by
1984). However, it is rarely used for the moni- SMALLet al. (1987) (MEYER,1988; SMITHand
toring of particular protein concentrations in CHANG,1983; GJERDEand FRITZ,1987). It is
168 5 On-Line Analysis of Broth
based on three different separation techniques: With the exchange of N a + for the H + of
ion-exchange (high performance ion chroma- the cation exchanger, the highly conductive
tography, HPIC, for larger ions), ion exclu- NaHCO, is converted into H2C03,which has a
sion (high performance ion chromatography low conductivity. On the other hand, NaCl
exclusion, HPICE), and ion pair formation and NaBr are converted into their highly con-
(mobile phase ion chromatography, MPIC) ductive acids. The highly conductive acids in
(WEISS, 1985). the weak conductivity H 2 C 0 3 solution permit
The equipment consists of a dispository for highly sensitive detection by the electrical con-
eluents, a pump, an injection valve, a column ductivity detector (especially suitable for tracer
for exchange reaction, a suppressor column, a analysis).
conductivity detector or photometer, and a re- Besides the potentiometric detectors, UV
corderhtegrator (DIONEXCORP.). and amperometric detectors are also used.
In contrast to HPLC columns, in which the However, the potentiometric detector remains
carrier is silica gel, the carrier in ion chromato- the standard. The use of the universal poten-
graphy columns often consists of polystyrene/ tiometric detector based on electrical conduc-
divinylbenzene (PS/DVB) resins. The latter tivity is hampered by the high-ion-strength
have a much better pH stability (they are sta- eluents. Therefore, either other detectors or
ble in the pH range 0 to 14) than silica (pH 1 low capacity resins are used for the separation
to 9). Eluents with extreme p H values must be (HADDADet al., 1986; HORVAIet al., 1988).
applied to bring compounds such as sugars Organic acids can also be separated by the
and alcohols to an ionogenic state. The degree anion exchange column. The separation of
of the cross-linking of the resin ranges between carbohydrates is difficult. They can only be
2 and 5 % to achieve the optimal porosity of separated by means of strong alkaline eluents
the carrier (WEISS, 1985). on a strong basic anion exchanger.
The anion exchanger of Dionex consists of Cation exchange chromatography columns
PWDVB particle cores (10-25 pm) with a sul- also consist of a surface-sulfonated PS/DVB
fonated surface and aminated porous latex resin core. On the surface of the core between
particles (0.1 pm), which adhere to the surface the SO; H groups and the cation M , the
+ +
Again, suitable suppressor systems and elec- ideal detector should have the same sensitivity
trical conductivity cells are used as detectors. for each analyte. If the analytes are not sepa-
Inorganic and organic anions can be separated rated, the ideal detector should detect only a
in a single run within 30 min in combination single analyte. It should be selective, i.e., it
with ion-exchange chromatography. should have a high chemical specificity.
Alternatively to ion exchange chromatogra- Furthermore, the ideal detector should have
phy, ion pair chromatography is gaining im- a short time constant and should be very sensi-
portance, since anions as well as cations can be tive to the analyte, but it should not be sensi-
separated and analyzed by this process. An or- tive to variations in temperature, pressure,
ganic ionic component, which forms an ion flow rate, or chemical composition of the ma-
pair with the counter-charged analyte of the trix (i.e., it should be noise-free).
sample, is added to the mobile phase. This ion Chromatographic techniques require non-
pair is a salt, but it behaves like a non-ionic specific detectors based, e.g., on the refractive
organic molecule. It can be separated by re- index (RZ),ultraviolet (UV) absorbance, fluo-
versed-phase chromatography. rescence emission, electrochemical reaction, or
One uses, e.g., alkylsulfonate for the analy- electrical conductivity. Continuous air-seg-
sis of cationic analytes and, e.g., tetrabutylam- mented and flow injection analyzers need
monium phosphate for the analysis of anionic chemically specific detectors.
analytes. Measurement of the concentration of sev-
The advantages of ion pair chromatography eral analytes is based on specific wet chemical
are: reactions that form easily detectable com-
pounds. However, inorganic ions can be di-
0 separation in a reversed-phase system is rectly detected by ion-selective electrodes
possible, (SCHINDLER and SCHINDLER, 1983), low mo-
0 mixtures of acids, bases, and neutral lecular weight organic compounds by enzymat-
components as well as amphoteric mole- ic reaction (BOWERS and CARR, 1980;
cules can be separated, SCHINDLER and SCHINDLER, 1983), and pro-
0 the selectivity can be controlled by teins by means of immunoassays (SCHARPEet
means of the counter-ion (TOMLINSON al., 1976; VORLAENDER,1980; SCHINDLER
et al., 1978; BIDLINGMEYER, 1980; and SCHINDLER, 1983).
GLOORand JOHNSON,1977). For ion-selective detectors see Chapter 1,
Section 6.
For more details see WEISS (1985) and The combination of transducers and bio-
MEYER(1988). chemical receptor membranes and their in situ
Ion chromatography is mainly used in in- use as biosensors are considered in Chapter 3
dustry for water analysis (WEISS,1985). Up to (Biosensors). However, this receptor-trans-
the present, no application for the analysis of ducer combination is often applied as a detec-
fermentation broth is known. This lack is tor for continuous air-segmented analyzers or
probably due to the complex matrix and high FIA-systems. In such a case, the biochemical
ion strength in the broths. receptor is not necessarily used as a membrane
It is expected that these difficulties will be directly connected to the transducer.
overcome and ion chromatography will be-
come a standard method for broth analysis. The following Combinations are used:
d) the receptor is immobilized on a mem- as on the volume and time constant of the re-
brane and combined with a separate ceptor-transducer system and the mass disper-
transducer. sion therein. Detectors consisting of thin re-
ceptor membranes with direct contact to the
The signal quality depends on the fractional transducer surface have the lowest time con-
conversion of the enzymatic reaction (which is stant.
related to the amount of enzyme present and The behavior of the electrochemical trans-
the time during which the analysis is run) as ducer is discussed in Chapter 1, that of the
well as on the mass dispersion during the anal- transistor in Chapter 3, and that of the ther-
ysis. A larger amount of enzyme can be immo- mistor and fluorometer in Chapter 6.
bilized on the surface of carrier particles than
on a tubing surface or a membrane. Further-
more, the reaction time in a cartridge or tubing
is longer than the contact time with a mem-
brane. 4 Prerequisites
The mass dispersion in a small cartridge
filled with carrier particles is very low (close to
for in situ Analysis
the plug flow) as long as the cartridge is uni-
formly filled. Because of the parabolic laminar On-line aseptic sampling is required for on-
velocity profile of the sample flow in the tub- line analysis of the medium composition out-
ing, the mass dispersion in it is higher than in side the reactor. Depending on the molecular
the cartridge. weight of the analytes, ultrafiltration or micro-
On the other hand, a comparison of disper- filtration membrane filters are used. The sam-
sion in the membrane and cartridge systems in- ple must be prepared for the analysis with con-
dicates that the mass dispersion is only lower tinuous air-segmented, FIA, or chromatogra-
in the cartridge than in the tubing if the flow phic techniques (removal of proteins, mixing
across the membrane is completely uniform. with reagents, derivatization, etc.). Further-
Generally speaking, when using a cartridge more, frequent calibrations and blank determi-
the highest signal can be expected with the nations are required to compensate for drifts
highest fractional conversion and the smallest (caused by variations of the activity of the
mass dispersion. In addition, the enzyme activ- chemically-specific receptor, of the transparen-
ity of the cartridge changes only slowly, thus cy of the optical detector windows, precipita-
its utilization time is much longer than that of tions, etc.).
a receptor immobilized on tubing or mem- Sample preparation, regular calibrations,
branes. and blank determinations are not possible for
Transducers are based on the variation of in situ analysis by sensors. Furthermore, when
p H [potentiometric p H electrodes, IS-FETs using monocultures for product formation, the
(pH-sensitive field effect transistors)], po, analyzer system must be sterilized. When us-
(amperometric po, electrodes), pco, (poten- ing, e.g., an enzyme electrode, the signal is in-
tiometric pH-electrodes), pNH,(potentiometric fluenced by the variation of the enzyme activi-
NH,-sensitive electrodes), Hz02(amperomet- ty due to enzyme loss, its deactivation (denatu-
ric electrodes) values, or the change of the lo- ration, decomposition by proteases or non-
cal temperature (thermistor, a minicalori- specific protein adsorption) or inhibition, as
meter; due to reaction enthalpy), fluorescence well as by the response of the transducer to the
intensity (fluorometer, e.g., due to cofactor- local and instantaneous variations of the prop-
NADH conversion), and color or luminescence erties of the medium (pH, p o z,pco,, turbidity,
intensity (photometer) (BOWERSand CARR, absorbance, fluorescence emission). These are
1980; KRICKAand THORPE,1986; TURNER^^ al., the reasons why biosensors cannot be used in
1987; RECHNITZ,1988; SCHMIDet al., 1987). situ for process control.
The performance of the detector also de- By measuring the local and instantaneous
pends on the transducer properties (time con- variation of broth properties by a second and
stant, sensitivity, selectivity, stability), as well perhaps a third transducer of the same type
Future Developments 17 1
close to the biosensor, and using this informa- Tab. 9. Important Aspects of FT-IR Technology
tion to correct the signal, the direct transducer (FINKand CHITTUR,1986)
response can be eliminated. However, the ~
It should be mentioned here that the combi- the windows, etc. (see Chapter 17). However,
nation of different chromatographic systems the automated system is not able to diagnose
(e.g., size exclusion chromatography with re- errors or subsequently to correct data or elimi-
versed-phase chromatography, HPLC and ca- nate the failures.
pillary GC, capillary GC and capillary SFC, For error diagnosis and failure correction, a
HPLC and capillary SFC) is an excellent meth- special closed-loop control expert system is
od for obtaining increased resolving power for necessary. The first steps in this direction have
multicomponent mixtures, especially if com- already been taken (FISHand F o x , 1988; espe-
pletely independent techniques are used (Lu- cially LUBBERTet al., 1988).
RIE, 1988). It is expected that in the near future several
expert systems will be developed for error di-
agnosis in on-line process analysis. This is a
5.4 Improvement of the Reliability prerequisite for the application of complex
of the Applied Techniques analyzer systems in process control.
by Means of Expert Systems
When using process analysis data for proc-
ess control, reliability is extremely important. 6 References
Several possibilities exist for malfunction of
the sterile sampling system as well as of the
analyzer system. In all analyzers, sample pre- ABC (1988), Sampling System. ABC Biotechnolo-
treatment (dilution, neutralization, filtration) gie/Bioverfahrenstechnik GmbH, P.O. Box 1151,
is necessary. The filtration, due to variation in D-8039 Puchheim.
the transmembrane permeation rate caused by AHARONOWITZ, Y.(1980), Nitrogen metabolite reg-
membrane fouling (usually protein adsorption) ulation of antibiotic biosynthesis, Annu. Rev.
or drift of the pumping rates on the different Microbiol. 34, 209-233.
channels can cause serious errors in the dilu- ALBERTI, J. C., PHILLIPS, J. A., FINK,D. J., WA-
SACZ, F. M. (1986), Off-line monitoring of fer-
tion of the analyte as well as in the p H of the mentation samples by FTIR/ATR: A feasibility
reaction mixture and the rate of the reaction study for real-time process control, Biotechnol.
needed for formation of the compound to be Bioeng. Symp. 15, 689-722.
detected. ALLEN,B. R., LULI,G. W. (1985), A gradient-feed
With the continuous air-segmented analyz- process for obtaining high cell densities for re-
er, the most common failures are caused by in- combinant E. coli, Paper presented at the Na-
complete phase separation, base-line drift due tional American Chemical Society Meeting, Chi-
to protein precipitation, and cell growth in the cago, September, 1985.
coils and in the detectors. ALWARTHAN, A., TOWNSHEND, A. (1988), Deter-
In FIA systems, the most common failures mination of tetracycline by flow injection with
chemiluminescence detection, Anal. Chim. Acta
are due to poor reproducibility of repeated in- 205, 261-265.
jections, sluggish peak response in reaching BAUER,B., FLOYD,T. A. (1987), Monitoring of
the base-line, base-line drift, air bubbles, glucose in biological fluids by Fourier-transform
abrupt noise on peaks, noisy signal, double infrared spectrometry with a cylindrical internal
peaks, and negative peaks (RUZICKAand reflectance cell, Anal. Chim. Acta 197, 295-
HANSEN,1988). 301.
In addition to the possible failures of FIA BAYER, T. (1987) Reaktionstechnische Untersu-
systems, base-line drift in HPLC systems may chungen zur Produktion von Cephalosporin C in
occur due to the precipitation of proteins. Mammutschlaufenreaktor, Dissertation, Univer-
sitat Hannover, FRG.
If such data are to be used for process con- BAYER, T., HEROLD,T., HIDDESSEN, R., SCHU-
trol, an automated system is necessary to re- GERL,K. (1986), On-line monitoring of media
calibrate the analyzer system, evaluate the components during the production of cephalo-
blank values, clean the channels periodically, sporin C, Anal. Chim. Acta 190, 213-219.
control the flow rates and the transparency of BAYER,T., ZHOU, W . , HOLZHAUER, K., SCHU-
174 5 On-Line Analysis of Broth
trol of glucose concentration, Biotechnol. HORVAI,G., PAL, H., NIEGREISZ,T., TOTH, K.,
Bioeng. 28, 119-121. PUNGOR, E. (1988), Electrochemical detectors in
GIBSON,T. D., WOODWARD, J. R. (1988), Contin- ion chromatography, LC-GC-International2 (l),
uous, reliable on-line analysis of fermentation 32-38.
media by simple enzymatic/spectrophotometric HORVATH, C. (Ed.) (1980),High Performance Liq-
assays, Anal. Chim. Acta 213, 61-68. uid Chromatography, Advances and Perspec-
GJERDE,D.T., FRITZ,J. S. (1987),Zon Chromato- tives. New York: Academic Press.
graphy, 2nd Ed. Heidelberg: Huthig. HUEN,J. M. (1984),HPLC analysis of amino acids
GLOOR,R., JOHNSON,E. L. (1977), Practical as- in beverages, IBL May/June, 50-53.
pects of reversed phase ion pair chromatography, HUSTEDT,H., KRONER, K. H., KULA, M. R.
J. Chromatogr. Sci. 15, 413-423. (1 989,On-line-Bestimmungen von Enzymen und
GORTON,L.,HEDLUND,A. (1988), A flow-injec- Proteinen zur Prozenkontrolle in der Enzymge-
tion method for the amperometric determination winnung, BTF-Biotech-Forum 2, 57-63.
of L-lactate with immobilized enzymes and a HUBNER, U. (1990),Dissertation, Universitat Han-
chemically modified electrode, Anal. Chim. Acta nover, FRG.
213, 91-100. IMAI,K., UEDA,E., TOYOKA,T. (1988), High-per-
GRAF, H. (1989),Dissertation, Universitat Hanno- formance liquid chromatography with photo-
ver, FRG. chemical fluorometric detection of tryptophan
GRAM,J., DE BANG, M., VILLADSEN,J. (19871, based on 4-fluoro-7-nitrobenzo-2-oxa-1,3-dia-
Flow injection analysers for on-line monitoring zole, Anal. Chim. Acta 205, 7-14.
of glucose isomerisation process, in: Proc. 4th IMMING,G., SCHALLER, K., MEINERS,M. (1982),
Eur. Congr. Biotechnology, Vol. 4, pp. 409-415 The use of automated analyser to control micro-
(NEISSEL,0 . M., VAN DER MEER,R. R., LUY- bial processes, in: Proc. First ZFAC Workshop
BEN, K. Ch. A. M., Eds.). Amsterdam: Elsevier on Modelling and Control in Biotechnical Proc-
Science Publishers B.V. esses, Helsinki (HALME,A., Ed.). Oxford-New
GRESSIN,J. C. (1988), On-line process liquid chro- York: Pergamon Press.
matography, a useful tool in biochemical proces- KARLBERG, B. I. (1983),Automation of wet chemi-
ses, BTF-Biotech-Forum 5, 38-43. cal procedures using FIA, Znt. Lab., Jan./Febr.
GUSTAFSSON, J. G., FREJ, A. K., HEDMAN,P. 82-90.
(1986),Monitoring of protein product formation KARUBE,I., SUZUKI,M. (1986),Novel immunosen-
during fermentation with fast protein liquid chro- sors, Biosensors 2, 343-362.
matography, Biotechnol, Bioeng. 28, 16-20. KENNEDY,J. F., RIVERA,Z. S., WHITE, C. A.
HADDAD,P. R., ALEXANDER, P. W., TROJANO- (1989), The use of HPLC in biotechnology, J.
WICZ,M. (1986), The application of a metallic Biotechnol. 9, 83-106.
copper electrode to potentiometric detection of KRICKA,L.J., THORPE,G. H. G. (1986), Immobil-
reducing species in ion chromatography, J. Liq. ized enzymes in analysis, Trends Biotechnol. 4
Chrornatogr. 9,777-789. (lo), 253-258.
HANSEN,E. H. (1985),Recent advances in flow in- KRONER,K. H., KULA,M. R. (1984),On-line meas-
jection analysis, Znt. Lab., October, 14-23. urement of extracellular enzymes during fermen-
HANSEN,E. H.(1989),Flow-injection enzymatic as- tation by using membrane techniques, Anal.
says, Anal. Chim. Acta 216, 257-273. Chim. Acta 163, 3-15.
HEARN,T. W., REGNIER,F. E., WEHR, C. T. KRONER,K. H., STACH, W., KUHLMANN,W.
(1983),HPLC of peptide and proteins, Znt. Lab., (1986), Kontinuierliches Probeentnahmeverfah-
JanJFebr., 16-35. ren fur die Bioprozenanalytik, Chem. Tech. 15,
HERSBACH, G. J. M., VAN DER BEEK,c. P., VAN 74-77*
DIJCK,P. W. M. (1984),The penicillins: Proper- KUHLMANN, W.,MEYER,H. D., BELLGARDT, K.
ties, bio-synthesis, fermentation, in: Biotechnolo- H., SCHUGERL,K. (1984), On-line analysis of
gy of Industrial Antibiotics, pp. 45-140 (VAN- yeast growth and alcohol production, J. Biotech-
damme, E. J., Ed.). New York: Marcel Dekker. nol. 1, 171-185. I
HOLST,O., HAKANSON, H., MIYABASHI, A., MAT- LARSSON,K. M., OLSSON,B., MATTIASSON,B.
TIASSON, B. (1988), Monitoring of glucose fer- (1987),Flow injection enzyme immuno assay, in:
mentation processes using a commercial glucose Proc. 4th Eur. Congr. Biotechnology, Amster-
analyser, Appl. Microbiol. Biotechnol. 28, 32-36. dam, Vol. 3, pp. 196-199 (NEISSEL,0. M., VAN
HOLZHAUER, K. (1987), Entwicklung und Anpas- DER MEERR. R., LUYBEN, K. Ch. A. M., Eds.).
sung chromatographischer Verfahren zur Verbes- Amsterdam: Elsevier Science Publishers BV.
serung der Antibiotikaproduktion, Diploma The- LATER,D. W., BORNHOP,D. J., LEE, E. D., HE-
sis, Universitat Hannover, FRG. NION,J. D., WIRBOLDT, R. C. (1987),Detection
176 5 On-Line Analysis of Broth
GERL, K. (1986), The use of an automatic on-line (1985b), Rapid on-line protein detection in bio-
system for monitoring penicillin cultivation in a technological processes by flow injection analysis
bubble-column loop reactor, Anal. Chim. Acta (FIA), Enzyme Microbiol. Technol. 7, 146-149.
190, 205-212. RHODES,P. M. (1984), The production of oxytetra-
NIKOLAJSEN,K., NIELSEN, J., VILLADSEN,J. cycline in chemostat culture, Biotechnol. Bioeng.
(1988), In-line flow injection analysis for moni- 26, 382-385.
toring lactic acid fermentations, Anal. Chim. RICHEY,J. (1983), FPLC: A comprehensive separa-
Acta 209, 139-145. tion technique for biopolymers, Int. Lab., Jan./
OH, T. K., KIM, S. H. (1986), Determination of cel- Febr., 50-75.
lobiohydrolase from culture filtrate of Trichoder- ROSS, A. (1986), Probeentnahmesystem zur An-
ma viride by the method of single immunodiffu- kupplung eines Autoanalysers an einen Fermen-
sion and enzyme-linked immunosorbent assay, ter, in: Technische Membranen in der Biotechno-
Biotechnol, Lett. 8, 403-406. logie (KULA,M. R., SCHUGERL,K., WANDREY,
OLSSON,B. (1988), A flow-injection system based C., Eds.), GBFMonographien, Vol. 9, pp. 221-
on immobilized penicillinase and a linear pH-buf- 224. Weinheim: VCH Verlagsgesellschaft.
fer for potentiometric determination of penicil- ROSS,A., SCHUGERL,K. (1988), Tetracycline pro-
lin, Anal. Chim. Acta 209, 123-133. duction by Streptomyces aureofaciens: the time
PAIGEN,K., WILLIAMS,B. (1970), Catabolite re- lag of production, Appl. Microbiol. Biotechnol.
pression and other control mechanisms in carbo- 29, 174-180.
hydrate utilization, Adv. Microb. Physiol. 4, 251, RUZICKA,J., HANSEN,E. H. (1988), Flow Injec-
PEDERSEN, K. M., HAAGENSEN, P., KUMMEL, M., tion Analysis, 2nd Ed., New York: John Wiley &
SOEBERG,H. (1985), Automated measuring sys- Sons.
tem for biological wastewater plant with biologi- SATOH,I., ARAKAWA, S., OKAMOTO,A. (1988),
cal removal of phosphate, in: Modelling and Flow-injection determination of glutathione with
Control of Biotechnological Processes, 1st IFAC amperometric monitoring of the enzymatic reac-
Symposium, Noordwijkerhout, Dec. 11-13, pp. tion, Anal. Chim. Acta 214, 415-419.
101-106 (JOHNSON,A,, Ed.). Oxford-New S C H A R P S.
~ , L., COOREMAN, W. M., BLOMME,W.
York-Toronto-Sydney-Frankfurt: Pergamon Press. J., LAEKEMAN, G. M. (1976), Quantitative en-
PETERSSON, B. A. (1988a), Amperometric assay of zyme immunoassay: Current status, Clin. Chem.
glucose and lactic acid by flow injection analysis, 22 (6), 733-738.
Anal. Chim. Acta 209, 231-237. SCHEIDEGGER, A., KUENZI,M. T., FIECHTER,A,,
PETERSSON, B. A. (1988 b), Enzymatic determina- NUESCH,J. (1988), Effect of glucose and oxygen
tion of urea in undiluted whole blood by flow in- on p-lactam biosynthesis by Cephalosporium
jection analysis using an ammonium ion-selective acremonium, J. Biotechnol. 7, 131-140.
electrode, Anal. Chim. Acta 209, 239-248. SCHELTER-GRAF, A., SCHMIDT,H. L., HUCK,H.
PLANT,A. L., LOCASCIO-BROWN, L., BRIZGYS, M. (1984), Determination of the substrates of dehy-
V., DURST,R. A. (1988), Liposome enhanced drogenases in biological material in flow-injec-
flow injection immunoanalysis, Bio/Technology, tion systems with electrocatalytic NADH oxida-
March, 266-269. tion, Anal. Chim. Acta 163, 299-303.
QUEENER, S. W., WILKERSON, S., TUNIN,D. R., SCHINDLER,J. G., SCHINDLER,M. M. (1983),
MCDERMOTT,J. P., CHAPMAN,J. L., NASH, Bioelektrochemische Membran-Elektroden. Ber-
C., PLATT,C., WESPHELING, J. (1984), Cepha- lin: Walter de Gruyter.
losporin C fermentation: Biochemical and regu- SCHMID,R. D., GUIBAULT,G. G., KARUBE,I.,
latory aspects of sulfur metabolism, in: Biotech- SCHMIDT,H. L., WINGARD, L. B. (Eds.) (1987),
nology of Industrial Antibiotics, pp. 141-170 Biosensors. International Workshop 1987. Wein-
(VANDAMME, E. J., Ed.). New York: Marcel heim: VCH Verlagsgesellschaft.
Dekker, Inc. SCHMIDT,W. J., MEYER,H. D., SCHUGERL,K.,
RANDALL,L. G., WAHRHAFTIG,A. L. (1978), KUHLMANN, W., BELLGARDT,K. H. (1984),
Dense gas chromatography/mass spectrometer On-line analysis of fermentation media, Anal.
interface, Anal. Chem. 50, 1703-1705. Chim. Acta 163, 101-109.
RECHNITZ,G. A. (1988), Biosensors, Chem. Eng. SCHMIDT,W. J., KUHLMANN, W., S C H ~ G E RK.
L,
News, Sept. 5 , 24-36. (1989, Automated determination of glucose in
RECKTENWALD, A., KRONER,K. H., KULA, M. R. fermentation broths with p-hydroxy-benzoic-acid
(1985a), On-line monitoring in downstream proc- hydrazide (p-HBAH), Appl. Microbiol. Biotech-
essing by flow injection analysis (FIA), Enzyme nol. 21, 78-84.
Microbiol. Technol. I, 607-612. SCHMIDT,W. J., AZZOPARDI, G., GRABOSCH, M.,
RECKTENWALD, A., KRONER,K. H., KULA,M. R. HEROLD,T., SCHUGERL,K. (1986), On-line
178 5 On-Line Analysis of Broth
analysis and control of production of antibiotics, TECATORAB, Brochure: FIAstar for Flow Injec-
Anal. Chim. Acta 213, 1-9. tion Analysis. P.O. Box 70, S-263 01 Hoganas,
SCHNEIDER, H. J., FOLDI,P. (1986), Aminosauren- Sweden.
Analyse, I und 11, GIT Fachz. Lab. 8, 783, 873- THOMPSON,R. Q., KIM, H., MILLER,C. E. (1987),
889. Comparison of immobilized enzyme reactors for
SCHUURS,A. H. W. M., VAN WEMEN, B. K. flow injection systems, Anal. Chim. Acta 192,
(1978), Immunoassay using antigen-enzyme con- 165-172.
jugates, FEBS Lett 15, 232-236. TOMLINSON,E., JEFFERIS,T. M., RILEY,C. M.
SCHUGERL,K. (1988), On-line analysis and control (1978), Ion pair high-performance liquid chroma-
of production of antibiotics, Anal. Chim. Acta tography, J. Chromatogr. 159, 315-358.
213, 1-9. TURNER,A. P. F., KARUBE,I., WILSIN G. S.
SHAFER,K. H., GRIFFITS,P. R. (1983), On-line su- (Eds.) (1987), Biosensors: Fundamentals and Ap-
percritical fluid chromatography/Fourier trans- plications. Oxford: Oxford University Press.
form infrared spectrometry, Anal. Chem. 55, VALCARCEL, M., LUQUEDE CASTRO,M. D., LA-
1939-1 942. Z A R ~ F.,
, RIOS, A. (1989), Multiple peak record-
SKALARANALYTIKA,Brochure, Breda, The Neth- ings in flow injection analysis, Anal. Chim. Acta
erlands. 216, 275-288.
SMALL, H., STEVENS,T. S., BAUMANN,W. C. VALERO,F., POCH,M., SERRA,A., BARTROLI,J.,
(1975), Novel ion-exchange chromatographic MEDINA,M. J . (1988), On-line monitoring of
method using conductometric detection, Anal. glucose fermentation by flow injection analysis,
Chem. 47, 1801. in: DECHEMA Biotechnology Conferences, Vol.
SMITH,A. (1985), Cephalosporins, in: Comprehen- 2, pp. 103-115. Lectures at ACHEMA '88,
sive Biotechnology. The Practice of Biotechnolo- Weinheim-Basel-Cambridge-NewYork: VCH.
gy. Current Commodity Products, Vol. 3, pp. VORLAENDER, K. 0. (1980), Diagnostik unter Ver-
163-185 (BLANCH,H. W., D ~ E wS., , WANG,D. wendung immunologischer Methoden. Stuttgart:
I. C., Eds.). Oxford: Pergamon Press. Georg Thieme Verlag.
SMITH,Jr., F. C., CHANG,R. C. (1983), ThePrac- VOSER,W. (1966), Verwendung flussiger Ionenaus-
tice of Ion Chromatography, New York: Wiley- tauscher in der automatischen Analyse organi-
Interscience. scher Verbindungen mit dem Technicon Auto-
SNYDER,L. R., KIRKLAND,J. J . (1979), Zntroduc- Analyser, in: Automation in the Analytical Che-
tion to Modern Liquid Chromatography, 2nd mistry, Paris, 2.-4. Nov., pp. 47-52.
Ed. New York: John Wiley & Sons. WARWICK,T. L. (1985), Culture monitoring sy-
SONNLEITNER, B., KAPPELI, 0. (1986), Growth of stem, Eur. Patent Application, Appl. No.
Saccharomyces cerevisiae is controlled by its lim- 85308834.2, Publ. No. 0 184 441, Date filing:
ited respiratory capacity: Formulation and verifi- 04.12.85 Appl. Eli Lilly & Company, Indianapo-
cation of a hypothesis, Biotechnob Bioeng. 28, lis, IN, USA.
921-937. WAZAUCo. Bio-filtrator, Bauart TU Berlin, Bro-
STULTS,C. L. M., WADE, A. P., CROUCH,S. R. chure. Dr.-Ing. Georg Wazau, Keplerstr. 12, D-
(1987), Computer-assisted optimization of an im- 1000 Berlin.
mobilized enzyme flow-injection system for the WEISS,J. (1985), Handbuch der Ionenchromatogra-
determination of glucose, Anal. Chim. Acta 192, phie. Weiterstadt, FRG: Dionex GmbH.
155-163. WENTZ, D. (1989), Dissertation, Universitat Han-
SWINDLEHURST (Koerner), C. A., NIEMAN,T. A. nover, FRG.
(1988), Flow-injection determination of sugars WHITE, R. L., ROBERTS,D. E., ATTRIGE,M. C.
with immobilized enzyme reactors and chemilu- (1985), Fourier transform infrared detection of
minescence detection, Anal. Chim. Acta 205, pyruvic acid assimilation by E. coli, Anal. Chem.
195-205. 51, 2487-2491.
SZIGETI,L., P k s , M., LOVREZ,G., PUNGOR,Jr., WONG,J. S., REIN, A. J., WILKS,D., WILKS,Jr.,
E., NYESTE,L., HOLLO, J . (1986), Measurement P. (1984), Infrared spectroscopy of aqueous anti-
and control of phosphate concentration of fer- biotic solutions, Appl. Spectrosc. 38, 32-35.
mentation processes, Hung. Sci. Instrum. 61, 57- YAO, T., WASA, T. (1988), High-performance liq-
64. uid chromatographic detection of L- and D-ami-
TAWFIK,A. M., MARDON,C. J. (1985), Automated no acids by use of immobilized enzyme electrodes
analytical method for the determination of indi- as detectors, Anal. Chim. Acta 209, 259-264.
vidual sugars in mixtures of glucose, fructose and YELLOW SPRINGSINSTRUMENTS, YSI 23A Bro-
sucrose, J. Sci. Agric. 35, 621-627. chure.
6 Determination of Cell
Concentration
and Characterization of Cells
KENNETHF. REARDON
Fort Collins, Colorado 80523, U.S.A.
THOMASH. SCHEPER
Hannover, Federal Republic of Germany
1 Introduction 181
2 Biomass Concentration 181
2.1 Optical Sensors 182
2.1.1 Nephelometric Methods 182
2.1.2 Fluorescence Sensors 185
2.2 Calorimetric Methods 188
2.3 Filtration Methods 190
2.4 Viscosity Methods 192
2.5 Electrochemical Methods 192
2.5.1 Impedimetric Methods 192
2.5.2 Potentiometric Sensors 193
2.5.3 Fuel Cell Type Sensors 193
2.5.4 Amperometric Sensors 194
2.6 Acoustic Methods 195
3 Cell Components and Characteristics 195
3.1 Cell Morphology 200
3.1.1 Coulter Counter 200
3.1.2 Flow Cytometry 201
3.1.3 Other Instruments 201
3.2 Viability 201
3.2.1 Flow Cytometry 202
3.2.2 Sample-Flow Analysis 202
3.3 Intracellular pH 202
3.3.1 Flow Cytometry 203
3.3.2 Nuclear Magnetic Resonance Spectroscopy 203
3.3.3 Sample-Flow Analysis 203
180 6 Determination of Cell Concentration and Characterization of Cells
c
wide variety of sensor types for application in v) / / l R W sratterinn
biotechnology . These sensors are non-invasive
and the response times are nearly instanta-
neous. The use of glass fiber technology makes
these sensors small, robust, and reduces their
cost. These advantages have led to the devel- V \
opment of a large number of commercially Cell concentration
available optical biomass sensors. Fig. 3. Sensor signals as a function of cell mass con-
centration for three different turbidity measurement
methods.
2.1.1 Nephelometric Methods
Nephelometric methods monitor the re-
sponse of turbid samples to light signals. In gy (VANOUS,1978). Most of the incident light
general, turbidity results from suspended solid is scattered in the forward direction. Different
particles in the liquid and is largely dependent nephelometer designs can be used to measure
on the physical size and concentration of the turbidity (Fig. 2). Although it is not the most
particles. In biological systems, suspended sensitive for concentration variations, the 90"
cells, solid particles, and air bubbles are the scattered light is usually measured because it is
major causes of turbidity. It is therefore not less sensitive to variations in particle size. The
possible to distinguish between viable and non- measurement of light scattering at other angles
viable biomass with nephelometric methods. is generally used for particle-size analysis.
Fig. 1 shows the scattering behavior of par- Most optical sensors in biotechnology meas-
ticles that are of sizes relevant for biotechnolo- ure transmitted light. The amount of transmit-
Biomass Concentration 183
ted light is reduced by the light scattering of biomass concentrations because the transmit-
the particles, and thus the turbidity can be ted light intensity decreases rapidly. The meas-
measured at low optical densities. At very high urement of transmitted light is only accurate
cell densities, both transmitted and scattered up to extinction values of about 0.5 (KOCH,
light are absorbed inside the measuring volume 1970). As shown in Fig. 1, most of the incident
(Fig. 3). This “inner filter” effect makes all de- light is scattered in the forward direction. At
tection designs other than retroreflective moni- higher concentrations, part of the scattered
toring of scattered light ineffective for high light will be scattered again by other cells and
biomass concentrations. The accurate use of reoriented so that it reaches the detector unit.
nephelometric techniques requires the correct These effects are responsible for deviations
application of light scattering analysis (includ- from the Lambert-Beer law (KOCH, 1970).
ing the limitations of the technique) as well as Higher biomass concentrations can be moni-
proper interpretation of the data obtained. tored by changing the length of the light path
As previously mentioned, the measurement or by diluting the sample. For continuous
of transmitted light can be used for cell mass monitoring, sample dilution would require a
estimation at low cell concentrations (below 1 waste of fermentation broth or a recycle of di-
g/L wet weight) (HANCHERet al., 1974). In luted broth. In order to avoid this, a side
the region of these low particle concentrations, stream from the cultivation can be pumped
the Lambert-Beer law is valid: through a flow cell with a variable light path
and recycled to the fermentor without dilu-
log Io/I=k.l*c (1) tion. Fig. 4a shows a schematic diagram of
such a flow cell (LEE and LIM, 1980) in which
where I is the length of the light path through the light path is adjusted by varying the diam-
the sample and k is the extinction coefficient. eter of the water-filled inner tube. With this
The ratio between the incident light (I,,) and device (outer diameter (0) 12 mm; inner diam-
the transmitted light (I) is commonly known as eter (d)8 mm), extinction measurements were
the optical density (OD). directly proportional to dry weight masses up
The ratio of scattered to transmitted light is to 2.0 g/L. This also shows the limitations of
also measured occasionally. As shown in Fig. nephelometric methods to relatively low bio-
3, this method becomes ineffective at higher mass concentrations.
Sample out
Optical
density
Here, forward-scattered light and transmitted Fiberoptic bundles are used to conduct the
light are measured at the same time. This flow- light signals, resulting in a small sensor that
through sensor must be placed in a fermentor fits into standard electrode ports of fermentors
bypass loop. By using the ratio of transmitted (MERTENet al., 1987).
to scattered light, the linearity of signal and
cell concentration should be better for this
class of sensors (VANOUS,1978). 2.1.2 Fluorescence Sensors
Another turbidometric sensor is described in
detail by HANCHERet al. (1974). In this case, Another important group of optical sensors
a retroreflective turibidometer unit was used in for biomass concentration is on-line, in situ
a flow-through cell coupled to a fermentor. By fluorosensors. During the last few years, these
measuring the back-scattered light, wet bio- sensors have gained importance for biotechno-
mass concentrations in the range of 1-60 g/L logical purposes not only for biomass estima-
for Escherichia coli cultivations could be mon- tion, but also for reactor characterization
itored. The principle of a commercial retrore- (EINSELEet al., 1978, BEYELERet al., 1983,
flective sensor (Aquasant) is shown in Fig. 7. GSCHWENDet al., 1983, SCHEPERand SCHU-
GERL,1986a), process monitoring (e.g., meta-
bolic state of the cells, see Sec. 3.6.1), and
process control (MEYERand BEYELER,1984).
The principle of these sensors is based on
the fluorescence studies of DUYSENS and
AMESZ(1957). They demonstrated that the re-
duced adenine dinucleotides (NADH and
NADPH) in living cells fluoresce at a wave-
length of about 460 nm when irradiated with
340-360 nm (UV) light. By measuring this
emitted light, the NAD(P)H pool inside living
microorganisms can be monitored. The inten-
sity of fluorescence is affected by the amount
of viable biomass, the metabolic state of the
Fig. 6 . Monitek turbidity probe in which forward- cells, and also by abiotic factors (e.g., bubbles
scattered and transmitted light are measured simul- and fluorescent medium components).
taneously. 1 Light source, 2 optical system, 3 me- Originally, large, complicated fluorometer
dium, 4 optical system, 5 transmission light detec-
tor, 6 scattering light detector, 7 light trap (re-
devices were interfaced with fermentors. Inci-
printed with permission of Monitek, Morsenbroi- dent as well as emitted light passed through the
cher Weg 200, D-2000 Diisseldorf 30). fermentor walls via quartz windows (HARRI-
SON and CHANCE, 1970; ZABRISKIEand
HUMPHREY,1978). During the next several
years, studies were performed with this type of
Incident light fluorometer equipment. For example, HARRI-
SON and CHANCE(1970) studied aerobic-anae-
robic transitions of cultivations of Klebsiella
aerogenes, and ZABRISKIEand HUMPHREY
(1978) showed that biomass estimation is pos-
sible on the basis of culture fluorescence. Stud-
Scattering light ies on feeding strategies for fed-batch cultures
of Candidu utilis were performed with this de-
Reflected light vice by RISTROPHet al. (1977), while EINSELE
Fig. 7. Schematic design of the Aquasant retrore- et al. (1978) used it for substrate uptake and
flective turbidity probe (reprinted with permission mixing time studies in bioreactors. All of these
of Aquasant Mentechnik AG, Hauptstrane 20, CH- studies clearly showed the potential of this
4416 Bubendorf). method.
186 6 Determination of Cell Concentration and Characterization of Cells
Excitation light
I I 1bO mm I
Cell
t MP
Fig. 8. Schematic design of a fluorescence probe for the simultaneous detection of two different
wavelengths. M mirror, L lamp, L1-3 lenses, F1-4filters, FR1-4 fiber cables, D1.3photodetectors,
H housing, MZ mixing zone of fibers, QP quartz plate, A amplifier, C cooling, MP microproces-
sor unit, PS power supply.
and adjusted to one single type of fermentor, of 1 g/L was followed by an increase of fluo-
they were able to show that the on-line moni- rescence intensity of about 21 mV. However,
toring of biotechnological processes on the ba- the intracellular metabolism changed com-
sis of the redox state of cells was possible. ZA- pletely at glucose concentrations of 15%. Un-
BRISKIE and HUMPHREY (1978) used Culture der these conditions, the NAD(P)H values dif-
fluorescence as an on-line estimator of viable fered totally from those at lower glucose con-
biomass during cultivation of Saccharomyces centrations. A linear increase of culture fluo-
cerevisiae, a species of Streptomyces, and a rescence with biomass was also found by
species of Thermoactinomyces. The culture LUONGand CARRIER(1986) for cultivations
fluorescence and biomass data could be linear- of Methylomonas mucosa and by BOYERand
ized by the following equation (see Fig. 9) (ZA- HUMPHREY (1988) for cultivations of Pseudo-
BRISKIE and HUMPHREY, 1978): monas putida.
Most NAD(P)H-fluorescence studies have
been performed in suspended cell culture. Do-
RAN and BAILEY (1987), REARDONet al.
During the next few years, several other au- (1986), and MOLLERet al. (1988) showed that
thors reported on the estimation of biomass the monitoring of culture fluorescence could
concentration from culture fluorescence data also be applied to immobilized cell systems.
(Tab. 1). A strict linear relation between bio- The growth of Clostridium acetobutylicum
mass and culture fluorescence was found for and Saccharomyces cerevisiae immobilized in
the growth of Zymomonas mobilis under non- calcium alginate was observed by these investi-
limited conditions (SCHEPERet al., 1987b). gators. Although no accurate calibration of
Fig. 10 show the results for cultivation at dif- fluorescence signal and biomass content could
ferent glucose levels. An increase of biomass be performed, the authors showed that the in-
Organism References
o 2 ‘1. glucose
x 6 % glucose
10 % glucose
A 15 ‘/a glUCOSe
Fig. 10. Culture fluorescence signals as a
function of biomass concentration at var-
ious glucose concentrations during cultiva-
Biomass ( g l l ) tion of Zymomonas mobilis.
can be caused by air bubbles, which lower the consumption via different catabolic pathways
sensor readings while displacing cell-contain- (LUONGand VOLESKY,1983). Different calo-
ing medium in front of the sensor’s observa- rimetric devices (external-flow microcalorime-
tion window. In addition, fluorescent medium ter, twin-type, and heat-flux calorimeter) and
components, especially in fed-batch cultiva- different calorimetric techniques (dynamic or
tions, affect the signal, as d o drastic changes continuous calorimetry) have been used to
in the metabolic state of the cells. Wall growth monitor heat evolution during cultivation
on the sensor’s observation window has not processes.
been found by any author. In general, accurate Most problems with this technique arise
biomass estimation should only be possible when the heat produced by the microorga-
when the NAD(P)H pool per cell is constant nisms must be separated from other heat-pro-
during the whole cultivation process. Howev- ducing or heat-consuming processes (e.g., aer-
er, under controlled conditions an accurate es- ation, stirring, and addition of nutrients or al-
timation is still possible even when metabolic kali/acid). An overall heat balance of the
changes occur, e.g., for aerobic yeast cultiva- whole fermentor/analysis system is necessary
tions, in which the metabolism changes from (LUONGand VOLESKY,1983).
oxidative-reductive to purely oxidative (ZA- In earlier calorimetric studies in biotechno-
BRISKIE and HUMPHREY,1978; see Fig. 9). logy, small (5-50 mL volume), but well-de-
The non-invasive monitoring of culture flu- fined calorimetric devices were used (DER-
orescence offers interesting insights into bio- MOUN and BELAICH,1979; ISHIKAWAet al.,
technological processes. The on-line estimation 1981; ISHIKAWAand SHODA, 1983). These
of viable biomass is only one of them. Al- twin-type heat conduction calorimeters utilize
though limitations for general usage as an on- two similar vessels. One is run as a cultivation
line biomass sensor d o exist, several applica- vessel (with microorganisms) and the other
tions to cultivation monitoring have shown the functions as the reference under the same con-
high analytical potential of this technique. ditions (without microorganisms). This tech-
Biomass Concentration 189
nique makes it possible to measure the heat mentor. Heat losses during pumping and wall
evolved by the cells without the influence of growth can also become problems.
the aforementioned disturbances. However, Dynamic calorimetry is another technique
sampling is nearly impossible for batch proc- for performing calorimetric measurements on
esses and thus large fermentors must be run in bench-scale fermentors (Mou and COONEY,
parallel as reference. 1976; WANGet al., 1978). Here, the heat that
Calorimetric techniques can be applied to is released during cellular metabolism is meas-
real fermentation processes by recirculating ured by the increase in temperature of the fer-
fermentation broth through an external flow mentation broth when the temperature control
microcalorimeter where the evolved heat is system is turned off. The heat accumulation is
monitored (DJAVAN and JAMES, 1980; NI- corrected for heat loss and gain in the fermen-
COLS and JAMES, 1981; ORIOLet al., 1987; tor system. Although real-time, on-line moni-
SAMSON et al., 1987; ROY and SAMSON, toring is not possible and the fermentation
1988). Problems with the flow calorimeter conditions are not strictly isothermic, this sim-
might arise from changes in cellular metabo- ple method has been successfully applied to
lism while the medium is pumped through the various bench-scale cultivations (Mou and
loop, especially if oxygen or nutrients are ex- COONEY,1976; WANG et al., 1978).
hausted during transport. In this case, the LUONGand VOLESKY(1980, 1982) have de-
broth inside the flow cell might not represent scribed the use of continuous calorimetry to
the actual state of cultivation inside the fer- study heat evolution in cultivations in larger
Air pre-
sat uratior
unit
--
I;
-
Air
in
mass-
Fig. 11. Schematic design of a “heat flux” calorimeter (reprinted with permission of MARISONand VON
STOCKAR,1986).
190 6 Determination of Cell Concentration and Characterization of Cells
fermentors (up to 14 L). Here, the fermenta- complete system and a heat balance are no
tion broth was overcooled using a defined longer necessary.
cooling water flow rate in the fermentor jack- Fig. 12 shows that the monitoring of heat
et. Heat produced by the microorganisms and evolution can be used for the on-line measure-
by a separately-controlled electrical immersion ment of biomass concentration (MARISONand
heater was responsible for the temperature of VON STOCKAR,1986, 1987; BIROU et al.,
the fermentor liquid. The temperature was 1987; BIROU and VON STOCKAR, 1989).
kept constant by controlling the immersion Evolved heat corresponded closely with bio-
heater. An overall heat balance was employed mass concentration. However, changes in the
to determine the amount of heat evolved by metabolic state of the cells caused by medium
the microorganisms (LUONG and VOLESKY, limitations or diauxies resulted in drastic
1980, 1982; LUONGet al., 1983). changes in the thermograms (MARISONand
The system shown in Fig. 11 is called a VON STOCKAR,1987). In such cases, the rela-
“heat-flux calorimeter” (MARISONand VON tionship between biomass and evolved heat is
STOCKAR,1986). A 1.8 L fermentor is temper- quite different.
ature controlled by a silicone-oil circulation
system composed of two different subunits.
These are connected via an electronic valve 2.3 Filtration Methods
that is controlled by a computer. One subunit
is run at a slightly higher temperature (electri-
cally heated) than desired for the cultivation It is well known that cells can be separated
process, while the other subunit is run at a from the medium by filtration. Thus, biomass
lower temperature. The reactor temperature is concentration can be estimated from the filtra-
monitored continuously and kept constant by tion cake volume, the filtation flow rate, the
computer-controlled mixing of both flows. A filtrate volume, and the pressure difference
rise in the reactor temperature, brought about across the filter system during the filtration
by heat evolution of the microorganisms, is process (SILVENNOINEN and KOIVO, 1982).
followed by a temperature decrease in the NESTAASand WANG(1981) demonstrated that
jacket oil. The temperature difference that is the biomass concentration can be calculated
necessary to overcome heat-transfer resistance from:
between the temperature controlling fluid in TI
in which X (g/L) is the biomass concentration, 105s (NESTAASet al., 1981, NESTAASand
v (L/g) is the specific cake volume, V, (L) is WANG,1983). After filtration, the filtrate and
the cake volume, and V, (L) is the filtrate vol- the filter cake volume were recycled to the fer-
ume. The authors showed that an automatical- mentor in order to minimize the loss of fer-
ly-operated filtration unit could be coupled to mentation broth. The probe had to be cleaned
a Penicillium chrysogenum cultivation for after each filtration, resulting in a sampling cy-
semicontinuous estimates of biomass concen- cle time of 30 min. A similar device was used
tration (NESTAASand WANG, 1981; NESTAAS by THOMASet al. (1985) for biomass calcula-
et al., 1981). Samples of 50-70 mL were with- tion during cultivation of P. chrysogenum.
drawn from the fermentor, loaded into the fil- The correlation of off-line data and the auto-
tration probe, and filtered under constant matic filtration unit was very good in the range
pressure. The filtrate volume could be meas- of 2-40 g/L biomass; however, problems arose
ured gravimetrically using a load cell, while the at lower biomass concentrations.
cake volume was determined optically (light In order to overcome difficulties with the
transmission). The filter cake volume can also backflushing of filter media and with filter
be measured by the pressure drop across the contamination, LENZet al. (1985) constructed
system (NESTAASet al., 1983). The measure- a computer-controlled filtration probe with re-
ment of the filter cake volume as a function of newable filters. The principle of one analysis
filtration time is therefore possible with this cycle is shown in Fig. 13 (REUSSet al., 1987).
device. When the specific filter cake volume is During sampling, the filtration unit was filled
nearly constant during cultivation, biomass with fermentation broth (approximately 100
can be measured accurately with this filtration mL). The measuring cycle was then started by
probe. The filtration time for one sample was applying a constant filtration pressure. Five in-
Filter band
I CI+ Pump
Waste
Balance
Fig. 13. Schematic diagram of one operation cycle of a computer-controlled filtration probe (reprinted with
permission of REUSSet al., 1987a).
192 6 Determination of Cell Concentration and Characterization of Cells
Biomass concentration
Penicillium chrysogenum Laboratory measurements
o Filtration measurements
1 1
I I I I I I I I I I I I
-
= I
10 I I 1 I I I I I I
1 I
I I
20 30 10 50 60 70 80 9
Time f i h ]
Fig. 14. Comparison of on-line filtration and off-line measured biomass data during cultivation of Penicil-
lium chrysogenum (reprinted with permission of REUSS et al., 1987b).
dependent retroflective light barriers were used et al., 1976). PERLEY et al. (1979) showed that
to monitor the growth of the filter cake, thus a capillary-type viscosimeter in a continuous
increasing the analytical accuracy. After the sample flow from the fermentor could be used
measuring procedure, the filtration unit was for monitoring broth viscosity. The viscosity
cleaned with water and dried with air while the of cell suspensions from cultivations of Hanse-
filter was renewed. Fig. 14 shows that the off- nula polymorpha increased in a non-linear
line gravimetric data correlated very well with manner with cell mass concentration. The ef-
data of the filtration probe (REUSS, 1987). In fect of cell mass on broth viscosity was small,
addition, the filtrate could be used for analysis and therefore biomass estimates from viscosity
of the cell-free medium. Although filtration measurements were rather inaccurate. The ap-
probes are relatively complex devices and exact plication of an in-line rotational viscometer
on-line monitoring of biomass concentration is for baker's yeast (SHIMMONSet al., 1976)
not possible with them, the technique could be showed better results, although these experi-
successfully applied to bioprocess monitoring ments were performed using yeast cells sus-
during cultivations of P. chrysogenum (NES- pended in buffer solution. No real data on
TAAS and WANG, 1981, 1983; NESTAASet al., growth have been supplied.
1981; LENZet al., 1985; THOMAS et al., 1985;
REUSS, 1987; REUSS et al., 1987) and to the
Pekilo process (SILVENNOINEN and KOIVO, 2.5 Electrochemical Methods
1982).
Different electrochemical methods can be
used for the estimation of biomass concentra-
2.4 Viscosity Methods tion (RAMSAYet al., 1985). Here, four of
these methods for on-line analysis will be re-
The viscosity of a fermentation broth is af- viewed.
fected by biomass concentration, among other
factors. Therefore, the cell concentration can
be derived from viscosity measurements when 2.5.1 Impedimetric Methods
no other influence such as extracellular mate-
rial, cell morphology changes, or temperature When a sinusoidal electrical field is applied
variations complicate this process (SHIMMONS to two electrodes between which a cell-contain-
Biomass Concentration 193
ing sample is placed, the impedance of the bio- 300 kHz frequency). A version of this type of
logical sample can be measured. Two paramet- in situ, real-time biomass sensor is commer-
ers, E, the electrical permittivity (for capaci- cially available (Aber Instruments Ltd); this
tance) and 0, the conductivity (for conduc- device measures the radio frequency (100 kHz-
tance) are necessary to describe the electrical 10 MHz), conductance and capacitance during
properties of this sample. Both are affected by cultivation processes.
the biomass concentration. In a carefully tem-
perature-controlled system under defined
growth conditions (e.g., medium composi- 2.5.2 Potentiometric Sensors
tion), the measurement of the conductivity can
be used for biomass monitoring. Since instru- This kind of sensor measures the potential
mentation (e.g., from Malthus Instruments that is developed at an electrode in a cell con-
Ltd, Bactomatic, or Bactobridge) is often taining a sample with respect to a reference
quite expensive, these methods are used in the electrode. The growth of microorganisms and
routine control of microbial contamination for the associated production of electroactive sub-
large numbers of off-line samples for clinical stances cause changes in the potential. WIL-
or food industry analyses (CADY, 1975; RAM- KINS et al. (1974) described such a biomass
SAY et al., 1985). Only a small number of re- sensor based on the electrochemical detection
ports on the analysis of higher biomass con- of hydrogen evolved by the microorganisms.
centrations in these commercially available in- This test system consisted of a platinum elec-
struments have been published. One of these trode and a reference electrode.The hydrogen
reports (MUNOZand SILVERMAN,1979) con- evolved was detected by a voltage increase in
tains data on measurements of Escherichia coli the cathodic direction. WILKINS(1978) and
concentrations in sewage effluents in the range WILKINSet al. (1978) showed that a wide vari-
of lo6 to 10’ cells per mL. ety of microorganisms could be measured us-
Routine conductivity meters can also be ing this method. It is very sensitive: one cell
used for the determination of biomass. An per mL could be detected (with a detection
example is given by TAYA et al. (1989) for time of 9 h) (WILKINS,1978). However, these
plant tissue cultures. A linear relation between sensors appear to be restricted to relatively low
conductivity and cell mass was found for bio- biomass concentrations, and a long analysis
mass concentrations up to 10 g/L for Coffea time is necessary (e.g., for Gram-negative
arabica, Nicotiana tabacum, Withania sonni- cells, lo6 cells/mL could be measured in 2 h)
fera, and Catharanthus roseus. In addition, (WILKINS,1978; WILKINSet al., 1978; JUN-
BLUTEet al. (1988) used conductivity measure- TER et al., 1980). Microorganisms as well as
ments for on-line cell mass estimation in hol- electroactive substances affect the sensor read-
low-fiber bioreactors. ings.
The measurement of the frequency-depend-
ent permittivity seems to be more promising
for biotechnological purposes. Simple devices 2.5.3 Fuel Cell Type Sensors
for off-line analysis have shown the potential
of this technique for monitoring biotechnolog- The principle of this method is shown in
ical processes (GENCERand MUTHARASAN, Fig. 15 (MATSUNGAet al., 1979; BENNETTOet
1979). HARRISet al. (1987) clearly showed that al., 1987). Two fuel cells are inserted into the
the permittivity at low radio frequencies (be- cell containing sample. The anodes of both
low 1 GHz) has a linear correlation with bio- cells are in contact with the sample. The cur-
mass concentrations in cell suspensions. A rent produced in probe I results from the oxi-
small probe suitable for use in standard elec- dation of microorganisms and electroactive
trode ports was used to measure the permittivi- substances, while the current of reference
ty and conductivity in fermentation broths. probe I1 results only from the oxidation of
KELL(1987) and HARRISet al. (1987) reported electroactive substances, because microorgan-
a good linear correlation between capacitance isms cannot permeate through the dialysis
and biomass in yeast cultures up to 100 g/L (at membrane (MATSUNGAet al., 1979). Thus,
194 6 Determination of Cell Concentration and Characterization of Cells
-Stirrer
ious types of microorganisms.
W
I
y v 1
A i:L: f i
Redox systems 2,l-Oichlorophenol-
Product $I in c e l l membrane
Redd;;d
Fig. 16. Principle of a mediated electron transfer between the redox systems in cell membranes and a
fuel-cell type sensor (adapted from NISHIKAWA et al., 1982).
Cell Components and Characteristics 195
'
probes (each containing a working, counter,
and reference electrode). One set was covered
with a dialysis membrane to measure the cell-
free medium as a reference. The measuring
times were in the range of 5 to 10 min. Again,
mediators increased the sensitivity of this am-
Plastic Ge
t
Silicon rubber I' Piezoelectric membrane
process development and control. These con- contribute to the overall behavior of a cultiva-
siderations are especially significant for culti- tion. In addition, it is possible to sort the sam-
vations of genetically modified cells, in which ple on the basis of the measured parameter to
the product often accumulates inside the or- acquire relatively homogeneous cell popula-
ganism. Although off-line techniques are, in tions.
many cases, readily available, they are far less The sample stream that enters the flow cy-
desirable for process development because the tometer must be a suspension of single cells
measured quantities are not necessarily those that contains as little debris as possible. Both
that would be found in situ. The metabolism of these requirements make the direct interface
and physiology of microorganisms have been of a flow cytometer with a fermentor some-
observed to change very rapidly in response to what problematic, although suitable pre-treat-
alterations in the cellular environment. The ment steps can be envisioned. In a typical cy-
use of off-line measurements is nearly worth- tometer (Fig. 18), the cell suspension is intro-
less for process control, since the time scale of duced into the center of a flow of sheath fluid
the changes occurring in the bioreactor, and in such a way that the cells flow one at a time
hence the required time scale of the control re- past a focused light source, usually a laser or
sponse, is often much shorter than the time mercury arc lamp. The light beam that hits a
needed for analysis. cell is scattered and absorbed, and, if the cell
Unfortunately, very few techniques or sen- has been stained, it is re-emitted as fluorescent
sors for truly on-line measurement exist, and light. The resulting light signals are collected
even fewer are commercially available. Thus, by various lenses and measured with photo-
the scope of this section has been expanded to multipliers or photodiodes, from which the
include instruments or methods that are quasi- output can be analyzed and processed.
on-line (i.e., directly interfaced with the bio- Flow cytometry has been used for many
reactor but providing non-continuous data) or years in medicine and biology, and many dif-
that seem applicable to on-line analysis. In ad- ferent cellular parameters have been measured
dition, the focus is on cultivations of sus- (Tab. 2). Although most of these assays re-
pended cells, primarily because most work has quire treatment with dyes and other reagents
involved such systems. and add some complications to automated
As a rapid perusal of the following pages re- bioreactor analysis, the wide spectrum of de-
veals, three instrument systems are capable of tailed measurements possible makes flow cy-
measuring many cellular characteristics of in- tometry a potentially powerful quasi-on-line
terest: flow cytometry, nuclear magnetic re- technique. Details on instruments, methods,
sonance spectroscopy, and systems based on and data analysis have been presented in a
sample-flow analysis (flow injection analytical number of reports (e.g., HORANand WHEE-
devices and autoanalyzers). Therefore, it is LESS, 1977; KRUTH, 1982; STEINKAMP, 1984)
useful to give a brief overview of each of these and books, including an excellent text by SHA-
techniques before discussing the measurement PIRO (1988).
of individual cellular quantities.
J Amplifier
I
-
Plotter )c Pulse height
analyzer *
Mainframe
computer CRT display
198 6 Determination of Cell Concentration and Characterization of Cells
Components Characteristics
as well as the local magnetic field to which the gations, because its spectra are simpler and yet
nucleus is exposed. Thus, the phosphorus in still able to provide information on a number
ATP can be distinguished from that in of important phosphate compounds (Fig. 19).
NADH. The frequency spectra are usually ex- In addition, 31Phas the advantage of being the
pressed relative to the frequency of a reference natural isotope.
compound by a quantity called the chemical Since NMR spectroscopy is a non-invasive,
shift. NMR spectra contain a great deal of in- non-destructive technique, it is a powerful
formation, and the technique has been used method to study fundamental physiological
with success for many years on solids, includ- and metabolic phenomena in vivo. The major
ing tissue samples. drawback of NMR analysis of microorganisms
Tab. 3 presents some of the nuclei that have is the low sensitivity of the technique, requir-
been used in NMR experiments on biological ing the use of high-density samples. This can
systems. From all of the listed components, lead to further problems with maintenance of
31Pis most commonly used for in vivo investi- cell viability. However, the power of this tool
Cell Components and Characteristics 199
Sample-Flow Analysis
This term refers not to an instrument but
rather to a type of analytical system. Many de-
tectors cannot be utilized directly in a fermen-
tor vessel because of probe fouling, chemical
interference of medium components, electrical
interference, calibration requirements, and be-
cause they cannot be sterilized. Other analyti-
cal techniques may require the addition of rea-
gents to the sample. All of these problems can,
in theory, be overcome by analyzing a sample
flow withdrawn from the bioreactor.
This type of system offers several advan- I -t
tages. Not only are the previously mentioned Fig. 20. Top: Scheme of a two-line injection system.
problems avoided, but calibration and probe C carrier, S sample, R reagent, RC reaction coil, D
maintenance are greatly simplified. There are detector, W waste.
also some potential problems associated with Bottom: Typical detector readout. “S” denotes the
the use of these systems, among them a higher samDle iniection (adaDted
. . from RUZICKA and
risk of culture contamination and the possibili- HANSEN,i988).
200 6 Determination of Cell Concentration and Characterization of Cells
vironment to which the sensor is exposed can cell crosses the field, and thus the voltage drop
be manipulated as desired. The main disadvan- across the aperture also increases. The increase
tage of this method is the requirement for rap- of the voltage drop is generally proportional to
id and sensitive sensor response and the dis- the fraction of the volume of the orifice occu-
continuous output of data. Still, the large pied by the cell. The detection limit of the
number of applications to the analysis of cul- commercially available instrument is about 0.5
ture broth reported in the literature (see Sect. pm, but modifications have led to detection
2.2) attest to the immense usefulness of these limits as low as 0.09 pm (DEBLOISand BEAU,
quasi-on-line FIA systems. To date, relatively 1970). Reviews of the instrument and tech-
few reports of the analysis of cell constituents nique are available (KUBITSCHEK, 1969; HAR-
or characteristics have been published, but RIS and KELL, 1985).
there are no significant obstacles to such appli- The Coulter counter technique suffers from
cations. Further details of FIA and other flow- several limitations, one of which is that cells of
sampling systems are provided in several re- all shapes are sized as spheres. Other problems
view articles (CLARKEet al., 1985; HANSEN, include false measurements caused by the si-
1988; RUZICKAand HANSEN,1988). multaneous passage of two or more cells
In the following sections, reported applica- through the orifice and the inability to discrim-
tions of these and other techniques for the on- inate between viable cells, non-viable cells, and
line, quasi-on-line, and potentially (quasi-)on- contaminating particulates. A more subtle
line analysis of a variety of cellular parameters problem that is often overlooked concerns the
are summarized. The parameters that are dis- magnitude of the applied electric field. Above
cussed are clearly not the only measurable ones a certain critical value, dielectric breakdown of
(see, for example, Tabs. 2 and 3), but are the cell membrane takes place, with the result
those of greatest general interest. that the measured cell size increases with in-
creasing field strength (JELTSCHand ZIMMER-
MA", 1979). Since most Coulter counter in-
3.1 Cell Morphology struments are capable of exceeding this critical
field strength, the actual cell size can easily be
The morphological characteristic most often underestimated. However, when adequate cau-
of interest in cell cultivation is the cell size, but tion is used, these problems can usually be
cell shape determination has also proven use- minimized or eliminated.
ful. Although these parameters may be investi- There have been many reports on the use of
gated with a microscope, several instruments the Coulter counter, but most of them involve
are available for rapid analysis, and cell size counting rather than sizing of particles. Size
distribution can also be obtained by some of determination is of interest because this pa-
them. The two mostly used instruments are the rameter can be related to the cell cycle or to an
Coulter counter and the flow cytometer. age distribution, both of which are needed to
understand the biological aspects of the cul-
ture. For example, ALBERGHINA et al. (1983)
3.1.1 Coulter Counter used Coulter counter measurements to study
the cell cycle of the budding yeast S. cerevi-
The ability of the Coulter counter to meas- siae. In another investigation, the size-discrim-
ure cell size is based on the disturbance of an inating capability of the instrument was used
electric field as it is traversed by small par- to distinguish and enumerate the different
ticles, in this case microorganisms. The cells populations in a cultivation of Tetrahymena
must be suspended in an electrically conduct- pyriformis growing on E. coli and Azobacter
ing fluid; most cultivation media would suf- vinelandii (DRAKEand TSUCHIYA,1973).
fice, although dilution may be required. This To date, no report of the quasi-on-line use
suspension is then drawn through a small ori- of a Coulter counter for monitoring of cul-
fice, across which an electric field has been ap- tures has been published, although it seems
plied. The resistance of the orifice increases possible to use a modified Coulter instrument
when a (relative to the liquid) low-conductivity for this purpose.
Cell Components and Characteristics 201
3.1.2 Flow Cytometry could clearly be seen in the flow cytometer re-
sults.
WITTRUPet al. (1988) used flow cytometry
Two types of flow cytometers have been uti- to investigate the morphological changes oc-
lized for cell-size measurements. In a single- curring in several strains of E. coli that formed
beam instrument, the intensity of narrow-an- inclusion bodies of plasmid-encoded proteins.
gle forward-scattered light resulting from the Both forward-angle (cell size) and right-angle
interaction of a cell with the focused light (internal structure) scatter intensities increased
beam can be related to the size of the cell, al- with foreign protein production and inclusion
though the difference between the refractive body formation. Thus, quasi-on-line monitor-
indices of the cell and the surrounding fluid ing of light scatter during growth of recombi-
also plays a role. Typically, the refractive in- nant cells would be both feasible and informa-
dices remain constant, and the value of the tive.
forward-scatter intensity can be used as a At this time, there have been no such re-
measure of cell size. There can be significant ports on the quasi-on-line use of flow cytome-
problems with particulate matter in the sample ters, although automatic sample injection and
fluid, and careful tuning is necessary to meas- calibration have been achieved. No pretreat-
ure size distributions in samples of small mi- ment steps are necessary for cell size determi-
croorganisms such as bacteria. nation, but the presence in the medium of par-
The second type of instrument utilizes a ticulates that are of the size range of the cells
split beam configuration. The two laser beams causes significant error.
are tightly focused at separate spots, allowing
calculation of the cell flow velocity. This vel-
ocity measurement can be combined with 3.1.3 Other Instruments
measurements of absorption (for cells larger
than 2 pm) or 90" light scatter (for bacteria or Nephelometric instruments are also able to
other cells smaller than 2 pm). measure cell-size distribution; these devices
The sensitivity of both types of flow cyto- have already been discussed with regard to the
meter can be improved by labelling the cells monitoring of biomass concentration. An
with a fluorescent dye, either a general label or example of one of these nephelometers is given
one which labels only viable cells (see Sect. by PREIKSCHAT (1987). The instrument uti-
3.2). lizes a scanning laser beam and measures the
Flow cytometry has been used to study cell resulting scattered light pulse, the width of
size for several applications. Batch and contin- which is proportional to the cell size. The
uous cultivations of S. cerevisiae have been in- maximum counting rate is reported to be
vestigated with this technique to study the ef- 300000 particles per second, with a size range
fects of dilution rate on cell size and protein of 0.5 to 250 pm. The published results were
content (RANZIet al., 1986). SCHEPERet al. measured off-line, but the author suggests that
(1987e) studied a different strain of the same a flow-through cuvet could be used and that he
species in oscillating continuous cultures and intends to develop an in situ probe.
in cultivations employing cell recycle.
Size measurements have also been perform-
ed on cultivations of recombinant E. coli 3.2 Viability
(DENNISet al., 1983, 1985; SCHEPERet al.,
1984, 1987d). These studies utilized the anti- It is clearly important to measure the con-
biotic resistance of the genetically modified centration of viable cells in a fermentor, as
cells to estimate the fraction of plasmid-con- these are the microorganisms that actually per-
taining cells. The addition of antibiotic to the form the desired bioconversion. As mentioned
medium blocked cell wall formation in plas- in the discussion of biomass sensors, most
mid-free cells and caused them to elongate, but methods in use today measure only the total
had no effect on plasmid-containing cells. The biomass concentration (and some also include
size differences between the two populations the mass of particulate matter).
202 6 Determination of Cell Concentration and Characterization of Cells
A major problem with the design of viabili- As previously mentioned, the use of flow
ty assays is the vagueness of the concept of a cytometry for quasi-on-line analysis has not
viable cell. The meaning of “viability” is not yet been achieved. The application of this tech-
clear - is a cell “dead” when it can no longer nique for quantifying viable cells requires the
reproduce? Or when it no longer shows meta- use of reagents, and is thus somewhat more
bolic activity? The best functional definition complex than cell-size measurements. An auto-
may vary with different types of cultivations. mated system for adding reagents has recently
Currently, the fraction of viable cells is de- been developed (PENNINGSet al., 1987), al-
termined off-line by three methods: vital stain- though it was not interfaced with a reactor.
ing, cell replication, and metabolic activity
(JONES, 1987). The most common of these is
probably the use of plate counting as a meas- 3.2.2 Sample-Flow Analysis
ure of reproductive ability. It is known, how-
ever, that plate counting underestimates the Although no such systems have been de-
number of viable cells (relative to vital stain scribed, the development of a sample-flow
methods) even in healthy cultures; stressed analysis technique for cell viability should be
populations yield even lower plate-count esti- possible. Since the general scheme of vital
mates. All of these off-line methods are time- staining and metabolic activity measurements
consuming. involves reagent addition and brief incubation,
followed by photometric or fluorometric de-
tection of the dye, a FIA-type system might be
3.2.1 Flow Cytometry used. The best system would involve a dye that
changes its characteristics once inside the cell,
Although it is not possible to assay for re- so that unused dye would not interfere with
productive ability in a straightforward manner the measurement. In addition, the incubation
by flow cytometry, many of the methods for period should be very brief.
vital staining and metabolic activity can be
adapted for use in a flow cytometer. Vital
stains usually measure the ability of a microor- 3.3 Intracellular pH
ganism to exclude a dye from its cytoplasm, al-
though a few require the uptake of a dye as an Many cellular processes, if not the majority,
indication of viability. Methods based on me- are pH-dependent, and the pH value of inter-
tabolic activity usually assay the ability of the est is thus not that of the medium, but rather
cell to enzymatically convert a non-fluorescent the intracellular pH (usually written pHi). Al-
dye to a fluorescent form; this conversion re- though the p H of the cytoplasm is the most
quires a functional energy metabolism. common pHi, the internal pH of various orga-
Many researchers have used dyes such as nelles (e.g., mitochondria or vacuoles) may
trypan blue, propidium iodide, and fluorescein also be relevant. Examples of important cellu-
diacetate to measure viability (by vital stain- lar activities that are pH-dependent include
ing) with flow cytometers (off line) (SHAPIRO, metabolic reactions such as glycolysis, as well
1988). For example, fluorescein diacetate as the transport of small molecules across the
(FDA) is a non-fluorescent dye that is cleaved cell membrane. In addition, the usual strategy
by intracellular esterases to form fluorescein, of controlling the pH of the medium is often
which is retained by viable cells. This stain has an ineffective means of maintaining a constant
been used to select viable erythroleukemia cells pHi, especially for microorganisms that pro-
(HAMORIet al., 1980), and both FDA and 4- duce weak acids. Thus, on-line measurements
methyl-umbelliferone cleaved by phosphatases of pHi would provide valuable information on
were used to detect viable leukocytes (MALIN- the progress of the cultivation.
BERDELand VALET, 1980). Recently, two new A number of off-line techniques are availa-
viability stains were introduced: vita blue dibu- ble for the determination of intracellular pH
tyrate (LEE et al., 1989) and calcofluor white values (NUCCITELLIand DEAMER, 1982).
M2R (BERGLUND et al., 1987). Methods suitable for on-line or quasi-on-line
Cell Components and Characteristics 203
adaptation include NMR spectroscopy and the the naturally abundant isotope, so no special
use of pH-sensitive dyes (e.g., in a flow cyto- substrates are required. In addition, several
meter). other phosphate-containing species of interest
can be studied at the same time. Using 31P,the
pHi is measured by determining the pH-de-
3.3.1 Flow Cytometry pendent chemical shift of intracellular inor-
ganic phosphate and relating it to a titration
There are many dyes available that change curve (GADIAN,1982).
color with changes in pH, and some of these Intracellular p H measurements by in vivo
have been utilized in flow cytometric pHi 31PNMR have become relatively common and
measurements (MUSGROVEet al., 1986; SHA- have been used to study metabolic processes in
PIRO, 1988). More accurate pHi values can be many different organisms. For example, the
obtained from ratiometric measurements, i.e., pH dependence of glycine-proton symport in
the ratio of two wavelengths at which absorp- S. cerevisiae has been reported (BALLARIN-
tion, excitation, or emission intensities change DENTI et al., 1984); pHi values were also
differently with pH. Examples of the off-line measured in a study of the effects of oxygen
use of these dyes include fluorescein for rat on glycolysis in the same organism (DENHOL-
bone marrow cells (VISSER et al., 1979), LANDER et al., 1981). An NMR investigation
2’,7’-bis(carboxyethy1)-5,6-~arboxy fluorescein showed that the pHi response to glucose pulses
(BCECF) for human leukemic T-cells (Mus- of E. coli strains with a high plasmid copy
GROVE et al., 1986), and 2,3-dicyanohydroqui- number was different from that of plasmid
none (DCH) for mouse Ehrlich ascites tumor free cells (AXE and BAILEY,1987). Both cyto-
cells (VALETet al., 1981). DCH and BCECF plasmic and vacuolar p H were measured in
have pH-dependent fluorescence emission pro- Catharanthus roseus and Daucus carota plant
files and are thus easier to use than fluores- cells (BRODELIUS and VOGEL, 1985). Intracel-
cein, which has a pH-dependent fluorescence lular p H has also been measured in immobil-
excitation spectrum (SHAPIRO, 1988). ized cells. Whereas the pHi of agarose- and al-
Measurements can also be made with distri- ginate-entrapped C. roseus cells was identical
butional pH-sensitive dyes, which are weak to that of free cells (VOGEL and BRODELIUS,
acids and bases that are distributed between 1984), GALAZZOet al. (1987) found that S.
the cytoplasm and the medium depending on cerevisiae cells immobilized in calcium alginate
the transmembrane p H gradient. In general, had a lower pHi than otherwise identical sus-
however, the ratiometric measurements dis- pended cells. This was correlated with higher
cussed above are considered to be more accu- rates of glucose uptake and ethanol produc-
rate for use with a flow cytometer because the tion.
ratio measurement is independent of fluoro- Since NMR measurements are not instanta-
phore concentration. neous, it was necessary to provide the cells in
such investigations with nutrients, including
oxygen, making them technically on-line meas-
3.3.2 Nuclear Magnetic Resonance urements. In most of these reports, the experi-
ments lasted only a few hours, but in the study
Spectroscopy involving CHO cells (GONZALEZ-MENDEZ et
al., 1982), the cells were maintained in the
Several different nuclei can be used to meas- NMR reactor for many days. Reactors for
ure intracellular p H by NMR, including ‘H NMR studies will be discussed in Sect. 3.7.
and 13C (ROBERTSand JARDETZKY, 1981),
14N and 15N(KANAMORI and ROBERTS,1983),
”F (OKERLUND and GILLIES,1988), and 31P 3.3.3 Sample-Flow Analysis
(GADIAN,1982). However, nearly all investi-
gations have used 3’P for this measurement. Several of the dyes discussed in Sect. 3.3.1
As mentioned in the general comments on were first used in non-flow cytometric applica-
NMR, this nucleus has the advantage of being tions, including the use of fluorescein diacetate
204 6 Determination of Cell Concentration and Characterization of Cells
Tab. 4. Some Stains for Use in Flow Cytometric Measurements of DNA and RNA (based on SHAPIRO,
1988)
~~
Ethidium bromide DS DNA; DS RNA Measurement of RNA or DNA individually requires use
(EtBr) of DNAse or RNAse (resp.)
Propidium iodide DS DNA; DS RNA As for EtBr; sharper peaks than EtBr; relatively short in-
(PI) cubation times possible
Mithramycin DS DNA Simple protocols;
Chromomycin A3 relatively short incubation times
Olivomycin
MithramycidEtBr DS DNA Less interference from RNA; higher fluorescence intensity
Hoechst 33342, 33258 DS DNA No permeabilization required
4' ,6'-Diamidino- DS DNA Very sharp peaks
2-phenylindole (DAPI)
Acridine orange DS DNA; SS RNA RNA and DNA complexes fluoresce at different 1s; some
(A01 difficulties with method
Hoechst 33342/ DS DNA; DS RNA RNA and DNA complexes fluoresce at different As; no
Pyronin Y permeabilization required
Cell Components and Characteristics 205
and thus treatment with DNAse may be unne- study continuous cultivations of Schizosaccha-
cessary. It is possible to obtain distributions of romyces pombe. In one of these studies, the
both DNA and RNA intracellular concentra- distribution of DNA and RNA of a population
tion simultaneously by using stain combina- undergoing synchronous growth was measured
tions such as Hoechst 33342 and pyronin Y (AGAR and BAILEY, 1982), and in another,
(SHAPIRO,1981). Other fluorochromes such changes in the DNA frequency as a function of
as DAPI are useful when only the DNA levels different dilution rates were measured (AGAR
are to be monitored (STOHRet al., 1977). Al- and BAILEY, 1981).
though the acridine orange method requires
more pretreatment and has been known to be
difficult to use (SHAPIRO,1988), it is the only 3.4.2 Sample-Flow Analysis
stain to provide information on total RNA
content (DS RNA is converted to SS (single- To date, no system for the analysis of nucle-
stranded) RNA before staining). In general, ic acids in a bioreactor sample stream has been
DS and SS RNA levels in cells parallel one an- developed. Although most chemical protocols
other, so DS RNA content measurements can for nucleic acid quantification are too time-
be useful. consuming and complex for use in such a sys-
A number of the dyes in Tab. 4 have been tem, an analytical scheme involving the fluor-
used in off-line studies of cultivations and are ochromes used in flow cytometry might be en-
briefly mentioned here as examples of the visioned. Alternatively, a scheme suggested by
manner in which such information can be uti- KOLIANDERet al. (1984) might be automated
lized. for quasi-on-line RNA analysis. In this meth-
The analysis of bacterial nucleic acid con- od, RNA is hydrolyzed rapidly in 5 mol/L
tent with a flow cytometer is more difficult NaOH, leaving guanylic acid, which is then
than that of mammalian cells, chiefly because measured by HPLC.
of the size difference. Nonetheless, several re-
ports providing intracellular distributions of
the nucleic acid population have been pub-
lished in recent years. Examples include a se- 3.5 Protein
ries of studies on batch cultivations of Bacillus
subtilis by BAILEYand coworkers (BAILEYet The on-line measurement of cellular protein
al., 1978; FAZEL-MADJLESSI and BAILEY, content, like that of DNA and RNA, is of in-
1979, 1980), in which propidium iodide (PI) terest because it provides detailed insights into
was used to measure total nucleic acids, and an growth and metabolism during a bioprocess.
investigation of the effects of several antibio- However, the monitoring of intracellular pro-
tics on the E. coli cell cycle using mithramycin tein levels is especially important in cultiva-
(STEENet al., 1982). Mithramycin DNA stain- tions of genetically modified cells, in which the
ing has also been combined with a mathemati- product is a protein that is retained in the cell
cal model in an investigation of the cell cycle or only partially secreted. In such cases, meas-
of a plasmid-containing E. coli strain (SEOand urement of the level of a specific protein
BAILEY,1987). would be desired. Among other uses, on-line
Yeast strains have also been the subject of or quasi-on-line quantification of the concen-
DNA analyses by flow cytometry. Since yeast tration of the protein product is valuable in de-
cells are larger than bacterial cells, the acquisi- termining the effects of cultivation conditions
tion of the histograms is simpler. SCHEPERet for process development and to determine the
al. (1987e) have reported the distribution of end point of a batch cultivation. It has been
DNA and RNA content for several different observed that the maximum intracellular prod-
batch and continuous cultivations of S. cerevi- uct concentration is often reached before the
siae; PI was used to stain both of the nucleic maximum cell density (Low, 1986).
acids (RNAse was added prior to the DNA Various methods are available for these
measurements). AGAR and BAILEY (198 1, measurements (e.g., LOWRY et al., 1951;
1982) used a similar staining procedure to BRADFORD,1976). The following discussion
206 6 Determination of Cell Concentration and Characterization of Cells
will focus on the monitoring of intracellular methods. Typically, the protein to be studied is
protein content, but will include (quasi) on- an enzyme, the activity of which is measured
line techniques for extracellular protein meas- with the use of substrates that are converted to
urement as well as off-line methods for intra- fluorophores. These substrates are often deri-
cellular proteins, since few on-line systems for vatives of fluorescein, 4-methylumbelliferone,
intracellular protein monitoring have been re- resorufin, and 7-bromo-3-hydroxy-2-naphtho-
ported. Hopefully, other systems will be s-anisidine (naphthol AS-BI) (KRUTH, 1982).
adapted for intracellular proteins and on-line Specific protein assays can be difficult to de-
use. velop because they require that the substrate
A more thorough review of methods for enter the cell and react to a fluorescent prod-
measuring extracellular proteins in cultivations uct that remains within the cell. Often, cell
is provided in Chapter 5 of this volume. membranes that have been permeabilized to al-
low the substrate to enter the cell are also
permeable to the product. In such cases, a
3.5.1 Flow Cytometry trapping reagent can be added to convert the
product into an insoluble compound.
A number of fluorescent, protein-binding Many studies have used flow cytometry
stains have been utilized for the flow cytomet- (off-line) to characterize the total protein dis-
ric measurement of the distributions of total tribution in cell populations, especially for
protein content. These include fluorescein iso- medical applications. Examples of published
thiocyanate (FITC), sulfaflavine, and several studies that are more directly related to culti-
rhodamine compounds such as rhodamine 640 vation processes include a series of papers by
(also known as rhodamine 101), sulforhoda- BAILEYand coworkers on batch fermentations
mine (SRlOI), tetramethylrhodamine isothio- of Bacillus subtilis (FAZEL-MADJLESSIand
cyanate (TRITC), substituted rhodamine iso- BAILEY,1979, 1980; FAZEL-MADJLESSI et al.,
thiocyanate (XRITC), and Texas Red (a deri- 1980), an investigation of the diauxic batch
vative of SR101). Of these, FITC is the most growth of Saccharomyces cerevisiae (GILBERT
commonly used; it binds covalently to proteins et al., 1978), and the measurement of changes
and has good spectral characteristics that al- of total protein content during synchronous
low it to be combined with DNA stains for growth of Schizosaccharomyces pombe (AGAR
dual-parameter analysis. However, the combi- and BAILEY, 1982). Additional studies with
nation of rhodamine 640 and Hoechst 33342 these two yeasts focused on the variations of
(for DNA) would be preferable for quasi-on- protein distributions of a population with
line application of flow cytometry because no changes in the chemostat dilution rate and
permeabilization step is required (CRISSMAN pulsewise addition of substrates. Protein con-
and STEINKAMP,1982). The spectral proper- tent was found to be very sensitive to such
ties of XRITC and Texas Red allow them to be changes (AGARand BAILEY,1981; ALBERGHI-
used with FITC in double protein-labelling NA et al., 1983; RANZI et al., 1986). In an
studies (TITUSet al., 1982). In some cases, the example of the applicability of flow cytometry
total protein content of cells that have been to the monitoring of recombinant bacterial
fixed (e.g., with glutaraldehyde) has been cultivations, SCHEPERet al. (1987d) measured
found to be proportional to measurements of the distribution of total cell protein in a geneti-
90" light scattering (SHAPIRO, 1988), a rela- cally modified E. coli strain before and after
tionship that could be exploited to shorten the thermal induction of plasmid transcription. In
pretreatment procedure in the application of all of these examples, proteins were stained
flow cytometry to process monitoring. with fluorescein isothiocyanate.
The distribution of a specific protein in a Procedures have also been published for the
population of cells can also be ascertained with measurement of the distribution of particular
flow cytometry. This permits one to detect dif- cloned-gene proteins in yeasts and bacteria.
ferences in the content of a specific protein WITTRUPand BAILEY(1988) reported an as-
among various subpopulations in a cultiva- say for /3-galactosidase in S. cerevisiae that uti-
tion, a phenomenon not measurable by other lizes resorufin-/3-D-galactopyranoside as the
CeN Components and Characteristics 207
substrate. In this protocol, the staining proce- based on the biuret assay (SHIDELERet al.,
dure was extremely rapid, the influence of re- 1980), the autoanalyzer (Kusov and KALIN-
leased resorufin was minimized by the addition CHUK,1978) and FIA (SALERNO et al., 1985)
of bovine serum albumin, and a quantitative versions of the Lowry assay, a Bradford-FIA
relationship between measured fluorescence method (RECKTENWALD et al., 1985a), and
and enzyme activity was obtained. An assay modifications of the bicinchoninic acid-based
for penicillin-G acylase has also been reported assay for use in FIA systems (DAVIS and
(SCHEPERet al. 1987d). The substrate in this RADKE,1987). Not all of these systems were
method is 7-phenylacetic-4-(trifluoromethyl)- developed for use with cultivation broth sam-
coumarinylamide, which is converted to a ples, the turbidity and medium composition of
fluorescent coumarin product. This assay was which might prevent accurate measurements.
used to study the penicillin-G acylase distribu- In addition, each system retained the charac-
tion in continuous cultivations of E. coli teristics of the assay upon which it was based,
SK(pHM12) cells at different dilution rates. e.g., protein-to-protein variations, sensitivity,
Both of these assays could be used to pro- and linearity.
vide information on the distribution of the Monitoring of a specific protein obviously
plasmid copy number during cultivations of requires different types of assays than those
genetically modified cells. The measurement of listed above. A variety of automated protein
total cell protein could be used to optimize and measurement schemes have been developed,
control many types of fermentations. including the autoanalyzer system of AHL-
MANN et al. (1986) for the monitoring of intra-
cellular penicillin-G acylase produced during
3.5.2 Sample-Flow Analysis cultivations of a genetically modified E. coli
0'3q
00 12 21 36
lime ( h l
18 I
and IWASE, 1976), an optoelectronic sensor
for serum albumin (GOLDFINCHand LOWE,
1980), and immunosensors for human serum
albumin (KARUBE,1988) and immunoglobulin
Fig. 22. Penicillin-G acylase production by Escheri- G (WEHMEYER et al., 1985). At present, how-
chia coli 5K (pHM12) during a fed-batch fermenta- ever, most of these sensors have several draw-
tion. The line indicates the data from the on-line backs that preclude their use for on-line or
system in Fig. 22, and the points represent off-line quasi-on-line monitoring of cultures, including
measurements (from AHLMANN et al., 1986). the limited functional stability of the biologi-
cal components and problems with mechanical
stability. It is hoped that these difficulties will
strain (Figs. 21 and 22). A thermistor-based be overcome in the near future.
system was also used to measure the amount Liquid chromatography, especially fast pro-
of this enzyme present in the cultivation broth tein liquid chromatography (FPLC), has also
(SCHEPERet al., 1984). The thermistor device been modified for use in the monitoring of
consisted of a nylon tube through which sub- protein levels in cultivations, although no re-
strate and buffer solutions were pumped con- port of an on-line application has been pub-
tinuously. Small amounts of sample were in- lished. One problem is the requirement for a
jected into this stream, and the temperature particulate-free sample. One such FPLC sys-
change was monitored by two thermistors. tem has been developed and used to measure
RECKTENWALD et al. (1985b) modified two concentrations of intracellular P-galactosidase
manual assay procedures for use in FIA sys- (Low, 1986; GUSTAFSSONet al., 1986), and a
tems and were able to monitor formate dehy- system for rapid affinity chromatography for
drogenase and leucine dehydrogenase levels in immunoglobulin G has also been described
sample streams from cell disintegration proc- (CHASE, 1986). Although these methods can
esses. be designed to detect specific proteins, their re-
With the exception of the above-mentioned sponse times are often rather long (up to one
thermistor-based system, all of the automated hour), which limits their usefulness in the
protein analysis systems developed to date uti- monitoring of quasi-on-line cell components.
lize photometric detectors. Many other types
of detectors are available for use with FIA or
autoanalyzer systems, e.g., fluorometers, lu- 3.6 NAD(P)H
minescence detectors, and various biosensors.
Of these, biosensors have the most potential Nicotinamide adenine dinucleotide (NAD +)
for future development; advances in all areas and the related compounds NADH, NADP , +
of biosensor development occur rapidly. and NADPH are the major cofactors involved
A biosensor can be thought of as an analyti- with the transfer of reducing equivalents in
cal device in which a biological component is metabolic reactions. As such, their intracellu-
coupled to a transducer to detect a chemical lar concentrations are closely related to the
species. A very wide spectrum of biological oxidation/reduction state of the cell, a param-
components (e.g., enzymes, antibodies, orga- eter that appears to affect several cellular proc-
nelles, and whole cells) and transducer types esses. Since the intracellular redox state is very
(e.g., ion-selective electrodes, field-effect tran- sensitive to changes in the environment of the
Cell Components and Characteristics 209
- - -
.
A
. 3.0-
A
3.5 2.0-
-12.0
>
O‘2.5-
.-=
0
L
f3.0-
.-o
-
II
.-1.0-
5
.Y
510.5
u
=a
S 1.0- 1.5- f -2.0-
c
6.0
acetobutylicum), or were subjected to various duction rate with NADH fluorescence measur-
environmental perturbations (e.g., changes of ements to form a sensitive control system that
dilution rate in a chemostat). In all of these was applied to yeast fermentations.
cases, the changes in the intracellular redox Fluorescence measurements can be com-
state were clearly measurable as changes in the bined with those of other intracellular parame-
fluorescence signal. ters to obtain more detailed information on
In several studies the rapid, sensitive re- the state of the cells. An experiment of this
sponse of the fluorescence signal to environ- type has been reported by SCHEPERet al.
mental changes has been used to control cer- (1987e) in which off-line flow cytometric
tain aspects of bioreactor operation. In one measurements of DNA, RNA, protein, and
case (Fig. 23), the addition of ethanol to a fed- cell-size distributions were augmented by on-
batch cultivation of Candidu utilis (for single line fluorosensor monitoring of a self-syn-
cell protein) was activated when the fluores- chronized S. cerevisiae culture. A more com-
cence signal decreased below a certain value plete image of the behavior of the cells was
because it had been observed that such a de- provided by the composite data set than could
crease indicated low levels of ethanol (RIS- be obtained from any individual quantity.
TROPH et al., 1977). In a second report, de- There are some problems with the interpre-
creases in the culture fluorescence were used to tation of data measured with fluorescence sen-
control the cellobiose addition to a fed-batch sors, because the observed fluorescence signal
cultivation of a Thermomonospora species can be influenced by a variety of phenomena
(MOREIRAet al., 1981). In both of these exam- such as inner filter effects and fluorescent me-
ples, fluorescence measurements provided a dium components. These have been mentioned
basis for control that would otherwise have previously in the discussion of the use of fluo-
been difficult to obtain, since on-line ethanol rescence probes for biomass estimation. Be-
and cellobiose sensors are not commercially cause of these interfering factors one cannot
available. MEYERand BEYELER(1984) devel- easily correlate a measured fluorescence inten-
oped the concept of reaction rate control and sity with a certain intracellular NAD(P)H con-
combined information on carbon dioxide pro- centration. However, this difficulty need not
Cell Components and Characteristics 21 1
detract at all from the usefulness of fluorosen- redox state (by NADH measurements) of the
sor monitoring. Rather, the user must merely cells simultaneously.
exercise caution when interpreting the fluores- Unfortunately, ATP and energy charge
cence signal. measurements are extremely difficult to per-
form, since ATP levels can change very rap-
idly, especially when the cells are stressed. It is
3.6.2 Biosensors therefore necessary to measure this quantity
on-line or to use a sampling protocol in which
At least two types of biosensors for the cells are immediately treated to stop all
NAD(P)H measurements have been develop- reactions involving ATP. To date, the availa.
ed. One of these is a system based on fiber op- ble techniques for ATP/energy charge measur-
tics, in which the tip of the optical fiber has ements include bioluminescence assays, the use
been coated with bacterial luciferase. NADH of biosensors, and 31PNMR. The first two of
in the test solution reacts at the fiber tip, and these methods require extraction and stabiliza-
the resulting bioluminescence is measured (AR- tion of ATP, as well as ADP and AMP, if the
NOLD and MEYERHOFF, 1988). The other type energy charge is to be measured.
of sensor is based on an enzyme-coated elec- Several studies have utilized various rapid
trode (SCHELLERet al., 1985). The use of extraction procedures and bioluminescence as-
either type could be somewhat problematic, as says (the luciferin-luciferase reaction) to deter-
the cells would have to be disrupted before the mine the level of ATP in different types of
measurement, a process that would most likely cells, although most of these measurements
change the levels of NAD(P)H very rapidly. were made off-line. An exception is the report
of an ATP autoanalyzer system that uses tri-
chloroacetic acid or benzalkonium extraction
3.7 Other Biomolecules and (SIROet al., 1982). Using this analyzer, moni-
toring of the ATP concentration in batch culti-
Metabolic Intermediates vations of S. cerevisiae showed that intracellu-
lar ATP levels peaked in the lag phase before
Naturally, it may be of interest to monitor increasing again during exponential growth.
the intracellular levels of compounds other Several types of ATP-sensitive biosensors
than those previously mentioned. Some of have been reported, including an optoelec-
these measurements, including those of ATP, tronic sensor based on bioluminescence
lipids, and various metabolites, will be dis- (SCHELLER et al., 1985) and an enzyme field
cussed in this section. effect transistor (ENFET) (GOTOH et al.,
1986). These sensors have only been tested in
solutions of ATP. As is the case with biolumi-
3.7.1 Adenosine Triphosphate nescence assays, it is necessary to employ a
rapid extraction technique if these sensors are
The intracellular concentration of adenosine to be used in whole-cell measurements. This
triphosphate (ATP) is an important measure requires the sensors to be unaffected by me-
of the energetic state of microorganisms, since dium components and any extraction chemi-
this compound functions as the cellular energy cals.
carrier. Actually, since adenosine diphosphate It is also possible to use 31Pnuclear magnet-
(ADP) can also donate energy to a reaction, it ic resonance spectroscopy to monitor intracel-
is more appropriate to quantify the bioenerget- lular ATP levels as well as the ratio [ATPI/
ic state in terms of the energy charge: [ADP]. These measurements have been per-
formed on-line in modified NMR tubes con-
[ATP] + 1/2 [ADP] taining perfused, suspended, or immobilized
energy charge = (4) cells. Examples of such studies can be found in
[ATP] + [ADP] + [AMP]
Tab. 6 of Sect. 3.7.3, which also includes a
It would be particularly interesting and valua- brief discussion of these test reactors.
ble to monitor both the energy charge and the
212 6 Determination of Cell Concentration and Characterization of Cells
vantages for longer-term cultivations (GON- obtain information on the levels of phosphory-
ZALEZ-MENDEZ et al., 1982; DRURYet al., lated metabolic intermediates, cellular energe-
1988). tics (ATP and ADP levels), pHi, and inorganic
In the majority of the reports presented in phosphate transport. The various phosphory-
Tab. 6, the authors utilized 31PNMR in their lated sugar compounds of the glycolytic path-
experiments. This nucleus is useful because it way and others appear as one large peak in 31P
is the sole naturally occurring isotope (see Tab. spectra and have usually been reported as a
3) and provides relatively simple, yet informa- combined “sugar phosphate” concentration.
tive spectra. With 31PNMR, it is possible to But a method recently described by SHANKS
214 6 Determination of Cell Concentration and Characterization of Cells
and BAILEY(1988) allows this peak to be de- nique to process research, perhaps the most
convoluted into the peaks of the individual likely area for flow cytometer use as well.
compounds, providing the investigator with The development of biosensors is proceed-
more information. ing at a high rate, but here again most sensors
Other nuclei are also useful in NMR studies, are not designed to measure intracellular com-
including 13C and "N. Since these nuclei are pounds. Those that are could be incorporated
not the naturally abundant isotopes, they are into a sample-flow analytical system (preceded
useful for metabolic pathway studies, in which by a cell-disruption step), as long as they are
the conversion of a small amount of labelled unaffected by medium components and are re-
substrate is followed with time. latively stable with time.
The most basic biological parameter in a
bioreactor is the cell density. Although many
systems can perform this measurement, few
are capable of determining the concentration
4 Future Developments of viable cells, and all have drawbacks.
Further research is clearly needed to develop
all of these systems and to design new sensors
The need for bioreactor monitoring for and analytical devices for monitoring of the
process research, development, optimization, cultivation.
and control is clear, and many analytical de-
vices have been developed to fill this need. Acknowledgements
However, the overwhelming majority of these
instruments quantify extracellular chemical The authors would like to thank K. Dane
and physical quantities. Information on the Wittrup, Jacqueline Vanni Shanks, and Gerd
biological components of the cultivation is of Wehnert for their valuable assistance.
vital importance.
Although the preceding discussion of sen-
sors and analytical devices for on-line and qua-
si-on-line monitoring of biomass and cell char-
acteristics during cultivations shows that very 5 References
few such systems are available at this time, it
should be clear from the descriptions of other
currently used off-line measurement methods AGAR,D. W., BAILEY,J. E. (1981), Continuous
that many of these could be developed for cultivation of fission yeast: classical and flow mi-
bioreactor monitoring. For example, many crofluorometry observations, Biotechnol. Bioeng.
manual assays can readily be modified for use 23, 2217-2229.
in sample-flow analytical systems (FIA or au- AGAR,D. W., BAILEY,J. E. (1982), Measurements
and models of synchronous growth of fission
toanalyzer), and the addition of an efficient, yeast in induced by temperature oscillations, Bio-
automated cell disruption step. The powerful technol. Bioeng. 24, 211-236.
technique of flow cytometry could potentially AHLMANN,N., NIEHOFF, A., RINAS,U., SCHEPER,
be applied to quasi-on-line monitoring by uti- T., SCHOGERL, K. (1986), Continuous on-line
lizing a FIA system for sample preparation. monitoring of intracellular enzyme activity,
The resulting output of the population-distri- Anal. Chim. Acta 190, 221-226.
bution would offer far more information than ALBERGHINA, L., MARIANI,L., MARTEGANI, E.,
the average-value output of other instruments. VANONI,M. (1983), Analysis of protein distribu-
The high cost of this instrument could be low- tion in budding yeast, Biotechnol. Bioeng. 25,
1295-1 3 10.
ered by simplifying its design for a particular ARMIGER, W. B., LEE, J . F., MONTALVO, L. M.,
application. FORRO,J. R. (1985), Fed batch control based
On-line NMR measurements are currently upon the measurement of intracellular NADH,
feasible, although the expense of the instru- Proc. 190th ACS Meeting, Chicago, MBTD 40.
ment and the restrictions on the design of the ARMIGER, W. B., FORRO,J. R., MONTALVO,L.
bioreactor tend to limit the use of this tech- M., LEE, J. F. (1986), The interpretation of on-
References 215
Eds.), pp 551-571, Oxford: Oxford Science Pub- sonance spectrometer, Biotechnol. Bioeng. 32,
lications. 966-974.
COONROD,J. D., KUNZ, L. J., FERRARO,M. J., DUYSENS,L. N. M., AMESZ, J. (1957), Fluores-
(Eds.) (1983), The Direct Detection of Microor- cence spectrophotometry of reduced phosphopy-
ganisms in Clinical Samples, London: Academic ridine nucleotide in intact cells in the near-ultra-
Press. violet and visible region, Biochim. Biophys. Acta
COONEY,C. L. (1981), Growth of microorganisms, 24, 19-26.
in: Biotechnology (REHM, H. J., REED, G., EINSELE,A . , RISTROPH,D. L., HUMPHREY, A. E.
Eds.), Vol. 1, pp 73-112, Weinheim-Deerfield (1978), Mixing times and glucose uptake meas-
Beach/Florida-Basel: Verlag Chemie. ured with a fluorometer, Biotechnol. Bioeng. 20,
CRISSMAN,H. A , , STEINKAMP, J. A. (1982), Rap- 1487- 1492.
id, one-step staining procedures for analysis of EINSELE,A,, RISTROPH,D. L., HUMPHREY, A. E.
cellular DNA and protein by single and dual laser (1979), Substrate uptake mechanisms for yeast
flow cytometry, Cytometry 3 (2), 84-90. cells. A new approach utilizing a fluorometer,
DAVIS,L. C., RADKE,G. A. (1987), Measurement Eur. J. Appl. Microbiol. Biotechnol, 6, 335-339,
of protein using flow injection analysis with Bi- FAZEL-MADJLESSI, J., BAILEY,J. E. (1979), Analy-
cinchoninic Acid, Anal. Biochem. 161, 152-156. sis of fermentation processes using flow micro-
DEBLOIS,R. W., BEAU,C. P. (1970), Counting and fluorometry: single parameter observations of
sizing of submicron particles by the resistive batch bacterial growth, Biotechnol. Bioeng. 21,
pulse technique, Rev. Sci. Znstrum. 41 (7), 909- 1995-2010.
915. FAZEL-MADJLESSI, J., BAILEY,J. E. (1980), Analy-
DEN HOLLANDER, J. A., UGURBIL,K., BROWN,T. sis of fermentation processes using flow micro-
R., SHULMAN,R. G. (1981), Phosphorus-31 nu- fluorometry: Amylase and protease activities in
clear magnetic resonance studies of the effect of Bacillus subtilis batch cultures, Biotechnol.
oxygen upon glycolysis in yeast, Biochemistry 20, Bioeng. 22, 1657-1669.
’ 5871-5880. FAZEL-MADJLESSI, J., BAILEY,J. E., MCQUITTY,
DEN HOLLANDER, J. A., UGURBIL,K., SHULMAN, D. N. (1980), Flow microfluorometry measure-
R. G. (1986), ”P and 13C NMR studies of inter- ments of multicomponent cell composition dur-
mediates of aerobic and anaerobic glycolysis in ing batch bacterial growth, Biotechnol. Bioeng.
Saccharomyces cerevisiae, Biochemistry 25, 212- 22, 457-462.
219. FERNANDEZ, E. J., MANCUSO,A., CLARK,D. S.
DENNIS,K., SRIENC,F., BAILEY, J. E. (1983), (1988), NMR spectroscopy studies of hybridoma
Flow cytometry analysis of plasmid heterogeneity metabolism in a simple membrane reactor, Bio-
in Escherichia coli populations, Biotechnol. technol. Prog. 4 (3), 173-183.
Bioeng. 25, 2485-2490. FERNANDEZ, E. J., CLARK,D. S. (1987), N.M.R.
DENNIS,K., SRIENC,F., BAILEY,J. E. (1985), Am- Spectroscopy: a non-invasive tool for studying
picillin effects on five recombinant Escherichia intracellular processes, Enzyme Microb. Technol.
coli strains: Implications for selection pressure 9, 259-271.
design, Biotechnol. Bioeng. 27, 1490-1494. FINGUERUT, J., GUARDA,E. T. L., CAMARGO,E.
DERMOUN,Z . , BELAICH,J. P. (1979), Microcalori- (1978), Influence of the growth on the spectro-
metric study of Escherichia coli aerobic growth: photometric determination of the yeast concen-
kinetics and experimental enthalpy associated tration in liquid hydrocarbon fermentations, Bio-
with growth on succinic acid, J. Bacteriol. 140 technol. Bioeng. 20, 1285-1286.
(2), 377-380. FLEISCHAKER, R. J., WEAVER,J. C. SINSKEY,A.
DJAVAN,A,, JAMES, A. M. (1980), Determination J. (1981), Instrumentation for process control in
of the maintenance energy of Klebsiella aero- cell culture, Adv. Appl. Microbiol. 27, 137-167.
genes growing in continuous culture, Biotechnol. FORRO,J. R., MAENNER,G. F., ARMIGER,W . B.
Lett. 2, 303-308. (1984), Monitoring cell activity by use of the cul-
DORAN,P. M., BAILEY,J. E. (1987), Effects of ture fluorescence, Proc. 188th ACS National
immobilization on the nature of glycolytic Meeting, Philadelphia, MBTD 79.
oscillations in yeast, Biotechnol. Bioeng. 29, 892- FORSBERG,C. W., LAM, K. (1977), Use of adeno-
897. sine-5’-triphosphate as an indicator of microbiota
DRAKE,J. F., TSUCHIYA, H. M. (1973), Differen- biomass in rumen contents, Appl. Environ. Mi-
tial counting in mixed cultures with Coulter crobiol. 33 (3), 528-537.
counters, Apl. Microbiol. 26 (l), 9-13. FOXALL,D. L., COHEN,J. S. (1983), NMR studies
DRURY,D. D. (1988), Oxygen transfer properties of of perfused cells, J. Mag. Res. 52, 346-349.
a bioreactor for use within a nuclear magnetic re- GADIAN,D. G. (1982), Nuclear Magnetic Reson-
References 211
ance and Its Applications to Living Systems. GUPTA,R. K., GUPTA,P., MOORE,R. D. (1984),
London: Oxford Unversity Press. NMR studies of intracellular metal ions in intact
GALAZZO,J. L., SHANKS,J. V., BAILEY,J. E. cells and tissues, Annu. Rev. Biophys. Bioeng.
(1987), Comparison of suspended and immobil- 13, 221-246.
ized yeast metabolism using 31Pnuclear magnetic GUSTAFSSON,J.-G., FREJ, A.-K., HEDMAN,P.
resonance spectroscopy, Biotechnol. Tech. 1 (l), (1986), Monitoring of protein product formation
1-6. during fermentation with fast protein liquid chro-
GALLAZZO, J. L., BAILEY,J. E. (1989), In vivo nu- matography, Biotechnol. Bioeng. 28, 16-20.
clear magnetic resonance analysis of immobiliza- HAMORI,E., ARNDT-JOVIN, D. J., GRIMWADE, B.
tion effects on glucose metabolism of yeast Sac- G., JOVIN,T. M. (1980), Selection of viable cells
charomyces cerevisiae, Biotechnol. Bioeng. 33, with known DNA content, Cytometry 1 (2), 132-
1283-1289. 135.
GEBAUER, A., SCHEPER,T., SCHUGERL,K. (1987), HANCHER, C. W., THACKER, L. H., PHARES,E. F.
Growth of E. coli in a stirred tank in an airlift (1974), A fiber-optic retroreflective turbidimeter
tower reactor with an outer loop, Bioprocess for continuously monitoring cell concentration
Eng. 2, 13-23. during fermentation, Biotechnol. Bioeng. 16,
GENCER,M. A., MUTHARASAN, R. (1979), Deter- 475-484.
mination of biomass concentration by capacity HANSEN,E. H. (1988), Flow injection analysis - re-
measurement, Biotechnol. Bioeng. 21, 1097- cent developments and future trends, in: Analyti-
1103. cal Uses of Immobilized Biological compounds
GILBERT,M. F., MCQUITTY,D. N., BAILEY,J. E. f o r Detection, Medical and Industrial Uses
(1978), Flow microfluorometry study of diauxic (GUILBAULT, G. G., MASCINI,M., Eds.), pp.
batch growth of Saccharomyces cerevisiae, Appl. 291-308, Norwell, MA: D. Reidel Publ. Co.
Environ. Microbiol. 36 (4), 615-617. HARAN,N., KAHANA,Z. E., LAPIDOT,A. (1983),
GILLIES,R. J., MACKENZIE,N. E., DALE, B. E. In vivo ''N NMR studies of regulation of nitro-
(1989), Analysis of bioreactor performance by gen assimilation and amino acid production by
nuclear magnetic resonance spectroscopy, Bio/ Brevibacterium lactofermenturn, J. Biol. Chem.
Technology 7, 50-54. 258 (21), 12929-12933.
GOLDFINCH, M. J., LOWE, C. R. (1980), A solid- HARRIS,C. M., TODD, R. W., BUNGARD,S. J.,
phase optoelectric sensor for serum albumin, LOVITT,R. W., MORRIS,J. G., KELL, D. B.
Anal. Biochem. 109, 216-221. (1987), Dielectric permittivity of microbial sus-
GONZALEZ-MENDEZ, R., WEMMER,G., HAHN,G., pensions at radio frequencies: a novel method for
WADE-JARDETZKY, N., JARDETZKY, 0. (1982), the real-time estimation of microbial biomass,
Continuous-flow NMR culture system for mam- Enzyme Microb. Technol. 9, 181-186.
malian cells, Biochim. Biophys. Acta 720, 274- HARRIS,C. M., KELL,D. B. (1985), The estimation
280. of microbial biomass, Biosensors 1, 17-84.
GOTOH,M., TAMIYA,E., KARUBE,I., KAGAWA, HARRIS,R. K. (1986), Nuclear Magnetic Reson-
Y. (1986), A microsensor for adenosine-5'-tri- ance: A Physiochemical View, Harlow, England:
phosphate pH sensitive field effect transistors, Longman House.
Anal. Chirn. Acta 187, 287-291. HARRISON, D. E. F., CHANCE,B. (1970), Fluoro-
GREENSPAN, P., MAYER,E. P., FOWLER,S. D. metric technique for monitoring changes in the
(1985), Nile red: a selective fluorescent stain for level of reduced nicotinamide nucleotides in con-
intracellular lipid droplets, J. Cell Biol. 100, 965- tinuous cultures of microorganisms, Appl. Mi-
973. crobiol. 19 (3), 446-450.
GROOM,C. A,, LUONG,J. H. T., MULCHANDANI, HATCH, R. T., WILDER,C., CADMAN,T. W.
A. (1988), On-line culture fluorescence measur- (1979), Analysis of control of mixed cultures,
ing during the batch cultivation of poly-hydroxy- Biotechnol. Bioeng. Symp. 9, 25-37.
butyrate producing Alcaligenes eutrophus, J. HEDEN,C. G., I L L ~ N Y T., (1975) (Eds.) New Ap-
Biotechnol. 8, 271-218. proaches to the Identlfication of Microorgan-
GSCHWEND,K., BEYELER,W., FIECHTER,A. isms, New York: John Wiley & Sons.
(1983), Detection of reactor nonhomogeneities by HEINZLE,E., GOLDSCHMIDT, B., MOES,J., DUNN,
measuring culture fluorescence, Biotechnol. I. J. (1986), Modelling of mixing phenomena ob-
Bioeng. 25, 2789-2793. served in Bacillus subtilis batch cultivations,
GUILBAULT, G. G., IWASE,A. (1976), Assay of Proc. 5th Yugoslavian-Austrian-Italian Chem.
cholinesterase in an electrode system with an im- Eng. Conf., Portoroz, Yugoslavia, pp. 525-534.
mobilized substrate, Anal. Chim. Acta 85, 295- HENDY,N. A., GRAY,P. P. (1979), Use of ATP as
300. an indicator of biomass concentration in Tri-
218 6 Determination of Cell Concentration and Characterization of Cells
choderma viride Fermentation, Biotechnol. (1983), ATP as a measure of active biomass con-
Bioeng. 21, 153-156. centration and inhibition in biological wastewater
HERRERO,A. A., GOMEZ,R. F., ROBERTS,M. F. treatment processes, Proc. Znd. Waste Conf., pp.
(1985), 31PNMR studies of Clostridium thermo- 751-759.
cellum, J. Biol. Chem. 260 (12), 7442-7451. KARL, D. M. (1986), Determination of in situ mi-
HONG, K., TANNER,R. D., MALANEY,G . W., crobial biomass, viability, metabolism, and
WILSON, D. J. (1987), A spectrophotometric growth, in: Bacteria in Nature (POINDEXTER, J.
method for estimating the yeast cell concentra- S., LEADBETTER,E. R., Eds.), Vol. 2, pp. 85-
tion in a semi-solid state fermentation, Process 176, New York: Plenum Press.
Biochem., 149- 153. KARUBE, I. (1988), Immunosensors, in: Analytical
HORAN,P. K., WHEELESSJr., L. L. (1977), Quan- Uses of Immobilized Biological Compounds for
titative single cell analysis and sorting, Science Detection, Medical and Industrial Uses (GUIL-
198, 149-157. BAULT, G. G., MASCINI,M., Eds.), pp. 267-279.
HUANG, T.-L., HAN, Y . W., CALLIHAN,C. D. Norwell, MA: D. Reidel Publ. Co.
(1971), Application of the Lowry method for de- KARCZMAR, G. S., KORTESKY,A. P., BISSELL,M.
termination of cell concentration in fermentation J., KLEIN,M. P., WIENER,M. W. (1983), A de-
of waste cellulosics, J. Ferment. Technol. 49 (6), vice for maintaining viable cells at high densities
574-576. for NMR studies, J. Mag. Res. 53, 123-128.
HYSERT,D. W., KNUDSEN,F. B., MORRISON,N. KELL, D. B. (1987), Forces, fluxes and the control
N. M., VAN GHELUWE,G . , LOM, T.(1979), Ap- of microbial growth and metabolism, J. Gen.Mi-
plication of a bioluminescence ATP assay in rrobiol. 133, 1651-1665.
brewery wastewater treatment studies, Biotech- KILBURN,D. G., FITZPATRICK,P., BLAKE-COLE-
nol. Bioeng. 21, 1301-1314. MAN, B. C., CLARKE,D. J., GRIFFITS,J. B.
ISHIKAWA,Y., SHODA, M,. MARUYAMA,H. (1989), On-line monitoring of cell mass in mam-
(1981), Design and performance of a new micro- malian cell cultures by acoustic densiometry, Bio-
calorimetric system for aerobic cultivation of mi- technol. Bioeng. 33, 1379-1384.
croorgnisms, Biotechnol. Bioeng. 23, 2692- KOCH, A. L. (1970), Turbidity measurements of
2640. bacterial cultures in some available commercial
ISHIKAWA,Y., SHODA, M. (1983), Calorimetric instruments, Anal. Biochem., 252-259.
analysis of Eschericha coli on continuous culture, KOLIANDER, B., HAMPEL,W., ROEHR,M. (1984),
Biotechnol. Bioeng. 25, 1817-1827. Indirect estimation of biomass by rapid ribonu-
ISHIMORI,Y.,KARUBE,I., SUZUKI,S. (1981), De- cleic acid determination, Appl. Microbiol. Bio-
termination of microbial populations with pie- technol. 19, 272-276.
zoelectric membranes, Appl. Environ. Microbiol. KRUTH,H . S. (1982), Flow cytometry: rapid bio-
42 (4), 632-637. chemical analysis of single cells, Anal. Biochem.
JELTSCH,A , , ZIMMERMANN, U. (1979), Particles 125, 225-242.
in a homogeneous electric field: A model for KUBITSCHEK, H. E. (1969), Counting and sizing mi-
the dielectric breakdown of living cells in a cro-organisms with the Coulter counter, in:
Coulter counter, Bioelectrochem. Bioenerg. 6, Methods in Microbiology (NORRIS,J. R., RIB-
349-384. BONS,D. W., Eds.), pp. 593-604, London-New
JENKINSON, D. S., LADD, J. N. (1981), Microbiol York: Academic Press.
biomass in soil: Measurement and turnover, in: KUCHENBECKER, D., BLEY, T., SCHMIDT, A.
Soil Biochemistry (PAUL, E. A., LADD, J. N., (1981), Short-periodic oscillations of culture flu-
Eds.), Vol. 5, pp. 415-471, New York: Marcel orescence in yeast cell populations, Stud. Bio-
Dekker. phys. 86 (2), 92-96.
JONES,R. P. (1987), Measures of yeast death and Kusov, Y. Y., KALINCHUK,N. A. (1978), Auto-
deactivation and their meaning, Process Bio- mated Lowry method for microdetermination of
chem. 22, 118-133. protein in samples and effluents containing non-
JUNTER,G. A., LEMELAND,J. F., SELEGNY,E. ionic detergents, Anal. Biochem. 88, 256-262.
(1980), Electrochemical detection and counting LEE, Y. H. (1981), Pulsed light probe for cell densi-
of E. coli in the presence of a reducible coen- ty measurement, Biotechnol, Bioeng. 23, 1903-
zyme, lipoic acid, Appl. Environ. Microbiol. 39, 1906.
307-316. LEE, C., LIM, H. (1980), New device for contin-
KANAMORI, K., ROBERTS,J. D. (1983), N-15 NMR uously monitoring the optical density of concen-
studies of biological systems, Acc. Chem. Res. trated microbial cultures, Biotechnol. Bioeng. 22,
16, 35-41. 639-642.
KANG, S . J., PUGH,L. B., BORCHARDT,J. A. LEE, L. G., BERRY,G. M., CHEN, C.-H. (1989),
References 2 19
Vita Blue: A new 633-nm excitable fluorescent MUTHARASAN, R. (1987), On-line measurement
dye for cell analysis, Cytometry 10, 151-164. of hybridoma growth by culture fluorescence,
LEIST, C., MEYER,H. P., FIECHTER,A. (1986), Biotechnol, Techn. 1 (4), 213-218.
Process control during the suspension culture of MALIN-BERDEL, J., VALET, C. (1980), Flow cyto-
a human melanoma cell line in a mechanically metric determination of esterase and phosphatase
stirred loop bioreactor, J. Biotechnol. 4, 235- activities and kinetics in hematopoietic cells with
246. fluorogenic substrates, Cytometry 1 (3), 222-
LENZ,R., BOELECKE,C., PECKMANN, U., REUSS, 228.
M. (1985), A new automatic sampling device for MANESHIN,S. K., AREVSHATYAN, A. A. (1972),
determination of filtration characteristics and the Change in the fluorescence intensity of NADH2
coupling of an HPLC to fermentors, Proc. Mod- in Candida guilliermondii in the transition from
elling and Control of Biotechnological Processes, an anaerobic to an aerobic state, Appl. Biochem.
1st ZFAC Symposium, Noordwijkerhout, 11-13 Microbiol. 8, 273-275.
December (JOHNSON,A., Ed.), IFAC Publica- MARISON,I. W., VON STOCKAR, U. (1986), The ap-
tions, New York: Pergamon Press. plication of a novel heat flux calorimeter for
LI, J., HUMPHREY, A. E. (1989), Kinetic and fluor- studying growth of Escherichia coli W in aerobic
ometric behavior of a phenol fermentation, Bio- batch culture, Biotechnol. Bioeng. 28, 1780-
technol. Lett. 11 (3), 177-182. 1793.
LIMA FILHO, J. L., LEDINGHAM, W. M. (1987), MARISON,I., VON STOCKAR,U. (1987), A calori-
Continuous measurement of biomass concentra- metric investigation of the aerobic cultivation of
tion in laboratory-scale fermenters using a LED- Kluyveromyces fragilis on various substrates, En-
electrode system, Biotechnol. Tech. 1 (3), 145- zyme Microb. Technol. 9, 33-43.
150. MATSUNAGA, T., KARUBE, I., SUZUKI,S. (1979),
Low, D. K. R. (1986), The use of f.p.l.cSRsystem in Electrode system for the determination of micro-
method development and process monitoring for bial populations, Appl. Environ. Microbiol. 37
industrial protein chromatography, J. Chem. (l), 117-121.
Tech. Biotechnol. 36, 345-350. MATSUNAGA, T., KARUBE, I., SUZUKI,S. (1980),
LOWRY,0. H., ROSEBROUGH, N. J., FARR,A. L., Electrochemical determination of cell popula-
RANDALL,R. J. (1951), Protein measurement tions, Eur. J. Appl. Microbiol. Biotechnol. 10,
with the Folinphenol reagent, J. Biol. Chem. 125-1 32.
193, 265-275. MERTEN,0.-W., PALFI,G. E., STEINER,J. (1986),
LUONG,J. H. T., VOLESKY,B. (1980), Determina- On-line determination of biochemical/physiolog-
tion of the heat of some aerobic fermentations, ical parameters in the fermentation of animal
Can. J. Chem. Eng. 58, 497-504. cells in a continuous or discontinuous mode,
LUONG,J. H. T., VOLESKY,B. (1982), Indirect de- Adv. Biotechnol. 6, 111-178.
termination of biomass concentration in fermen- MERTEN, 0.-W., PALFI, G. E., STAHELI, J.,
tation processes, Can. J. Chem. Eng. 60, 163- STEINER,J. (1987), Invasive infrared sensor for
167. the determination of cell number in a continous
LUONG,J. H. T., VOLESKY,B. (1983), Heat evolu- fermentation of hybridomas, Dev. Biol. Stand.
tion during the microbial process-estimation, 66, 357-360.
measurement and applications, Adv. Biochem. METZ, H. (1981), Kontinuierliche Triibungsmessung
Eng. Biotechnol. 28, 1-40. in Bioreaktoren, Chem. Tech. 10, 691-696.
LUONG,J. H. T., CARRIER,D. J. (1986), On-line MEYER,C., BEYELER, W. (1984), Control strategies
measurement of culture fluorescence during culti- for continuous bioprocess based on biological ac-
vation of Methylomonas mucosa, Appl. Micro- tivities, Biotechnol. Bioeng. 26, 916-925.
biol. Biotechnol. 24, 65-70. MEYER, H.-P., BEYELER,W., FIECHTER, A.
LUONG,J. H. T., YERUSHALMI, L., VOLESKY,B. (1984), Experiences with the on-line measurement
(1983), Estimating the maintenance energy and of culture fluorescence during cultivation of Ba-
biomass concentration of Saccharomyces cerevi- cillus subtilis, Escherichia coli, Sporotrichum
siae by continuous calorimetry, Enzyme Microb. thermophile, J. Biotechnol. 1, 341-349.
Technol. 5 , 291-297. MOREIRA,A. R., PHILLIPS, J. A., HUMPHREY, A.
MACBRIDE,W. R., MAGAE,J. A., ARMIGER,W. E. (1978), Method for determining the concentra-
B., ZABRISKIE, D. W. (1986), Optical Apparatus tion of adsorbed protein and cell biomass in cel-
and Method for Measuring the Characteristics of lulose fermentations, Biotechnol. Bioeng. 20,
Materials by their Fluorescence, U. S. Patent, 1501-1505.
4577 110. MOREIRA,A . R., PHILLIPS, J. A . HUMPHREY, A.
MACMICHAEL, G., ARMIGER,W. B., LEE, J. F., E. (1981), Utilization of carbohydrates by Ther-
220 6 Determination of Cell Concentration and Characterization of Cells
momonospora sp. grown on glucose, cellobiose, OKERLUND, R. S., GILLIES,R. J. (1987), Determi-
and cellulose, Biotechnol, Bioeng. 23 (6), 1325- nation of Ca + + in whole cells by ”F NMR spec-
1338. troscopy, in: NMR Spectroscopy of Cells and Or-
Mou, D. G., COONEY,C. L. (1976), Application of ganells (GUPTA,R. K., Ed.), Boca Raton, FL:
dynamic calorimetry for monitoring fermenta- CRC Press.
tion processes, Biotechnol. Bioeng. 18, 1371- ORIOL, E., CONTRERAS,R., RAIMBAULT,M.
1392. (1987), Use of microcalorimetry for monitoring
MULLER,W., WEHNERT,G., SCHEPER,T. (1988), the solid state culture of Aspergillus niger, Bio-
Fluorescence monitoring of immobilized micro- technol. Tech. 1 (2), 79-84.
organisms in cultures, Anal. Chim. Acta 213,47- PENNINGS, A., SPETH,P., WESSELS,H., HAANEN,
53. C. (1987), Improved flow cytometry of cellular
MUNOZ,E. F., SILVERMAN, M. P. (1979), Rapid, DNA and RNA by on-line reagent addition, Cy-
single-step most-probable-number method for tometry 8, 335-338.
enumerating fecal coliforms in effluents from se.- PERLEY,C. R., SWARTZ,J. R., COONEY,C. L.
wage treatment plants, Appl. Environ. Micro- (1979), Measurement of cell mass concentration
biol. 37 (3), 527-530. with a continuous-flow viscosimeter, Biotechnol. ,
MUSGROVE,E., RUGG, C., HEDLEY,D. (1986), Bioeng. 21, 519-523.
Flow cytometric measurement of cytoplasmic PREIKSCHAT, E. (1987), New inline method to
pH: A critical evaluation of available fluoro- measure cell count and cell size in fermentors us-
chromes, Cytometry 7, 347-355. ing a focussed laser beam, Proc. 4th Eur. Congr.
NESTAAS,E., WANG,D. I. C. (1981), A new sen- Biotechnol., Vol. 3, pp. 122-125.
sor, the “filtration probe”, for quantitative h a r - RAMSEY,G., TURNER,A. P. F., FRANKLIN, A.,
acterization of the penicillin fermentation. I. My- HIGGINS,I. J. (1985), Rapid bioelectrochemical
celial morphology and culture activity, Biotech- methods for the detection of living microrgan-
nol. Bioeng. 23, 2803-2813. isms, Proc. Modelling and Control of Biotechno-
NESTAAS,E., WANG,D. I. C. (1983), A new sen- logica Processes, 1st IFA C Symposium, Noord-
sor, the “filtration probe”, for quantitative char- wijkerhout, 11-13 December, New York: Perga-
acterization of the penicillin fermentation. 111. mon Press.
An automatically operating probe, Biotechnol. RAMSAY,G., TURNER,A. P. F. (1988), Develop-
Bioeng. 25, 1981-1987. ment of an electrochemical method for the rapid
NESTAAS, E., WANG, D. I. C., SUZUKI,H., determination of microbial concentration and
EVANS,L. B. (1981), A new sensor, the “filtra- evidence for the reaction mechanism, Anal.
tion probe”, for quantitative characterization of Chim. Acta 215, 61-69.
the penicillin fermentation. 11. The monitor of RANZI,B. M., COMPAGNO,C., MARTEGANI,E.
mycelial growth, Biotechnol. Bioeng. 23, 2815- (1986), Analysis of protein and cell volume dis-
2824. tribution in glucose-limited continuous cultures
NICOLS,S. L., JAMES,A. M. (1981), A calorimetric by budding yeast, Biotechnol. Bioeng. 28, 185-
determination of maintenance energy require- 190.
ments during the aerobic growth of Klebsiella RAO, G., MUTHARASAN, R. (1989), NADH levels
aerogenes, Biotechnol, Lett. 3, 119-124. and solventogenesis in Clostridium acetobutyli-
NISHIKAWA, S., SAKAI,S., KARUBE, I., MATSUN- cum: New insights through culture fluorescence,
GA, T., SUZUKI, S. (1982), Dye-coupled electrode Appl. Microbiol. Biotechnol. 30, 59-66.
system for the rapid determination of cell popu- REARDON,K. F., SCHEPER,T., BAILEY,J. E.
lations in polluted water, Appl. Environ. Micro- (1986), In situ fluorescence monitoring of immo-
biol. 43 (4), 814-818. bilized Clostridium acetobutylicum, Biotechnol.
NUCCITELLI, R., DEAMER,D. W. (1982), Intracel- Lett. 8 ( l l ) , 817-822.
lular pH: Its Measurement, Regulation and Utili- REARDON,K. F., SCHEPER,T., BAILEY,J.E.
zation in Cellular Functions. New York: Alan R. (1987), Metabolic pathway rates and culture fluo-
Liss., Inc. rescence in batch fermentations of Clostridium
OHASHI,M., WATABE,T., ISHIKAWA, T., WATA- acetobutylicum, Biotechnol. Progr. 3 (3), 153-
NABE, Y., MIWA, K., SHODE,M., ISHIKAWA, 167.
Y., ANDO, T., SHIBATA,T., KITSUNAI,T., RECKTENWALD, A,, KRONER,K.-H., KULA,M.-R.
KAMIYAMA, N., OIKAWA,Y. (1979), Sensors (1985a), Rapid on-line protein detection in bio-
and instrumentation: steam-sterilizable dissolved technological processes by flow injection analysis
oxygen sensors and cell mass sensor for on-line (FIA), Enzyme Microb. Technol. 7, 146-149.
fermentation system control, Biotechnol. Bioeng. RECKTENWALD, A., KRONER,K.-H., KULA, M.-R.
Symp. 9, 103-116. (1985b), On-line monitoring of enzymes in down-
References 221
stream processing by flow injection analysis rescence studies on aerobic continuous cultures
(FIA), Enzyme Microb. Technol. I , 607-612. of C. cerevisiae, Appl. Microbiol. Biotechnol.
REUSS, M. (1987) Process computer coupled sub- 23, 440-444.
strate feeding for fermentation processes, in: Bio- SCHEPER,T., GEBAUER, A., SAUERBREI, A., NIE-
chemical Engineering (CHMIEL,HAMMES,BAI- HOFF, A., SCHUGERL,K. (1984), Measurement
LEY, Eds.), pp. 149-168, Stuttgart-New York: of biological parameters during fermentation
Gustav Fischer Verlag. processes, Anal. Chim. Acta 163, 111-118.
REUSS,M., BOELCKE,C., LENZ, R., PECKMANN, SCHEPER,T., LORENZ,T., SCHMIDT,W., SCHU-
U. (1987), A new automatic sampling device for GERL, K. (1986), Measurement of culture fluores-
determination of filtration characteristics of bio- cence during the cultivation of P. chrysogenum
suspensions and coupling of analyzers with in- and Z . mobilis, J. Biotechnol, 3, 231-238.
dustrial fermentation processes, BTF-Biotech-Fo- SCHEPER,T., LORENZ,T., SCHMIDT,W., SCHU-
rum 4, 2-12. GERL,K. (1987a), On-line measurement of cul-
RINK,T. J., TSIEN,R. Y., POZZAN,T. (1982), Cy- ture fluorescence for process control of biotech-
toplasmic pH and free Mg2+ in lymphocytes, J. nological processes, Ann. N Y Acad. Sci. 506,
Cell Biol. 95, 189-196. 431-445.
RISTROPH,D. L., WATTEEUW, C. M., ARMIGER, SCHEPER, T., WEHNERT, G., SCHUGERL, K.
W. B., HUMPHREY, A. E. (1977), Experiences in (1987b), in: DECHEMA Biotechnology Confer-
the use of culture fluorescence for monitoring ences, Vol. 1, pp. 63-66, Weinheim-Deerfield
fermentations, J. Ferment. Technol. 55, 599- Beach/Florida-Basel: VCH.
ROBERTS,J. K. M., JARDETZKY, 0. (1981), Moni- SCHEPER, T., GEBAUER,A., SCHOGERL, K.
toring of cellular metabolism by NMR, Biochim. (1987c), Monitoring of NADH-dependent culture
Biophys. Acta 639, 53-76. fluorescence during cultivation of Escherichia
ROY, D., SAMSON,R. (1988), Investigation of coli, Chem. Eng. J. 34, B7-Bl2.
growth and metabolism of Saccharomyces cere- SCHEPER,T., HITZMANN,B., RINAS,U., SCHU-
visiae (baker’s yeast) using microcalorimetry and GERL,K. (1987d), Flow cytometry of Escherichia
bioluminometry, J. Biotechnol. 8, 193-206. coli for process monitoring, J. Biotechnol. 5,
RUZICKA,J., HANSEN,E. H. (1988), Homogeneous 139-148.
and heterogeneous systems, flow injection analy- SCHEPER,T., HOFFMANN,H., SCHUGERL,K.
sis today and tomorrow, Anal. Chim. Acta 214, (1987e), Flow cytometric studies during culture
1-27. of Saccharomyces cerevisiae, Enzyme Microb.
SAKATO,K., TANAKA,H., SAMEJIMA,H. (1981), Technol. 9, 399-405.
Electrochemical measurement of cell popula- SEO, J.-H., BAILEY,J. E. (1987), Cell cycle analysis
tions, Ann. NY Acad. Sci. 369, 321-334. of plasmid-containing Escherichia coli HB 101
SALERNO,R. A., ODELL, C., CYANOVICH, N., populations with flow cytometry, Biotechnol.
BUBNIS,B. P., MORGES,W . , GRAY,A. (1985), Bioeng. 30, 297-305.
Lowry protein determination by automated flow SHANKS,J. V., BAILEY,J. E. (1988), Estimation of
injection analysis for bovine serum albumin and intracellular sugar phosphate concentrations in
hepatitis B surface antigen, Anal. Biochem. 151, Saccharomyces cerevisiae using 31Pnuclear mag-
309-314. netic resonance spectroscopy, Biotechnol.
SAMSON,R., BEAUMIER, D., BEAULIEU, C. (1987), Bioeng. 32, 1138-1152.
Simultaneous evaluation of on-line microcalori- SHANKS,J. V., BAILEY,J. E. (1990), Comparison
metry and fluorometry during batch culture of of wild type and reg1 mutant Saccharomyces cer-
Pseudomonasputida ATCC 11172 and Saccharo- evisiae metabolic levels during glucose and galac-
myces cerevisiae ATCC 18824, J. Biotechnol. 6, tose metabolism using ”P NMR, submitted.
175-190. SHAPIRO,H. M. (1981), Flow cytometric estimation
SANTOS,H., TURNER,D. L. (1986), Characteriza- of DNA and RNA content in intact cells stained
tion of the improved sensitivity obtained using a with Hoechst 33342 and pyronin Y, Cytometry 2
flow method for oxygenating and mixing cell sus- (3), 143-150.
pensions in NMR, J. Mag. Res. 68, 345-349. SHAPIRO,H. M. (1988), Practical Flow Cytometry,
SCHELLER, F. W., SCHUBERT, F., RENNEBERG, R., New York: Alan R. Liss, Inc.
MULLER,H . G . (1985), Biosensors: Trends and SHIDELER,C. E., STEWART,K. K., CRUMP,J.,
commercialization, Biosensors 1, 135-160. WILLS, M. R., SAVORY,J., RENCE, B. W.
SCHEPER,T., SCHUGERL,K. (1986a), Characteriza- (1980), Automated multiple flow injection analy-
tion of bioreactors by in-situ fluorometry, J. Bio- sis in clinical chemistry: determination of total
technol. 3, 221-229. protein with biuret reagent, Clin. Chem. 26 (lo),
SCHEPER,T., SCHUGERL,K. (1986b), Culture fluo- 1454-1458.
222 6 Determination of Cell Concentration and Characterization of Cells
SHIMMONS, B. W., SVRCEK,W. Y., ZAJIC,J. E. troscopic probes generated in situ, Biochemistry
(1976), Cell concentration by viscosity, Biotech- 18 (ll), 2210-2218.
nol. Bioeng. 18, 1793-1805. THOMAS,D. C., CHITTUR,V. K., CAGNEY,J. W.,
SILVENNOINEN, E., KOIVO,H. N. (1982), Estima- LIM, H. C. (1985), On-line estimation of mycelial
tion of the biomass concentration in the Pekilo cell mass concentrations with a computer-inter-
process from the filtrate flow rate of the concen- faced filtration probe, Biotechnol. Bioeng. 27,
trator, in: Modelling and Control of Biotechnical 129-742.
Processes (HALME,A., ed.), pp. 219-224, IFAC- TITUS,J. A., HAUGLAND, R., SHARROW,S. O.,
Publications. SEGAL,D. M. (1982), Texas Red, a hydrophilic,
SIRO, M.-R., ROMAR,H., LOVGREN,T. (1982), red-emitting fluorophore for use with fluorescein
Continuous flow method for extraction and bio- in dual parameter flow microfluorometric and
luminescence assay of ATP in baker’s yeast, Eur. fluorescence microscopic studies, J. Zmmunol.
J. Appl. Microbiol. Biotechnol. 15, 258-264. Methods 50, 193-204.
SOLOMON,B. O., ERICKSON,L. E., YANG, S. S. TURNER,A. P. F., RAMSAY,G., HIGGINS,I. J.
(1983), Estimation of biomass concentration in (1982), Applications of electron transfer between
the presence of solids for the purpose of parame- biological systems and electrodes, Biochem. SOC.
ter estimation, Biotechnol Bioeng. 25, 2469- Trans. 11, 445-448.
2477. VALET, G., RAFFAEL, A., MORODER, L.,
SRINIVAS,S. P., MUTHARASAN, R. (1987a), Cul- WUNSCH,E., RUHENSTROTH-BAUER, G. (1981),
ture fluorescence characteristics and its metabolic Fast Intracellular pH determination in single cells
significance in batch cultures of Clostridium ace- by flow-cytometry, Naturwissenschaften 68, 265-
tobutylicum, Biotechnol. Lett. 9 (2), 139-142. 266.
SRINIVAS, S. P., MUTHARASAN, R. (1987b), Inner VANOUS,R. D. (1978), Understanding nephelomet-
filter effects and their interferences in the inter- ric instrumentation, Am. Lab. 6, 38-46.
pretation of culture fluorescence, Biotechnol. VISSER,J. W. M., JONGELING, A. A. M., TANKE,
Bioeng. 30 (6), 769-714. H. J. (1979), Intracellular pH-determination by
STEEN,H. B., BOYE, E., SKARSTAD, K., BLOOM, fluorescence measurements, J. Histochem. Cyto-
B., GODAL,T., MUSTAFA,S. (1982), Applica- chem. 21, 32-35.
tions of flow cytometry on bacteria: cell cycle ki- VOGEL,H. J., BRODELIUS, P. (1984), An in vivo
netics, drug effects, and quantitation of antibody 31P NMR comparison of freely suspended and
binding, Cytrometry 2 (4), 249-257. immobilized Catharantus roseus plant cells, J.
STEINKAMP, J. A. (1984), Flow cytometry, Rev. Biotechnol. 1, 159-170.
Sci. Znstrum. 55 (9), 1375-1400. WALKER,C. C., DHURJATI,P. (1989), Use of cul-
STOHR, M., EIPEL, H., GOERTTLER,K., VOGT- ture fluorescence as a sensor for on-line discrimi-
SCHADEN,M. (1988), Extended application of nation of host and overproducing recombinant
flow microfluorometry by means of a dual laser Escherichia coli, Biotechnol. Bioeng. 33, 500-505.
excitation, Histochemistry 51, 305-313. WATTEEUW, C., ARMIGER,W. B., RISTROPH,D.,
TAYA,M., AOKI,N., KOBAYASHI, T. (1986), Esti- HUMPHREY, A. E. (1979), Production of single
mation of microbial mass concentration based on cell protein from ethanol by fed-batch process,
fluorometric measurement of cell protein, J. Fer- Biotechnol. Bioeng. 21, 1221-1237.
ment. Technol. 64 (5), 411-417. WANG, N. S., STEPHANOPOULOS, G. N. (1984),
TAYA, M., HEGGLIN, M., PRENOSIL,J. E., Computer applications to fermentation proc-
BOURNE,J. R. (1989), On-line monitoring of cell esses, CRC Crit. Rev. Biotechnol. 2 (l), 1-103.
growth in plant tissue cultures by conductometry, WANG, H., WANG, D. I. C., COONEY,C. L.
Enzyme Microb. Technol. 11, 170-176. (1978), The application of dynamic calorimetry
THAKUR,M. S., PRAPULLA, S. G., KARANTH,N. for monitoring growth of Saccharomyces cerevi-
G. (1989), Estimation of intracellular lipids by siae, Eur. J. Appl. Microbiol. Biotechnol. 5,
the measurement of absorbance of yeast cells 207-214.
stained with Sudan black B, Enzyme Microb. WEHMEYER, K. R., HALSALL,H. B., HEINEMANN,
Technol. 11, 252-254. W. R. (1985), Heterogeneous enzyme immunoas-
THOMAS,J.A., COLE,R. E., LANGWORTHY, T. A. say with electrochemical detection: competitive
(1976), Intracellular pH measurements with a and “sandwich”-type immunoassays, Clin.
spectroscopic probe generated in situ, Fed. Proc., Chem. 31 (9), 1546-1549.
Fed. A m . SOC.,Exp. Biol. 35, 1455. WILKINS,J. R. (1978), Use of platinum electrodes
THOMAS,J. A., BUCHSBAUM, R. N., ZIMNIAK, A., for the electrochemical detection of bacteria,
RACKER,E. (1979), Intracellular pH measure- Appl. Env. Microbiol. 36 (5), 683-687.
ments in Ehrlich ascites tumor cells utilizing spec- WILKINS,J. R., STONER,G. E., BOYKIN,E. H.
References 223
(1974), Microbial detection method based on tion in recombinant Escherichia coli, Bio/Tech-
sensing molecular hydrogen, Appl. Microbiol. nology 6 , 423-426.
27, 949-952. ZABRISKIE,D. W. (1979), Use of culture fluores-
WILKINS,J. R., YOUNG, R. N., BOYKIN,E. H. cence for monitoring of fermentation systems,
(1978), Multichannel electrochemical microbial Biotechnol. Bioeng. Symp. 9, 117-123.
detection unit, Appl. Environ. Microbiol. 35 (l), ZABRISKIE, D. W . , HUMPHREY,A . E. (1978), Esti-
214-21 5. mation of fermentation biomass concentration
WITTRUP,K. D., BAILEY,J. E. (1988), A single-cell by measuring culture fluorescence, Eur. J. Appl.
assay of P-galactosidase activity in Saccharo- Microbiol. 35 (2), 337-343.
myces cerevisiae, Cytometry 9, 394-404. ZABRISKIE, D. W . , ARMIGER,W. B., HUMPHREY,
WITTRUP,K. D., MANN, M. B., FENTON, D. M., A. E. (1975), Estimation of cell biomass and
TSAI, L. B., BAILEY,J. E. (1988), Single-cell growth rate by measurement of culture fluores-
light scatter as a probe of refractile body forma- cence, Proc. ASMMeeting, New York, p. 195.
7 Bioreactor State Estimation
system behavior
Measurements
I
-
I -t
outputs
measurements
Mathematical Formalism
Noise -eliminated
Estimates of directly
227
uncertainty unmeasurable
identification system state
model
incompleteness I- Identification of
uncertain portion of
system dynamics
Fig. 1. Schematic de-
scription of system iden-
tification.
issue of observability. The latter is, of course, Taking the average of consecutive measure-
related to the number and type of available ments is an ad hoc procedure and is a subopti-
measurements and implies that only such var- mal solution to this problem. Such averaging
iables should be included in the state vector discards much valuable information about the
that affect the measurements directly or indi- process, yields unstable estimates, and is equi-
rectly but always in a distinguishable manner, valent to the arbitrary filtering of process
and that are, in this context, observable. The measurements. Identification theory accom-
requirement for observability limits the use of plishes the bioreactor estimation objective in
structured models for identification applica- an optimally balanced manner. Furthermore,
tions. Compared to chemical reactors, fermen- it produces optimal estimates of an array of
tations are supported by considerably more in- additional variables and parameters which are
strumentation for on-line analysis and off- not directly measurable but are observable
line assays. Yet, fermentations are significant- through their indirect effect on the measured
ly less reproducible, controllable, and under- variables.
stood due to the complexities of the cell bio- After the state variable vector, x, has been
catalyst. selected, the next step is to derive the model
In order to illustrate the procedure of model equations for the system dynamics. As men-
formulation, consider a simple bacterial fer- tioned earlier, these models are basically mate-
mentation, in a defined limiting medium, pro- rial balances of the state variables, and as such
ducing extracellular products. Measurements they include other models of growth and meta-
typically include biomass, substrate, and prod- bolic rates. In those cases where reliable model
uct concentrations, off-gas analysis, dissolved structures and parameter values are available,
oxygen, pH, and pH-control-related parame- they are incorporated directly in the balances.
ters. Since no intracellular components are If, however, as is often the case, such models
measured, the use of structured models is not are not available, or are known with a limited
justified. The overall fermentation model con- degree of confidence, provisions need to be
sists rather of material balances based either made for their determination in the formula-
on fully known growth models or models with tion of the model. This is accomplished by ex-
unknown growth and yield parameters that plicitly introducing a parameter vector 8 as un-
need to be identified. The state vector only known in the model equations. This vector 8
contains those variables that can be observed contains all unknown kinetic parameters ap-
from the measurements. Including additional pearing in growth and production rates, yield
measurements, such as the generated heat or expression, stoichiometric coefficients, trans-
rate of nitrogen addition, does not justify in- port coefficients, and other quantities that ent-
creasing the model structure, but it may allow er into the balances. Sometimes, in the absence
the identification of additional parameters. of such model structures, identification of the
Fermentation identification in this case con- quantities (i.e., specific growth rate, etc.)
sists of estimating the concentrations of bio- themselves can be accommodated by including
mass and other metabolites, as well as their them directly in the vector 8.
rates and yields, during the course of fermen- Under these conditions, a suitable general
tation. representation of the system dynamics is
It may be argued that the above task can be
achieved by simple algebraic calculations and +W)
~(t)=f[x(t),B(t),u(t),tI (1)
without any special identification method. The
growth rate, for example, may be obtained Besides vectors x and 8, the model of Eq. (1)
from the difference between two consecutive also contains a control vector u encompassing
biomass measurements, and the specific the substrate feed, oxygen supply, agitation,
growth rate can be determined by dividing the and other control variables of the fermentor.
total growth rate by the average biomass con- Vector w represents random noise sequences to
centration between the two measurements. account for all random variations in the bal-
This approach, however, does not account for ances as written. These variations are intro-
the noise which is present in all measurements. duced by unanticipated factors and also in-
Mathematical Formalism 229
clude uncertainties in the structure of the mod- known, the optimal state is chosen as the one
els and the values of the parameters. that maximizes f. The spread, or variance, of
Of the various measurements made in a fer- the density function is a measure of the relia-
mentation, some are direct measurements of bility of the estimate.
the state variables, but others are only related In on-line system identification, one would
to one or more of the state variables. Off-gas like to update the distribution f as new meas-
analysis, for example, or measurements of urements become available. This can be ac-
heat generation or ammonia addition for p H complished by applying the following recursive
control, d o not directly address any of the rule (Bayes theorem):
components of the state. However, they con-
tain useful information about a number of f x ( r , ) , B ( r , ) lZ(r,), model =
state variables, and the latter influence these
measurements. The relationship between meas- -
- fr (t,) Ix(r,),W,hZ(t,- ,),model .
urements z(t) and state variables is expressed f z ( f , ) l Z ( f , - ,),model
The objective of system identification is to Then, the on-line density function update is
produce the most reliable estimate of the state
and parameter vectors, denoted as 2 and 8, f x ( r , ) lZ(t,),model =
from the available measurements and system
model dynamics. This objective is accom-
plished by maximizing
If one assumes that the noise vectors are in-
f x ( t , ) , B ( f , ) lZ(t,),model dependent random sequences of Gaussian dis-
tribution, then the distribution of state and
which represents the conditional probability measurement vectors is also Gaussian. This
distribution of the system state and parameters density function can be defined by the mean
conditioned upon the inputs, namely the mod- vector and covariance matrix as
el and the measurements, z(t,).The latter is
the collection of all measurements up to time
. .., z(to)]. The optimal esti-
t,, [z(ti),z(t,-l),
mate is the one that is most likely to be true.
Therefore, if the conditional probability is where x, and Qx are the mean and covariance
230 7 Bioreactor State Estimation
of the x-vector distribution, respectively, and n The above equations along with Eq. (9) and
is the dimension of the vector. A similar equa- Eq. (10) form the celebrated Kalman filter, the
tion can be written for the density of the pa- optimal linear on-line data processing algo-
rameter vector. If the estimate at time t,-l, rithm. In identifying the system state vector at
conditioned by the model and the measure- time ti, it uses both the i ( t i ) ,which is the pre-
ments up to t , - ] , is designated as a(t,-,)and dicted state estimate based on the known sys-
its covariance as P(t,- then the density func- tem dynamics, and (HTH)-'HTz(fi), which is
tions on the right-hand side of Eq. (6) can be the estimate of the state according to the meas-
expressed as follows: urements. The above knowledge and informa-
tion are used to an exten_tproportional to their
fx(t,) Iz(r,_,),model=Nx(r3(i(fr);P(tr)) (8) reliability, defined by P(ti)- and HTR - H ,
respectively. This is a rather special case of
In Eq. (8), i ( t , ) is the predicted estimate of system identification in which the system be-
x ( t , ) based on the measurements up to tfp1, havior is known accurately enough. In the fol-
and P(t,) is its covariance matrix. These are lowing sections, the more general situations
obtained by integrating, from t,-l to t,, the are considered in which the system model
following two equations which describe the contains uncertain portions parametrized for
time propagation of the mean and covariance on-line identification, as shown in Fig. 1.
based on Eq. (4): The Kalman filter derived in this section
offers a good starting point for this generaliza-
k(t) =F(t)P(t) =w-
w,- 1) 1) (9) tion.
It has often been the practice to determine
+
P(t)= F ( t ) P ( t )+ P ( t ) F T ( t ) Q state variables directly from the measure-
P ( t , - , ) = P ( t , - , ) (10) ments, without any prior filtering or data pre-
processing. The simultaneous measurement,
where Q is the covariance matrix of w ( t ) . for example, of nitrogen uptake rate and off-
The other density functions appearing in gas analysis in combination with elemental
Eq. (6) are similarly expressed as balances can yield estimates of the biomass
concentration. This approach leaves the ob-
fi(r,)~x(r,),Z(f,_,),model =Nz(r,)(Hx(t,); R, ( l l ) tained estimates entirely exposed to measure-
fi(r,)lZ(f,-l),model =
ment errors and process disturbances. This can
be seen in Eq. (14), which, assuming that
=N,(,,)(NP(t,);NP(t,)HT +R ) (12) measurements alone are relied upon for the es-
fx(r,)IZ(r,),model = N x ( t , )
(f(ti); P(ti)) (13) timation of the state, yields
where R is the covariance matrix of the meas- $ ( t i ) = (HTH)-lH=Z(fi) (18)
urement noise vector u ( t ) . Upon substitution
of Eq. (8), Eq. (ll), and Eq. (13) into Eq. (6), The above equation shows that the state esti-
one obtains the following expressions for the mates are completely vulnerable to any error
state estimate and its variance: that may affect the value of z(ti)that is used in
this estimation. Furthermore, the current state
2(ti)= [&ti) - ' + HTR - 'H-J- value has no bearing on the prediction of the
[P(ti)- i(ti)+ H ~ -R z (ti)] (14) next time instant, and any additional knowl-
edge about the process is simply disregarded.
P(tj)-'=P(ti)-'+H*R-'H (15) As a result, estimates obtained this way will be
noisy, unstable, and inaccurate. The average
Eq. (14) can be equivalently expressed in a of several consecutive measurements can yield
more familiar form using the gain matrix as smooth estimates, but this imposes arbitrary
rules on measurement processing solely for the
a (ti) +
= i( t i ) K (t;)[z(ti) - H i (ti)] (16) sake of smoothing. Besides conflicts that may
arise if this averaging is incompatible with the
K(ti)=P(t;)H=[HP(ti)H=
+ R ]- I (17) process, valuable information contained in the
Extended Kalman Filter 23 1
measurements will be lost in the smoothing This yields the following equations for the
process. propagation of the mean and the variance:
prediction before the update. Use of a high in- are increased to adapt to the changes in fer-
tensity for the process noise, (, has the effect mentation dynamics. This auto-tuning with
of diffusing with time the covariance matrix of adaptive noise estimation was proposed by
the predicted parameter vector in the course JAZWINSKI(1970).
predicted by Eq. (22). This, in turn, tends to As a final note, it should be mentioned that
make the updates rely more on the new meas- the inclusion of the parameters as additional
urements. The estimator is thus alerted to state variables will render a linear problem
changes in the process dynamics and it adapts nonlinear. Furthermore, the computational
accordingly. When it is nearly certain that the load in the EKF implementation of the aug-
model structure is accurately formulated and mented model increases significantly. It is
contains 19 as time-invariant parameters, the therefore advisable to carry out a sensitivity
diffusional intensity of the noise ( may be re- analysis beforehand in order to select those pa-
duced. rameters that d o vary during the process and
Although straightforward, the implementa- that also have a decisive effect on culture iden-
tion of EKF is actually not a simple matter. As tification.
seen in Fig. 1, the filter is designed to compro-
mise the disagreement, if any, between the
model prediction and the incoming measure- 3.2 EKF Applications
ments based on their relative reliability as pre- to Bioreactor Identification
sented by the inverse of their variances. How-
ever, the question of “how unreliable the mod-
el representation is” raises some difficulties Case Study I
with its real implementation. This issue is re-
lated to “filter tuning” and is particularly im- An interesting application of the EKF to a
portant in fermentation processes. Suppose, hypothetical enzymatic process of cellulose hy-
for example, a metabolic variable in the model drolysis is reported by CAMINALet al. (1987).
is not well known and it is parameterized and The main objective of this work was to identi-
allowed to vary with time. It is not easy, in fy the bioreactor state in terms of concentra-
such a case, to know a priori how rapidly this tions and extent of hydrolysis. This was ac-
parameter will change as fermentation pro- complished from measurements of reactant
gresses, or how inaccurate the model descrip- and product concentration, but in the face of
tion will be while this parameter is still being uncertainties in the rate model parameters and
identified. This makes it difficult to predeter- slow enzyme deactivation.
mine the state noise covariance matrix, Q. Fur- The process considered consists of two se-
thermore, there are no straightforward guide- quential enzymatic steps for the overall hydro-
lines to optima1 tuning of the filter. Filter tun- lysis of cellulose to glucose:
ing is specific to individual processes and
usually requires preliminary computer simula-
tion. Sometimes it is possible to add an adap-
cellulose
cellobiose
-
-
cellulase
cellobiose
P-glucosidase
glucose
tive tuning mechanism which adjusts the mag-
nitude of the state noise variance. During the
fermentation, for example, if the model equa- Both steps are carried out in the same well-
tions predict the system dynamics fairly well, mixed reactor from which samples are with-
then the variance of the residuals (i.e., the dif- drawn periodically and analyzed off-line for
ference between the measurements and the es- the concentrations of the three key substances.
timates) gives a n estimate of the variance of Enzymatic assays are employed which yield in
the measurement noise. As the newly calcu- this particular case direct measurements of all
lated residual exceeds the average noise level, state variables. However, due to interfering
the discrepancy is attributed to the change in noise and assay inadequacies, the above meas-
the system dynamics instead of just being neg- urements are corrupted and d o not constitute
lected. In proportion to the discrepancy, the an accurate representation of the real concen-
variances of the state variables and parameters trations. Hence, identification theory was ap-
Extended Kalman Filter 233
plied to extract the best concentration esti- of the Michaelis and inhibition constants are
mates from these measurements. not available, especially during the initial
Following the mathematical formulation of stages of the process. CAMINALet al. (1987)
Eq. (2), the measurements are related to their subjected all the model parameters to identifi-
actual values as follows: cation. Following the approach described in
Sect. 3.1, the parameters are modelled as non-
stationary stochastic processes
e= 5 (34)
where tlT is defined as [r,, rA, K,, KA, (K,/
Ki), (KA/K91T and t' as [L, C5, ( 6 , (7, (8, (91.
where CA, CB, and Cc are the actual levels of Note that by treating the parameters as addi-
cellulose, cellobiose, and glucose, respectively. tional state variables the algorithm of Eqs. (19)
The measured concentrations denoted by the to (25) can be used to obtain the best estimates
subscript m differ from the corresponding true of the extended state vector, except for the as-
values by the random noise sequences, vl, u2, signment of the tuning factors.
and v 3 , respectively. Intuitively, rA and r, appear to show more
Assuming Michaelis-Menten kinetics with rapid variation than other kinetic constants,
competitive inhibition for the rates of enzy- since these two parameters reflect the enzyme
matic reactions, the dynamics of the state is thermal deactivation which is certainly antici-
described by the following material balances: pated. Thus, the noise processes for these two
parameters C4 and C5 require higher diffusional
intensity of their distribution than in the case
of other parameters. In this work, the tuning
matrix is determined by simulation as
.;]
3
E5
0 0
0
In another application of EKF to fermentor
identification, SAN and STEPHANOPOULOS
(1984a) attempted the on-line identification of
batch, fed-batch, and continuous fermentation
of baker's yeast on glucose-limiting nutrients.
This is the first time that rigorous identifica-
tion procedures were applied to bioreactors,
and three novel contributions of this applica-
-%,---,
Time (mini
tion should be noted. First, a n expanded fer-
mentor state was used, including the usual fer-
mentor variables as well as the culture parame-
ters mentioned above. Second, no models were
employed for the dependence of the latter on
the concentration of biomass and components
k 2 of the abiotic phase. Instead, culture paramet-
;I ers were identified by employing an adaptive
exp. No. 2
U
estimation algorithm. Third, several of the
OO 200 400 600 measurements employed were only indirectly
Time (min)
related to the fermentor state. Specifically,
Fig. 2. Case Study 1 of EKF: Raw measurements of ammonia addition for p H control was meas-
cellobiose (0) and glucose (0) concentrations and ured, and oxygen uptake and carbon dioxide
their estimates by EKF (curves). Experiment 1 (top): evolution rates were monitored by off-gas
initial concentrations of cellulose and cellobiose are analysis. The measurements were combined
9.95 g/L and 0.0 g/L, respectively. Experiment 2 with stoichiometric balances which allowed the
(bottom): 0.0 g/L and 5.54 g/L, respectively. Fig- calculation of some key variables and the cast-
ures from CAMINAL et al. (1987). ing of the problem into the form of Eqs. (1)
and (2). In the same context, proton balances
have been employed in other applications in
order to relate ammonia addition and p H con-
trol to the production of acidic and basic prod-
ucts (SAN and STEPHANOPOULOS, 1984b).
In this example, the state vector comprised
biomass, substrate, and product concentra-
tions as well as the specific growth rate, p , and
the overall substrate and product yields. The
reactor was equipped for the on-line measure-
0 0 1
0 100 200 300 400 500 600 ment of pH, temperature, and the amount of
Time I min I ammonia added for pH control. Off-gas anal-
ysis also produced on-line measurements of
Fig. 3. Case Study 1 of EKF: Estimation of parame-
ter rA, the enzymatic activity of P-glucosidase, by the oxygen uptake and carbon dioxide evolu-
EKF. The EKF algorithm identifies the decrease in tion rate, OUR and CER, respectively.
the enzymatic activity during the reaction. (A) ex- In order to relate the above measurements
periment 1 and (A) experiment 2. Figures from to the state variables, stoichiometric relation-
CAMINAL et al. (1987). ships were invoked based on the following re-
presentation of the fermentation process
(WANGet al., 1977):
are believed to convey additional information
about the culture. Such parameters, termed ~C6H1206+ b02 + c N H ~ +
jecting the parameters themselves to adaptive mates by simple moving-average and intermit-
identification. The choice depends basically on tent off-line measurements are also shown for
the availability of a reliable model structure. If comparison. The estimates for biomass, glu-
one is available, its inclusion leads to a more cose level, and ethanol level show smooth
robust filter performance at the expense of in- trends and appear to agree well with the off-
creased complication with the implementation line measurements. As predicted from the lack
algorithm. If there are doubts about the validi- of explicit dynamic structures for the substrate
ty of the model, its use is inconsequent and the and ethanol yields, the estimates of these pa-
estimation algorithm relies entirely on the rameters are mostly dependent on the incom-
measurements. ing measurements. Thus, the estimates evi-
In this case study, it was assumed that n o dently reflect the noise of the measurements.
model was available and the culture parame-
ters were treated as independently identifiable
quantities and allowed to vary with time ac-
cording to the dynamics
- 090
k = 91 (0 (43)
y, = 93 (0 (45)
The above correspond to Eq. (27) of the pre- I
vious section for the augmented state vector. 100 200 300 400 500 600
With the state balances (40),(41), and (42) and Time [ h I
the measurement equations (37) to (39), they
Fig. 4. Case Study 2 of EKF: Biomass concentra-
can be used in the filter implementation ac-
tions in a fed-batch fermentation of baker’s yeast:
cording to Eqs. (19) to (25) of the previous sec- (0) off-line measurements; (. . .) estimates by EKF;
tion. (-) estimates by moving average. Note that EKF es-
The operation of the implemented filter can timates agree well with off-line measurements as fer-
be summarized as follows: The available raw mentation progresses. Figures from SAN and STE-
measurements are mapped to a more relevant PHANOPOULOS (1984b).
form, the culture parameters, by using the fer-
mentation equation (36). Filtering by Eqs. (43)
to (45) follows to eliminate noise. Then they
8 00
are fed to the deterministic material balances,
Eqs. (40) to (42), to compute the changes of
the biomass, the substrate level, and the prod- I
PI
Y)
100 200 300 400 500 600 1.00 2:OO 3.00 4.00 500 6:OO
Time( h 1 Timelh)
Fig. 6. Case Study 2 of EKF: Ethanol concentra- Fig. 8. Case Study 2 of EKF: Product yield as a
tions in a fed-batch fermentation of baker's yeast: function of time in a fed-batch fermentation of bak-
(0) off-line measurements; ( * * .) estimates by EKF; er's yeast: (. . .) estimates by EKF; (-) estimates by
(-) estimates by moving average. Figures from SAN moving average. Note that EKF gives smooth esti-
and STEPHANOPOULOS (1984b). mates. Figures from SAN and STEPHANOPOULOS
(1984b).
0.4C
3
.-W
A
W
0.30
c
e
e
VI
n
a
wl
0.20
I
100 200 300 4M3 500 600 100 200 300 400 500 600
Time I h 1 Time ( h I
Fig. 7. Case Study 2 of EKF: Specific growth rate as Fig. 9. Case Study 2 of EKF: Substrate yield as a
a function of time in a fed-batch fermentation of function of time in a fed-batch fermentation of bak-
baker's yeast: (. * * ) estimates by EKF; (-) esti- er's yeast: (. . .) estimates by EKF; (-) estimates by
mates by moving average. Note that EKF gives moving average. Figures from SANand STEPHANO-
noise-free estimates. Figures from SAN and STE- POULOS (1984b).
PHANOPOULOS (1984b).
However, the estimate of the specific growth straightforward implementation of EKF, the
rate shows substantially noise-free behavior, recursive identification of Eq. (3) may not give
whereas simple averaging manifests unsteady the same solution as the off-line batch identifi-
fluctuation. Large non-recurring measurement cation process due to approximations intro-
pertubations are rejected by the filter. duced by the nonlinearity. This sometimes
As mentioned earlier, the state vector aug- makes it difficult to establish satisfactory con-
mentation results in a nonlinear estimation vergence. This point is elaborated further in
problem. Therefore, despite the flexibility and Sect. 5 .
238 7 Bioreactor State Estimation
el) and its future projection using the model . ~ , ( t ~ ; e[)Z-( ~t J ) - ~ x ' ( t J ; e ) i
equation will result in small prediction errors.
Conversely, this property can be used for iden- At time t,, the optimal state/parameter esti-
tification of the parameters in the model equa- mates are obtained by solving the following
tion if they are a priori unknown. The best pa- two simultaneous equations.
rameter values are the ones which result in the
smallest prediction errors. This is the underly-
ing concept of the system identification meth- (48)
od dealt with in the following section. Algo-
rithms for the implementation of the concept
of recursive parameter estimation are present- (49)
ed in Sect. 4.2.
minimizes the sum of the prediction errors. date, the parameter values which maximize the
Second, the state estimation, which is the solu- likelihood function (or minimize the sum of
tion of Eq. (49), is obtained by filtering the the least square prediction error) can be lo-
measurements with the aid of the updated cated.
model by the new parameter values. The sec- At this point it is important to pause and re-
ond step is identical to the Kalman filter dis- flect upon the enormity of the problem sug-
cussed in Sect. 2.2. Therefore, the focus of this gested by Eqs. (50) and (51). At some sampling
approach to system identification is on the de- time ti, one needs to first formulate the sum of
velopment of a tractable recursive least square quadratic terms of Eq. (50). This requires the
prediction error algorithm for parameter iden- state predictions R(t,), O s k c j , i.e., at all pre-
tification. vious sampling points. These predictions are
obtained by the model equations and depend
on the parameter values used in the current
4.2 Recursive Least Square Error iteration. After forming the derivative sug-
gested by Eq. (50) and the Gradient and Hes-
Algorithm sian matrices of the likelihood function, one
needs to iterate Eq. (51) until satisfactory con-
The development of the previous section is vergence is achieved. Furthermore, as the pa-
conceptually straightforward and not substan- rameters are updated at every iteration, the
tially different from typical static (batch) pa- computation of the Gradient and Hessian ma-
rameter estimation by least squares. However, trices becomes heavily loaded. This is appre-
it leads to a sizable implementation problem so ciated by looking at the change of the state
that a simplifying recursive version of Eq. (48) prediction f ( t , ) resulting from changes in the
is sought in this section. parameter vector. This prediction depends on
First it is assumed that the covariance ma- the parameters when it is projected from the
trix Q, does not depend on the parameter vec- state estimate of the previous point 2 ( t k - l ) .
tor. This allows one to discard the normalizing The state estimate, in turn, depends on the pa-
factors from the density functions in Eq. (48) rameter vector which is obtained as seen in
which can now be written equivalently as Eqs. (9), (lo), (16), and (17). The entire proc-
ess is repeated at the next sampling point to
J= ---z
U
ae 2
l
[z(tj)-HB(tj;8)lT/1-1(fj).
yield the parameter values that minimize the
sum of the prediction error up to that time. If
one considers the very large number of sam-
[z(t,)- m ( t j ; ell = oT (50) pling points, the strong dependence of Gra-
dient and Hessian matrices on the parameters,
z(tj) is the measurement at time tj and R(tj) is and the recursive nature of the state prediction
the prediction of the state at time tj from the equations, it becomes clear that the rigorous
state at time tjPl using the model equation. A approach for parameter estimation is time-
is the covariance matrix of the prediction error consuming and very intensive, making it inap-
(denoted as Q, in Eq. (47)), whose dependency propriate for on-line implementation.
on the parameter vector is neglected. The pa- One approximation that greatly simplifies
rameter vector that satisfies Eq. (48) and the search for the optimal set of parameters is
solves this nonlinear least square error prob- to replace the iterative index in Eq. (51) with
lem can be found by the Gauss-Newton itera- the recursive time index. This means that in-
tive searching method, as stead of iterating batch with a static set of
measurements, one now does a dynamic itera-
8,= 8,- 1 + S[d,- 11 - I G[8,- 11 (51) tion, moving from measurement to measure-
ment in time and updating the parameters by
where N is the iteration index and G and S an equation analogous to Eq. (51). The Gauss-
are the first (Gradient) and second (Hessian) Newton search for the parameter vector is exe-
derivatives of the likelihood function with cuted only once at each sampling time starting
respect to 8. By iterating on the parameter up- from the previous sampling time. This ap-
240 7 Bioreactor State Estimation
proach is successful with slowly varying pa- parameter values and is defined as
rameters. Compared to the rigorous approach
of Eqs. (50) and (51), it will converge more d
y'(t.) = - [Hx'(tj)]
slowly on the correct parameter values. How- -dB
ever, it may be rendered more adaptive by
some mechanisms which will be discussed later In order to determine the y/-matrix, a state
in this section. In general, it works well with prediction equation is first formed by project-
systems that exhibit smooth variations in time ing the state estimate at the previous time in-
such as fermentors. stant with the transition matrix @ ( t i ,t i - 1) de-
The new parameter updating equation that rived from the model equation
is obtained by replacing the iterative index
with the recursive time index is 2(tj) = @ ( t i , ti- ,)?(ti_ 1) (57)
S It,, e(tl- 91 =
=~[t,-l,e(t,-*)I+y/(fJA -l(tJ!UT(t,) (54)
Matrix A (t,),which is the covariance matrix of
the prediction error, is approximately updated (59)
by the recursive formula
The subscript, I J ( ~ , - , ) , indicates that the corre-
( i > j i ( t , ) = ( i - l ) j i ( t , - l ) + [ ~-
(tl) sponding quantities are evaluated at the pa-
- H W l [Z(t,) -HW,)lT (55) rameter vector estimate available at time t,-l.
Eq. (59) is a rigorous equation for the time
Note that Eq. (52) indicates that when a discrep- evolution of the y/-matrix. However, the com-
ancy exists between the prediction and the ac- putation of y/(tl- I o(r,- ,) causes considerable
tual measurement (the prediction error), the complications. It is generally accepted to use
parameter vector readjusts itself in an optimal- y / ( t , - . J I i ~ r , - ~ ) instead of v(t,-~)Id(~,-,)
in Eq.
ly balanced manner. The readjustment is pro- (59), leading to an approximate recursive up-
portional to the predictionn error ( z - H f l , and date of the y/-matrix which, however, can be
the gain matrix ( S - I y A - ' ) , which is calcu- implemented on-line.
lated on a statistical basis. The algorithm of Eqs. (54) through (59) can
Probably the most important quantity in be summarized as follows. First, the y-matrix
this recursive parameter estimation algorithm is determined before the measurement at time
is the sensitivity matrix, also known as y/-ma- t, is made. Using Eq. (57) for the state predic-
trix. It appears in Eqs. (53) and (54) in deter- tion and the current measurement z(t,), the
mining the Gradient and Hessian matrices of prediction error ( ~ ( t-HR(t,)) ,) and its covar-
the likelihood function. It expresses the sensi- iance matrix A (t,) are calculated. Subsequent-
tivity of the state prediction to changes in the ly, the Gradient and Hessian matrices are cal-
Parameter Estimator on State Observer: Sequential State/Parameter Estimation (SSPE) 24 1
f(t-11 E(tl
As mentioned earlier, the above algorithm is covariance of old data to become larger. This
applicable to systems with adequate models discounts the reliability of old data and de-
for which the parameters are not expected to pends more on recently collected findings for
vary significantly with time. If system dynam- the determination of the parameter estimates.
ics exhibit significant variation with time so Different tracking schemes can be accommo-
that the model parameter vector needs to be dated by selecting different weighting se-
updated accordingly, the above algorithm may quences, such as Fixed Windows and Forget-
not show satisfactory tracking capability. This ting Factor (GOODWIN and SIN, 1984). The re-
is caused by the continuous increase with time sult is often a more flexible and robust param-
of the Hessian matrix, Eq. (54)jresulting in a eter estimation algorithm.
diminishing contribution of incoming meas-
urements to the update of the parameters. Put
differently, excessive confidence is gradually 4.3 SSPE Application to Bioreactor
being built on current parameter estimates, Identification
which makes the algorithm place less weight
on more recent observations.
In order to capture important information Case Study I
contained in current measurements and also to
improve the ability of the algorithm to handle Use of the SSPE algorithm for regulating
time-varying parameters, the algorithm is glucose concentration in yeast fermentation
modified to prevent the overconfidence has been reported (PARKand RAMIREZ, 1989).
founded on current parameter estimates. This In a fed-batch reactor, changes in the glucose
is accomplished by modifying the prediction level can be represented by the following mate-
error covariance matrix through the use of a rial balances
242 7 Bioreactor State Estimation
vector are inserted as their values are available accuracy of the model equations to describe
at this time. A numerical derivative method is the actual fermentation dynamics. As such,
used to obtain the matrix value of the term. the performance of the feedforward control is
Otherwise, a lengthy derivation is required to directly coupled to that of the system identifi-
obtain an explicit expression for the last term cation algorithm. In addition, a PI feedback
in Eq. (72) due to the complicated dependence control is supplemented as a backup in the
of the Kalman gain matrix on the parameter event the feedforward control does not per-
vector. The second step is the update of the form well. This approach of control law design
parameter values. With the ty-matrix obtained offers the opportunity of a precise control per-
using Eq. (72), the current estimates of A, S, formance with reliable on-line system identifi-
and G are produced by employing Eqs. (56), cation. Also, if the system identification fails,
(54), and ( 5 9 , respectively. Finally, the pa- disastrous consequences are prevented by
rameter vector update was made by using Eq. shifting to the PI-backup control. It should be
(53). The third step is the state vector estima- noted that the performance of the PI control is
tion by Kalman filter Eq. (58) using the renew- not as satisfactory as that of the optimally per-
ed model equations that jncorporFte the up- forming feedforward control.
dated parameter values 8, and &. Starting Typical results of the performance of the
from an initial guess of the parameter values, combined system identification (SSPE) and the
the above procedure is repeated at each sam- described control law are shown in Figs. 11 to
pling moment. 14. The glucose concentration is regulated at a
In Sect. 4 it was noted that the underlying setpoint of 0.2 g/L in a 20 L fermentor. In this
basis for the SSPE algorithm is the distinctive-
ly lower rate of parameter variation compared
to that of the state variables. Since ,D and Y
can be strongly dependent upon culture condi-
tions such as the glucose level, dissolved oxy-
gen level, temperature, and pH, the parametri-
zation of these two metabolic variables may
not be appropriate for SSPE application.
However, if the culture conditions are regul-
ated at constant levels, it is reasonable to ex- 0.0 I ' . O
pect that these parameters do not show consid-
erable variation. PARKand RAMIREZ(1989)
used the SSPE algorithm coupled with regula- -0.1
tion of culture conditions, including a model- - I N
based optimal glucose-level regulator. -0.2 2
Successful regulation of glucose concentra-
tion requires accurate estimates of the state -0.3
variables and, furthermore, reliable prediction
of bioreactor future behavior. The importance
of a reliable model becomes even more pro- 0 1 2 3 4 5 6 7
Time I h 1
nounced at low glucose concentrations due to
rapid culture dynamics that may lead to glu- Fig. 11. Case Study 1 of SSPE: Glucose and bio-
cose depletion. The objective of robust glucose mass concentration changes in a glucose controlled
regulation at low levels was met by designing a fed-batch yeast fermentation. The glucose concen-
control law which appropriately utilized the tration is regulated at set-point 0.2 g/L by a control
law which incorporates the SSPE algorithmqigures
outputs of the system identification algorithm. from PARKand RAMIREZ (1989) (above),
In short, the control scheme consists of a feed- Fig. 12. Case Study 1 of SSPE: Adaptation of pa-
forward and a Proportional-Integral (PI) feed- rameter estimates of B,(parl) and 02(par2) by the
back control. The feedforward control is de- SSPE algorithm. Note the rapid adaptation fol-
termined by using model predictions of the lowed by stabilization of the estimates. Figures from
state variables. Thus, it depends heavily on the PARKand RAMIREZ (1989) (below).
244 7 Bioreactor State Estimation
0000 1 2 3 54 5 6 7 L8
Time I h I
Fig. 13. Case Study 1 of SSPE: Feedforward (qff)
and feedback (qp,) control actions during a glucose-
regulated fed-batch fermentation. Feedforward con-
trol is mainly determined by the SSPE system iden-
tification algorithm. Figures from PARKand RAMI-
rez (1989) (above).
Fig. 14. Case Study 1 of SSPE: Profile of the ratio
(qratio= I qPI/qffI), the regulation is done primarily
by feedback (qp,) and feedforward (qff) control ac-
tion. Note that the contribution of qpI decays as
time progress, indicating that the SSPE algorithm
performs the identification task satisfactorily. Fig-
ures from PARKand RAMIREZ(1989) (below).
1.o
experiment, the Kalman filter gain matrix for 0 1 2 3 4 5 6 7 8
Time I h I
state estimation was set relatively high so that
the system identification algorithm is more Fig. 15. Case Study 1 of SSPE: Performance of glu-
sensitive to incoming measurements. An active cose level control when SSPE activity is repressed.
update of the system dynamic model is indi- Comparison with Fig. 11 demonstrates the impor-
cated by the extensive change in the estimated tance of active system identification for satisfactory
fermentor control. Figures from PARKand RAMI-
parameter values. For the initial parameter rez (1989) (above).
values, those which were calculated indepen- Fig. 16. Case Study 1 of SSPE: Profile of parameter
dently at nearly the same culture conditions estimates of &(parl) and 02(par2) when SSPE activ-
were used. However, significant variations ity is repressed. Note the distinctively slower adapta-
during fermentor operation were observed. tion of estimates compared with those in Fig. 12.
Based on the updated model, the feedforward Figures from PARKand RAMIREZ(1989) (below).
control action regulates the culture glucose lev-
el precisely at the designated set point. As
shown by the ratio of the two control actions, Case Study 2
I qPl/qffI , the regulation is done primarily by
the feedforward control and the PI-backup An algorithm for the detection of bioreactor
control is not activated for most of the time. contamination was reported by CHATTAWAY
Parameter Estimator on State Observer: Sequential State/Parameter Estimation (SSPE) 245
and STEPHANOPOULOS (1989). The behavior identify the corresponding specific growth
of contaminated cultures is distinguishably dif- rates.
ferent from that of uncontaminated ones. By The non-observability problem can be
operating on-line a system identification algo- solved by recalling that the algorithm has a
rithm, this type of erroneous fermentor per- two-fold objective, namely, accurate tracking
formance and its dynamics can be identified. for as long as a pure fermentation is main-
Combined with an appropriate decision criteri- tained and providing an alarm signal at the
on, this algorithm can then be used for the earliest possible moment after the onset of
timely detection of contaminating species in a contamination. One can then assign a known
fermentor. and constant value, b, to the specific growth
Consider the following two growth equa- rate of the contaminant and subject x, and xc
tions for the desired species and the contami- to identification, a definitely legitimate task.
nant in a batch culture: By choosing a large value for b, in fact larger
than the maximum possible specific growth
(73) rate of the desired strain, the growth of the lat-
ter is distinguished mathematically from that
xc. I * 1 = (1 + TPC)XC,f + cc, f (74) of the contaminant and, to a certain extent,
physically also. If the reactor is not contami-
Subscripts m and c designate the desired spe- nated, system identification reveals that xc,
cies and the contaminant, respectively. The stay close to zero, for the very rapid growth
specific growth rate of each species is denoted kinetics of the contaminant is incompatible
as p and the sampling interval as T. Since the with the observed fermentor dynamics as ex-
specific growth rates are parametrized, the pressed by the measurements of the biomass
noise sequences c are added to compensate for and growth rate. Under such conditions (i.e.,
model inaccuracy. Eqs. (73) and (74) form the xc,t= 0), the system is observable and accurate
state model equations. estimates of ,urnand x, are obtained. Upon the
Using the stoichiometric balances of Sect. appearance of contaminants growing at a high
3.2 (Case Study 2), the data from off-gas anal- rate, a non-zero biomass must be invoked to
ysis can be transformed to growth rate meas- explain the faster changes in biomass and total
urements. Furthermore, through nitrogen bal- growth rate. This produces an alarm signal
ance and ammonia addition measurements, when the state and parameter estimates
on-line biomass concentrations were obtained. beyond this point deviate from the true values,
Accounting for the usual noise, these measure- unless the selected value of b happens to be
ments are cast as follows in the measurement close to the true specific growth rate of the
equations for the identification formalism: contaminant. For more sensitive detection, a
small value of b should be used.
(75) The SSPE algorithm was used for the iden-
tification of the specific growth rate p, and
the estimation of the biomass concentration.
To this end, a recursive equation for the up-
Eqs. (75) and (76) relate the measurements to date of the ty-matrix is required. Combining
the state variables x, and xc. It should be Eqs. (57) and (58) yields the following equa-
noted that in the absence of a distinguishing tion for the propagation of the state predic-
characteristic for either of the species, neither tion:
the biomass nor the growth rate measurement
can be partitioned between the desired strain Z(t,+ 1) + @K(ti)Z(ti) (77)
= @ [I-K(ti)Hl2(ti)
and the contaminating species. In this context,
the system of Eqs. (73) through (76) is non-ob- The equation for the propagation of the meas-
servable, which, means that the two measure- urement prediction is obtained by substituting
ments of growth rate and biomass do not R(ti+,) by H -' Z ( ti+ J and ?(ti) by H-'Z(ti).
contain sufficient information to estimate Upon substitution and rearrangement, one ob-
the individual biomass concentrations and to tains the following recursive prediction equa-
246 7 Bioreactor State Estimation
It can be seen that the dependence of From the above, the equation for the recursive
H @ H - ' on the parameter ,urnis linear. How- update of the ipmatrix is given as
rn off-line
..........on - line
-)Y) estimate I b = 0.5 h"1
+CCIC( estimate 1 b = 0.9 KII
71' 150
'5
Main population estimates
-
751
50
Contaminant estimates
Fig. 20. Case Study 2 of SSPE: Esti-
mation of biomass concentration of
Escherichia coli in E. coli contami-
-
OI
VI
rn
VI
s
5 25
-
- Xz est. lb.0.5 h-')
Xz est. lb.0.9 h-ll
nated yeast culture by the SSPE algo-
rithm with different settings of the b-
value. (m) off-line measurements of to-
tal biomass; (. * .) on-line measure-
ments of total biomass using Eq. (36);
(0) off-line measurements of E. coli
biomass. Note that the algorithm de-
tects the contamination more sensitive-
0 ly with a smaller value of b selected.
0.0 5.0 10.0 Figures from CHATTAWAY and STE-
Time Ih I PHANOPOULOS (1989).
AXELMUNACK
Braunschweig, Federal Republic of Germany
It may be argued that a suitable choice of ple configuration of such a control system.
sampling period may not be a crucial task for The controlled variable y ( t ) is compared to the
biotechnical systems. The processes show a reference input w ( t ) to form the error e ( t ) .
very slow dynamic behavior, which allows From this continuous signal only a sampled se-
measurements to be made quasicontinuously. quence is supplied to a digital computer that
However, this fact should not induce an imme- performs the necessary calculations to control
diate jump to conclusions. It should be kept in the plant. These computations result in a dis-
mind that data from bioreactor measurements crete time output signal u (k T ) that is fed into
show quite different dynamic behavior; the re- the digital/analog converter, which holds the
sponse of the gas holdup and the oxygen con- output steady during the next sampling inter-
centration in the liquid phase to changes in the val, resulting in u ( t ) = u ( k T ) ,k T s t < ( k + l)T.
aeration rate is rather rapid and cannot be Thus, the process is driven by a staircase
controlled precisely by taking samples once a function which better approximates a contin-
minute. This shows that the non-biological uous signal, the smaller the sampling interval
control loops for a bioreactor may show fast T is. The purpose of the control loop is two-
behavior and must be sampled with sufficient fold: on the one hand, the controlled variable
frequency. should follow changes in the reference input
On the other hand, one must consider that almost precisely and with good dynamics
most of the biological data can only be meas- (tracking problem), and, on the other hand,
ured with some lag time and with relatively the influence of various disturbances n ( t ) on
low sampling rates. This is true even when the controlled variable should be kept very
auto-analyzers are used and is even more so small (disturbance rejection).
for manual analyses. Thus, the situation arises A rigorous means to study the behavior of
that, although the biological processes are very discrete-time systems is provided by the z-
slow, sampling may be so infrequent that a transform. This is a transformation of signals
good reconstruction of the biological state of in the time domain onto complex functions. It
the system is not possible. permits one to treat systems in the transformed
Since manual analyses are costly and auto- domain in a manner similar to that calculus
analyzers also require large amounts of materi- which is provided by the Laplace transform
al and manpower, one must plan experiments for continuous-time systems. The transformed
and the type and number of samples carefully. signal representation for a discrete-time signal
This chapter provides some basic considera- x ( k T ) is given by
tions for a suitable choice of sampling time,
m
some advanced calculations for optimization
of sampling, and applications to parameter i$(x(k))=x(z)= x(kT)*z-k (1)
k=O
identification problems for models of biotech-
nical processes. where z = ers.
It is impossible and at the same time beyond
the scope of this chapter, to provide a general
background of the theory of these systems. An
excellent textbook on this topic was written by
1 General Remarks ACKERMANN ( 1985).
In the specific field of application treated
Concerning a Suitable here, however, at least one fundamental ques-
Choice of Sampling Time tion should be answered: is it possible to re-
construct the continuous signal x ( t ) from the
samples x ( k T)? Equivalently, the question
For a discussion of the various aspects, may be posed: under what conditions does the
which must be considered for a reasonable inverse z-transform give an unambiguous re-
choice of the sampling interval, one should sult? The answer is provided by the sampling
first understand some basic principles of sam- theorem (SHANNON,1949), which states that a
pled-data control systems. Fig. 1 shows a sim- reconstruction is possible, provided the Fou-
- - computer Hold Process
~~
rier transform X u w ) of x ( t ) holds X u w ) = 0 sidered. This is the set of all initial states x(to)
for I o I 2 om,and the sampling period is cho- which may be transferred to the zero state
sen such that T s d o , , resulting in a Sam- within a certain number of time steps by a
pling frequency w , z 2 o , . This gives a first bounded input sequence u (k). ACKERMANN
hint as to how the sampling time may be cho- (1985) gave several examples of computed con-
sen. However, a different rule of thumb for trollability regions in which the area of the re-
the choice of the sampling interval is provided gion is taken as a measure, depending on the
by considering not only the signals but also the step number. These calculations, normalized
controlled system. The concept important in to that sampling period r where controllability
this context is that of controllability, which is entirely lost, demonstrate that a great in-
refers to the following property: A state x(to) crease in the controllability region occurs for
of a linear system is controllable if there exists T = r / 2 , r/3 and r / 4 . Afterwards, the region
a finite time instant tl>to and an input u ( t k ) , grows further (in the limit - as T approaches
O s k < l , such that the state x(to) is transferred zero - to the controllability region of the cor-
to the zero state x ( t l ) = 0. responding continuous-time system). The
The concept of state-space representation of amount of additional increase, however, is not
dynamical systems has not yet been intro- very large.
duced. Roughly speaking, the state of a sys-
tem, denoted by the state-space vector x, may
be interpreted as the actual charge of all mass,
energy, or momentum reservoirs of the system tims
(or a transformation of it). For example, the
biological growth process in Sect. 2 is fully de-
scribed by the state x = (x,s)=, x and s denot-
ing cell and substrate concentration, respec-
tively (however, this is a nonlinear system).
For a completely controllable linear dis- Fig. 2. Determination of the sampling period.
crete-time system it can be shown that n steps
are always sufficient to transfer the system to
the origin, n being the dimension of the state
vector. However, the corresponding input sig- Therefore, as a rule of thumb, the sampling
nal u ( k ) will often show large amplitudes in frequency should be chosen four times larger
order to achieve this goal, which does not than the frequency at which controllability is
match the real situation of bounded inputs (see lost, which corresponds to the frequency
Fig. 1) where u is the output of a digital/ana- o,= 2w, of the sampling theorem. The result-
log converter which is always bounded. Of ing rule for design of sampled-data control
course, further bounds are implied by elements systems is summarized as follows (Fig. 2):
of the loop, e.g., input flow pumps, valves,
etc. This means that the controllability region 0 For a continuous process with eigenval-
of systems with bounded inputs has to be con- ues si, one may draw the smallest possi-
254 8 Optimization of Sampling
ble circle in the s-plane around the ori- The class of models Acontains continuous-
gin s = 0 such that all the poles si are en-
time systems with a nonlinear dynamic state-
circled. The radius of this circle is r; space description and a nonlinear static output
then o,= 8r is chosen, which means that relation. The structure of the system is fixed
.T= 71/43. by a-priori knowledge, which means that only
system parameters are treated as unknown. In
Further recommendations of ACKERMANNorder to keep the presentation relatively sim-
refer to anti-aliasing filters or neglected dy- ple, lumped-parameter systems are treated; ex-
namics. These are relevant to mechanical sys- tension to the distributed-parameter case is not
tems but need not be considered for applica- crucial, cf. the case study in Sect. 3.
tion to biotechnical processes.
d ( P ) :2 = f ( x ,t , u ,P ) , x(0)= xo ( P ) (24
A (E{(p*-P )(p*-P )'}) = A ( V)-'min (6) denotes the output sensitivities with respect to
parameter variations, evaluated along the
where A:IRpxp+IR is a functional used to nominal output trajectories. The state sensitiv-
weight the covariance V . ities X p are computed by solving the linear sys-
To solve this problem, questions regarding tem
optimization of the information content of
measurements require discussion. This leads to
an optimized choice of the weighting matrices
Qi and a discussion of the experimental condi-
tions when taking the measurements. For dis- where the derivatives are taken along the nom-
tributed-parameter systems the question of the inal trajectories x o ( t , U,P ) , with the initial
best position of the sensors may also be condition
solved.
Some assumptions have to be made con- dX0
Xp(0) = -
cerning the noise that disturbs the measure- dP
ments. Zero-mean Gaussian white noise is con-
sidered with a diagonal covariance matrix If one chooses the weighting matrix Q(ti) as
C(ti),which means that the 'true' system out- the inverse of the covariance matrix C, then
puts y are additively disturbed to give the the matrix formed by evaluating the sum in
measurements: brackets of Eq. (8) is just the Fisher informa-
tion matrix F of the estimation problem
(LJUNG, 1987). This matrix is the inverse of
the parameter estimation error covariance ma-
where trix of the best linear unbiased estimator
(BLUE). Therefore, it gives an upper bound
E { ~ ( t ~ ) } = i0=,1 , ...,N and (7b) for the precision of the estimate obtainable by
a certain experiment. Eq. (8) gives:
E { E (ti)E ' ( t j ) } = 6ij * C ( t J (7c)
i,j = 1,. . .,N E { JI(u,P + Sp)}= Sp'F(u, P ) S p + N m (11)
a non-informative experiment, det (F)= 0 or the case when using other types of approxima-
Amin (F)= 0, Amin being the smallest eigenvalue tions of functions, such as polynomial ap-
of F, could be used as an indicator. Moreover, proximations.
not only these qualitative but also quantitative Some further remarks on the measurement
statements may be obtained. To provide the noise must be made. Here this noise reflects
largest possible distance to the singular non-in- not only the classical noise generated by sen-
formative case, det (F)(D-criterion) or Amin(F) sors but also takes into account the precision
(E-criterion) can be maximized. Other criteria of the measurements (though a Gaussian dis-
known from the literature refer to a maximiza- tribution with a zero mean may be somewhat
tion of the trace of F (simplified A-criterion) crude to describe this effect). Nevertheless, in
or to a minimization of det(S) (D-criterion), some cases one is not able to model these er-
A,, (S) (E-criterion) or tr (S) (A-criterion), rors by a constant noise level; on the contrary,
where S = F - ' . (In the case of the BLUE, a relative error would be more adequate. This
V = S holds.) The A-criterion may be inter- leads to a consideration of semi-relative sensi-
preted as a minimization of the identification tivities:
errors in an arithmetic mean, and the D-criteri-
on optimizes the geometric mean. The E-crite-
rion, however, minimizes the largest error.
Whereas the application of the D- or E-criteri-
on is possible with respect to the covariance which may also be incorporated into Eqs. (8)
matrix S or the information matrix F (with the and (1 1) by setting
same results), the A-criterion gives different
results. The simplest criterion tr(F)+max
should not be used, since it may lead to non-
informative experiments (GOODWIN,1987).
Thus, the functional of the optimization prob- For small values of the nominal output (yo),
lem, Eq. (6), is seen as a D- or E-criterion. this relative weighting may lead to overly op-
The quantitative evaluation of the informa- timistic results, since the error approaches zero
tion content of measurements from an experi- as the trajectory itself goes to zero. Therefore,
ment offers the opportunity to optimize the ex- a modification should be used instead of
perimental conditions, which means finding Eq. (13):
input functions u ( t ) that lead to the most in-
formative experiments. In the dynamical case,
this problem is in principle an infinite-dimen- (14)
sional optimization problem, since functions
in the time domain must be optimized. In or- A further modification may be carried out,
der to reduce the numerical burden, the input which takes into account the fact that in gener-
functions may be discretized, such that only al the desired accuracy of the estimated param-
values at a finite number of node points must eters is not an absolute value but a relative
be determined. Between the node points, the one. This is taken into consideration by (fully)
function may be considered as linear in time (if relative sensitivities, when Eq. (12) is modified
such a permanently changing feed-rate is pos- to give
sible with the installed process equipment), or
as a constant. Both possibilities provide simple
means for transforming the problem of com-
putation of an optimal input function in the
time domain into a parameter optimization Below, an application of the above formulated
problem. However, since usually various re- theoretical considerations to a simple biotech-
strictions must be observed (e.g., input flows nological growth process is considered, follow-
are always non-negative and bounded), these ing the results presented by MUNACK(1989).
may be easily incorporated in such a simple The process is described by two ordinary dif-
representation of functions. This would not be ferential equations,
Criteria to Optimize Measurements for Model Identification 257
x ( t )= p (s) * x ( t )-pD * x ( t ) (16)
1
k(t)= --*p(s)'x(t)
YX, s
where x denotes the cell concentration and s
the substrate concentration. Both states form
the state vector Y = ( X , S ) ~ , which - at the in-
stant of measurement - directly gives the un-
disturbed output vector of the system. pD is a
decay rate, and YX,+is the yield coefficient.
For ,u(s) a Michaelis-Menten-type nonlinear
relation is assumed to hold, which gives
Fig. 4. Plot of the undisturbed identification func-
tional for a batch experiment.
O-
2
r 4
4
6
6
Time i h I
8
8
10
I0
12
12
tifiable, a fact which was reported first by
HOLMBERG(1982). All estimates result in
good agreement of measured calculated data.
The identified parameters, however, are bio-
logically meaningless.
Fig. 3. Trajectories of the batch cultivation. Now the above described procedure for op-
timization of experimental conditions is ap-
plied to the process. A batch process offers
Tab. 1. Parameters Used for the Batch Cultivation only one usable degree of freedom to optimize
of Fig. 3 the experiment in order to permit better esti-
mates of the parameters. This is the initial sub-
pm =0.5 h - '
pD =0.05 h-' x(O)= 1 g . L - 1
strate concentration s(t=O). The main effect
Yx,s= 0.6 ~ ( 0 ) = 5 0g . L - ' on the performance of the estimates of param-
Ks = 3 g * L - ' eters, however, is provided by turning the
process to fed-batch operation and taking the
258 8 Optimization of Sampling
1
- -* p (s) . x ( t ) + - * q ( t )
SR
s(t) =
yx, s V
3 Application
to Cultivation in a Tower
Loop Reactor
The application of the methods of Sect. 2 to
distributed-parameter processes has already
been mentioned. Those are processes which
show distinct profiles of their state variables
along the spatial coordinate(s). For most bio-
technological applications, one may think of
concentration profiles which occur in reactors
that are not of the stirred-tank type. However,
temperature profiles may also have to be con-
sidered, when induction by temperature shift is
carried out in tubular reactors. Distributed-pa- Fig. 8. Schematic diagram of the tower loop reac-
rameter processes are naturally modelled by tor.
partial differential equations (PDEs). Most au-
thors agree in preserving the distributed nature
of the process in the calculations as long as used double-reactor system (column and loop)
possible. A ‘rapid lumping’, which means a di- is shown in Fig. 8. Since only the column is
rect approximation of the spatial partial deri- gassed from the bottom, gas phase (G) bal-
vatives in the PDE by difference formulas and ances are formulated only for this part, while
a further treatment of the resulting system of liquid phase balances appear in the column (F)
ordinary differential equations, may some- and in the bypass (B). For modelling the
times be worse than the immediate formula- growth process, only mass balances are taken
tion of a compartmented model. This is due to into account, since isothermal conditions are
the nonlinear terms which usually occur in the assured by control, and momentum is negligi-
description of biotechnical processes. There- ble. The task of the controller is to achieve
fore, we are led to the alternatives either to maximum cell production at minimal cost,
treat partial differential equations as long as which means lowest possible substrate (S) and
possible or to describe the process directly by oxygen (0)consumption.
formulating balances for compartments and A general model, based only on the nu-
their mutual exchange of mass and energy. trients O2 and S, the products X (cell concen-
In the following, the first strategy is applied tration) and C02, and the spatially varying gas
to a microbial growth process in a tower loop velocity uo leads to a nonlinear system of six
reactor (bubble column). For a more detailed parabolic partial differential equations and
treatment of this type of reactor, we refer to five plug flow equations (LUTTMANN,1980;
SCHUGERL (1985). A schematic diagram of the LUTTMANN et al., 1985). For control purposes
260 8 Optimization of Sampling
o
to be justified for the pilot plant used, the
model is drastically reduced to a quasi-linear
PDE of parabolic type, describing the dis- Normalized dissolved oxygen concentration in
solved oxygen concentration cOFin the liquid the loop
phase of the column, Eq. (21), an algebraic COB(2) - 2
equation for the oxygen concentration cOGin
COB (2) - cOB(O) + ko*ln-
COB (O)
- - qo/xrn-cx
US
the gas phase of the column, Eq. (22), an im-
plicit algebraic equation for the dissolved BC: COB@) = C O F ( ~ ) (23)
oxygen concentration cOB in the loop, Eq.
(23), and an ordinary differential equation for
the biomass concentration cx, Eq. (24). Normalized concentration of cells
IC:cx(O)= 1 (24)
The system equations are coupled via the states
and the boundary conditions. This means that
cx enters into the dissolved oxygen equations,
and, on the other hand, C O and
~ COB enter into
the cell equation. cOF and COG are also cou-
pled, and the physical connection of reactor
column and loop is modelled via coupling
=O terms in the boundary conditions.
Application to Cultivation in a Tower Loop Reactor 261
Four parameters have turned out to be un- adjoint state equations were used to compute
known and/or temporally varying. These are the gradient of the functional with respect to
two fluid-dynamic parameters, kLaEand K,,, the unknown parameters.
both describing oxygen transfer from the gas Details concerning these problems will not
phase into the liquid phase along the reactor be discussed here. Instead, emphasis will be
column, and two biological parameters, the placed on problems concerning the optimiza-
metabolic quotient qo/xm and the yield coeffi- tion of sampling of measurement data. The
cient Y,,,. Measurements can be taken of dis- questions to be solved in this context are as
solved oxygen concentration cOF at distinct follows:
points in the reactor column, cell concentra-
tion and outlet gas mole fractions, allowing How many sampling points for measur-
computation
- of the overall oxygen transfer ing the dissolved oxygen concentration
rate OTR. This computed value is treated as a along the reactor column should be
further measurement. Thus, the identification used?
functional (output least squares error criteri- Which are the best positions along the
on) is formulated as follows, where denotes A reactor column to install these sensors?
model outnuts with estimated Darameter set Which type of measurement is most val-
uable for identification of the unknown
parameters? Which data set is essential
~r (ri> = 0J I
~ ( t )
[iyl
(xi,
+ 0 (OTR ( t )-OTRM(t))' +
*
t>- C& (xi, t)12 +
for identification?
11
+ ~ * ( & ( t ) ) - c ? ( t ) ~dt (25)
distributed parameter process, this method has
been discussed in the literature by QURESHIet
al. (1980). A great variety of related methods
The numerical treatment of this problem of has been compiled in the survey paper of KUB-
parameter identification, in particular the for- RUSLY and MALEBRANCHE (1983).
mulation of very efficient algorithms for its The D-criterion was used to evaluate the in-
on-line solution, has been discussed in detail formation contents of the different possible
by MUNACK(1986). The algorithms refer to measurement data sets. The results are sum-
theoretical work on identification of parame- marized below. Fig. 9 shows the computed de-
ters in PDEs by CHAVENT (1974), in which the terminants as a function of the position of the
cOFsensor(s) along the reactor column. Curve
(1) gives information which may be obtained
by using a single cOFsensor as the only meas-
3 4x10
urement device. It can be seen that this sensor
2+x+0
is best placed at the bottom of the reactor (the
1+x+0 normalized position being x / L R= 0). This re-
sult does not change when further measure-
- 5-
LL - ments are additionally used, namely, the
curves (1 + X ) for additional cell concentration
measurements, (1 + 0) for additional OTR
~
OTR measurement.
When the number of oxygen sensors is in-
creased, the optimal locations are not spread
uniformly over the whole spatial area of the
reactor. Using more and more sensors partly
leads to the same positions as already occupied
by the other sensors. This is due to an assump- -25
tion made for the calculations stating that the 0 1
O5 X/LR
noise of different measurements is not corre-
lated. One would have to be careful with the Fig. 11. Information contents of various sensor
installation of the sensors in order to fulfill combinations as a function of the cOF sensor posi-
this condition in practice. tion in case of cyclic identification.
Extension of the Results for Process Control and optimization 263
__
the high information content of OTR measure- Then the partial derivatives
ments in the example treated here.
The case study carried out in this chapter
demonstrates the benefits gained by a suitable
choice of equipment for measurements in-
?I p , i=l,...,p
stalled at a bioreactor. It turned out - that a denote the influence of changes in the control
'lumped' measurement - in this case OTR - performance due to small changes of the sys-
may be of highest value for identification of tem param;ters around the estimated parame-
distributed-parameter systems. This fact would ter vector P. It is obvious that parameters that
have been underestimated in most cases, even influence the performance very strongly
if experts had been asked. In critical situa- should be identified with higher precision than
tions, particularly for on-line identifications, parameters which result in smaller deviations.
an analysis of the information provided by dif- This may be achieved by weighting the infor-
ferent measurements should, therefore, be car- mation matrix before carrying out the optimi-
ried out. However, the computational effort zation (TAKAMATSU et al., 1971), giving
required to gain the results reported here is
quite high. There is a definite lack of standar- i"(U,P) =
dized software to perform the calculations.
Further research is needed.
KARL-HEI N Z BELLGARDT
Hannover, Federal Republic of Germany
1 Introduction 269
2 Principles of Model Building for Biotechnological Processes 270
2.1 Kinetics and Types of Fermentations 270
2.2 Types of Biological Models 270
2.3 The Biological System Cell in the Modelling View 271
2.4 Identification of Parameters 272
3 Unstructured Models on the Population Level 273
3.1 Simple Growth and Substrate Uptake Kinetics 274
3.2 Substrate-Independent Growth Kinetics 275
3.3 Substrate and Product Inhibition 275
3.4 Multiple Limitations 277
3.4.1 Essential Substrates 277
3.4.2 Growth Enhancing and Alternative Substrates 278
3.4.3 Combinations of Essential and Alternative Substrates 279
3.4.4 Comments on Multi-Substrate Kinetics 279
3.5 Deviations from the Constant Yield Case 280
3.6 Influence of Other Variables on Growth 282
3.6.1 Mechanistic Description of Temperature Effects 282
3.6.2 Black-Box Models for Unknown Mechanisms 282
3.7 Primary Metabolite Formation 283
4 Structured Models on the Cellular Level 284
4.1 General Formulation 284
4.2 Examples of Structured Models 285
4.2.1 Comparison of an Implicitly Structured Model with a Two-Compartment Mod-
el 285
4.2.2 Compartment Model for Antibiotic Production 288
4.3 Cybernetic Models 290
4.4 The Metabolic Regulator Approach 292
5 Conclusion 296
6 References 297
268 9 Cell Models
Introduction 269
composition of the biomass and the state of The physical and biochemical processes in
the intracellular systems. But the simplifica- bioreactors are described by balances for
tion can also cause errors in the static and dy- mass, impulse, and energy, and by the related
namic behavior of the model. Particularly, conservation laws (ROELSet al., 1978). For a
time averaging results in sizeable errors in the unified modelling approach the equations are
presence of metabolic regulation or synchro- written in vector notation. Let C be the col-
nous growth of the culture, in which metabolic umn vector of concentrations in the liquid
variation during the division cycle becomes phase of the reactor, then in the above exam-
globally visible. Model errors from population ple C is defined by:
averaging can appear if the cells cannot adapt
at once to new growth conditions. c=(CS,c,, co, cc,CXlT
For inclusion of some of the above effects,
and if the cellular relaxation time is of the Generally the concentration vector can be div-
same magnitude as the duration of the process, ided into concentrations of the biotic phase,
the internal state of the cells must also be con- X , and of the abiotic phase, 2
sidered. This leads to so-called structured
models. This type of model is presented in
Sect. 4. The models can be structured on the
basis of biomass components such as concen-
c= [3 (1)
0 Cell mass, ,C
, autocatalytically synthe-
sized from the provided substrates, dVL(t) - FI(t)-Fo(t)
--
0 substrates, C, as energy and nutrient dt
suppliers,
0 products of metabolism, C,, inside or where FI and Fo are the inflow and outflow
outside of the cells, rate of the reactor, VL is the liquid phase vol-
0 dissolved gases, mainly oxygen, Co,and ume, CI is the concentration in the inflow, Q is
carbon dioxide, Cc, which are connected the vector of reaction rates in the liquid phase,
to the gas phase by mass exchange. and P is the mass exchange vector with the gas
272 9 Cell Models
phase. With the above model equations differ- time-dependent vector p ( t ) , but usually p is
ent kinds of processes can be described: constant.
Throughout this chapter, the models will be
Batch fermentations with given for specific reaction rates. For a more
FI(t)=Fo(t)=O,VL=const, convenient notation, a distinction will be made
fed-batch fermentations with FI( t )# 0, between elementary, independent specific reac-
Fo ( t )= 0, VL # const, tion rates, r , for which kinetic expressions
semicontinuous fermentations with have to be specified, and the entire reaction
F,(t)#Fo(t)#O, VL#const, vector, q , which can be calculated from r by
continuous fermentations with stoichiometry. Due to conservation laws and
D . VL = FI( t )= Fo ( t )# 0, V, = const. stoichiometric laws, the relation between the
globally visible net reaction rates in Eq. (6)
After one has chosen the vector of concentra- and the intrinsic specific reaction rates,
tions, C, and established the corresponding r(C(t),p ( t ) ) , within the microorganisms, can
balance equations, the next step of modelling be written as:
is to describe the biological reaction, Q , by the
cell model. The aim of this chapter is to intro-
duce the principal ideas for models of the bio-
logical system. The transport step, P , will not where Y is the matrix of stoichiometric or yield
be considered, and it will be assumed that no coefficients, which may vary with time but is
other reactions than those catalyzed by the mi- usually taken as constant. In this nomencla-
croorganisms take place. ture C ( t )andp(t) are the input variables of the
A general characteristic of microbially cata- biological model, and the net reaction rates
lyzed reactions is that the total volumetric q ( t ) are the output variables. In Fig. 2 the cor-
reaction rate is proportional to the amount of responding structure of a model of a biotech-
active biomass. The biological models are ad- nological process is shown. It consists of the
vantageously formulated using specific reac- liquid phase model, a gas phase model, and a
tion rates, defined by: cell model. Taking the biological model as a
local microkinetic model, it should be indepen-
dent of the type of reactor and mode of opera-
tion, and also independent of the reactor mod-
els. In the cell model, the microorganisms are
Besides the concentrations, other time-varying represented by kinetic equations for the reac-
operating parameters can also influence the tion vector r. Unstructured cell models are al-
microbial reactions. This is considered by the gebraic ones, while structured models also in-
clude further differential equations for the
biological state vector X .
Reactor model
2.4 Identification of Parameters
Model building is always combined with the-
'
I--
- Q
Reactions
___-----A
model
oretical studies and practical experiments. The
establishment of complex models is a very iter-
ative process (MOSER,1981). Since the prob-
lems in parameter identification and model
verification increase rapidly with the complexi-
Intrinsic con,ce?trations
ty of the model, one should begin with maxi-
I YBaiances mally simplified assumptions and withdraw
L________ J
them step by step in case that the model quali-
Fig. 2. Structure and elements of mathematical ty is not sufficient. Modelling includes not
models for biotechnological processes. only the selection of the correct model struc-
Unstructured Models on the Population Level 273
Process
ture, but also the quantitative determination tion techniques in conjunction with numerical
of the model parameters. Unfortunately, their optimization methods have to be used.
values are often not known or only inexactly Parameter identification by output error
known in advance and must be determined by methods does not evaluate the structure of the
fitting the model to experimental data. For model and the accuracy of the estimated pa-
simple models, regression methods such as rameters. Thus, a good fit may be achieved
Lineweaver-Burk plots can be used, but for even with wrong parameters, especially for
complex models more general methods of pa- non-linear systems, high model order, and
rameter identification have to be employed. only a few measured values. The limited accu-
The task of parameter identification is to es- racy of measurements not only hinders param-
timate unknown model parameters by compar- eter identification, but also discriminates be-
ing process (experiment) and model according tween different models. This should always be
to a given performance criterion. The identifi- kept in mind during model building.
cation procedure should be able to eliminate
disturbances of the measured variables. Off-
line methods for parametric models are parti-
cularly interesting for model building. Usually
the very general method of minimizing the var- 3 Unstructured Models
iance of the output error (least square estima-
tion) is used (FASOL and JORGL, 1980). A on the Population Level
schematic diagram of the method is shown in
Fig. 3. The parameter identification serves to In unstructured models, the biological reac-
minimize the performance criterion: tion depends directly and solely on macroscop-
ic variables that describe the conditions in the
N
fermentor. The only biological state variable is
J(e,p) = 2 e'(ti)* w.e(ti)L m i n (8) the cell mass concentration, .C, Nevertheless,
i= 1
many phenomena in biotechnological proc-
which depends on the parameter vector p and esses can be covered by this type of model. Be-
the error signal: sides the cell mass, only those other variables
have to be considered in the model that show
great variation during the fermentation and
which have significant influence on microbial
Here, N is the number of measured values, ti behavior. In the following, the focus will be on
are the measuring times, yMthe model predic- models for substrate uptake kinetics, including
tions, yo the measured values, and W is a multiple limitations, inhibition effects, and
weighting matrix. The usual regression method product formation. Due to the extensive omis-
is derived from Eq. (8) as a special case. For sions and simplifications of cell metabolism,
complex non-linear systems, numerical simula- these kinetics should be taken as formal kinet-
274 9 Cell Models
ics which give only an integral description of growth rate have been proposed, as summar-
the process and not a detailed picture of the ized in Tab. 1. In many cases the ratio of cell
underlying mechanisms. mass formed to substrate consumed, the yield
coefficient, Yxs,is constant as a good approxi-
mation. For a single substrate and cell mass as
3.1 Simple Growth and Substrate the only product, C is given by:
Uptake Kinetics c=(Cx,WT
During growth, new cell mass is formed au- In this simple model there is only one intrinsic
tocatalytically from substrate with the specific reaction, r = p ( C s ) , and therefore Eq. (7) be-
growth rate p: comes:
Blackmar
* 1.2
u
Maser r: 0.5
Fig. 4. Normalized p(Cs)characteristics of
of
substrate uptake kinetics 7 versus substrate
concentration C,, normalized to the half-satu- 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
ration constant. Normalized substrate concentration C5
teristics of the kinetics is given in Fig. 4. In Such kinetics can be interpreted as an inhibi-
practice, because of measurement errors, it is tion by the cell concentration or a space limita-
difficult to discriminate between the different tion. For this type of fermentation, the kinet-
kinetics, especially when only using batch cul- ics of FRAMEand H u (1988) can give a better
ture data. Therefore, the Monod kinetics is description than the logistic law because it ex-
generally a good choice for the model. hibits a sharper transition to the stationary
phase. Another application of these kinetics is
for processes in which information on the lim-
3.2 Substrate-Independent Growth iting substrate is not available, e. g., the limita-
Kinetics tion of an unidentified component in complex
media.
In certain cases the application of substrate-
independent kinetics given in Tab. 2 can make
sense. The origin of the kinetics can be found 3.3 Substrate
in modelling of natural populations. Also in
some fermentation processes, the growth rate and Product Inhibition
does not mainly depend on a limiting sub-
strate. For example, in cultivations of mam- Besides substrate limitation, inhibition by
malian cells on micro-carriers the cells need substrates or products is quite often found in
free surfaces to colonize and grow. Growth biotechnological processes. Both have been
stops if no unoccupied surface is available. examined by many authors. A short review is
given by HANand LEVENSPIEL (1988). Tab. 3
gives a list of normalized inhibition kinetics.
Tab. 2. Substrate-Independent Growth Kinetics
Most of the applied kinetics are extensions of
the Monod equation and have been derived
Name Year Normalized Kinetics 5 from enzyme inhibition kinetics (DIXONand
WEBB, 1967). These kinetics cannot predict
Logistic zero growth for a finite inhibitor concentra-
law
1938 1 -- c x tion. Therefore, other empirical equations
Cx,max
have been proposed for the description of this
1 -- c x behavior. The last five equations in Tab. 3 pre-
CUI and Cx,ma.x dict a zero growth rate at C,=K,. Product,
1982
LAWSON
1 -- c x substrate, and cell inhibition can be obtained
Cx,lirn by choosing the variable C, as C,, C,, or C,,
respectively. The competitive type only applies
FRAMEand 1988
Hu for substrate inhibition. For non-toxic sub-
strates there is also no evidence of non-compe-
216 9 Cell Models
HALDANE 1965a
(competitive type) (1930)
WEBB 1963a
(competitive type)
IERUSALIMSKY 1965a
(non-competitive type)
EDWARDS 1970
CS I
YANOand KOYA 1973
(generally non-competitive type)
Teissier type
GHOSEand TYAGI 1952"
DAGLEYand HINSHELWOOD
LEVENSPIEL 1980"
titive substrate inhibition. The given equations where all terms psi have the form of Eq. (11).
can be generalized for multiple inhibitions by It should be noted that Eq. (14) is equivalent
combining the normalized kinetics to a prod- to a multiple Blackman kinetics if one of the
uct with several factors. terms is set to a constant value, e.g.,
,uSl =,urn,,. In practice there are also only slight
differences between the interacting and non-in-
3.4 Multiple Limitations teracting models. As an example, both are
compared in Fig. 5 , which shows the simula-
The above kinetics are valid for a single lim- tion of a glycerol- and oxygen-limited culture
iting compound. Several attempts have been with both of the following kinetics (from SIN-
made to develop models for multiple limita- CLAIR and RYDER,1984):
tions. Most of them are extensions of the ele-
mentary kinetics in Tabs. 1 and 3. There are Interacting model,
several possibilities for the interrelation of
growth rate and substrate concentration, de-
pending on the substrate itself and on the mi-
croorganisms. All substrates can be essential,
meaning there is no growth if only one of the
substrates is lacking; or one of the substrates is
sufficient for growth and others are used up in Non-interacting model,
parallel or in sequence. In any case, compared
with simple kinetics, the model has to consider
much more information on the structure of
metabolism. In the following, the extension of
single substrate kinetics to multiple limitations
is shown for the mentioned cases. The limits of
this approach and how these restrictions can In vector notation the entire cell model takes
be overcome will be discussed as well. the form:
'C
on glucose and maltose. Experi-
mental data (symbols) from
HOPF in BELLGARDTet al.
(1988), concentration of glucose
C, (x), maltose C,, (A), cell 00 3 6 9
mass C , (a), and specific 12 15
growth rate p. Time ( h I
As an example, the simulation of diauxic where I,K , L , and N are the numbers of essen-
growth of Klebsiella terrigena on glucose and tial substrates in the ith class.
maltose is plotted in Fig. 6 . The sequential up-
take of the substrates can be described in prin-
ciple, but the model cannot be fitted to the 3.4.4 Comments on Multi-Substrate
diauxic lag phase. In this model the delay of
growth on the second substrate is due only to Kinetics
an unrealistically low inhibition constant, K I l .
For comparison, a simulation of the same While the multi-substrate kinetics Eqs. (13)
process with a structured model is shown in and (14) for essential substrates seem to give a
Figs. 12 and 13. good description of observed phenomena, the
case for alternative substrates is not so clear.
To the author's knowledge, experimental data
3.4.3 Combinations of Essential to support Eqs. (17) and (19) are very rare.
The reason might be a too-simple approxima-
and Alternative Substrates tion of the growth kinetics due to the neglect
of the regulatory response of the microorgan-
The above kinetics can be combined to a isms. As already mentioned, a sequential up-
more general form, as proposed by TSAOand take is often found with the presence of excess
HANSON(1975). They grouped the substrates substrate. As long as the preferred substrate
into m classes of essential substrates, and in can support a sufficient growth rate, it sup-
each such class there were several alternative presses and inhibits the uptake system of other
or enhancing substrates. The entire kinetics is substrates. After it has become limiting, the
then given by: enzymes for the substrate with the next lower
preference are induced or derepressed. The re-
sult is a remarkable lag phase for the adaption
to the next substrate, which can only be de-
scribed by structured models. To make the sit-
uation more complicated, the same substrates
which are used up sequentially at high concen-
trations are also used up in parallel if the con-
centrations are low enough. Furthermore, the
uptake of several alternative substrates can be
affected by one or more essential substrates
280 9 Cell Models
such as oxygen. A well-known example is the or the degradation rate of intracellular storage
growth of yeast in batch and continuous cul- material.
ture (see Sect. 4.5). This means that many reg- PIRT(1965) and IERUSALIMSKY (1967) re-
ulatory phenomena have to be considered for a ferred the phenomena of the decreasing yield
model of multi-substrate kinetics that are very to an additional need for substrate consump-
closely associated with the pathways of the mi- tion for maintenance of the cell structure, ex-
croorganism. It can be concluded that Eqs. pressed by the specific rate m,. Thus, if the
(17) and (19) on the one hand are too simple uptake rate of substrate used only for growth
and on the other hand too general to cover the is qSG, then the total substrate uptake qs is:
difficult adaptation of the cells. A promising
idea of dealing with these phenomena is the cy- qS= qSG- mS (23)
bernetic modelling approach reported in Sect.
4.4. (note that production is defined as positive
rate) and the model in vector notation
Fc = P G - P E (20)
In the usual vector notation the model is: By comparing Eqs. (22) and (28) one can see
that both are equivalent in the regions of nor-
mal growth, if
PE
ms=-
YXSG
CXG(p) -- a ( p )
--
When describing the substrate demand of the CXD(P) I - d P )
growing fraction by the maintenance model
Eq. (25), the observed growth yield with re- and combining this with Eq. (35) gives:
spect to the total cell mass is:
(1987). The identified parameter values are volved enzymes. For a fast inactivation reac-
KD=0.0064 h-', Yxs,=0.55, and ms=0.045 tion, there is a n equilibrium between the ac-
h-I. tive, cEA, and inactive, cEI,form of the enzyme
given by:
tions between growth activity and certain pa- As an example, the contour plot for cell yield
rameters. In such cases a black-box approach from glucose is shown in Fig. 9. The corre-
is the easiest way to build a mathematical sponding correlation function is:
model directly by fitting it to experimental
data. A very flexible choice for the unknown r,, (e) = 0.495
functional relations is a sum of powers of the -0.015 *AT-0.194*ApHZ
..
parameters. For 8= (el,02,. ,On,.. . ,ON) and -0.057.AT *ApH
p = ( P I,PZ,. . .,PI,
. . .,pL),the general function + 0.006. A T2*ApH + 0.039. AT * ApH'
can be + 0.134 * AT .ApH3-0.144. ApH3
-0.006.AT .ApHZ-0.018*AT2*ApH3
PI(8) =
II Kl
J,
with AT=T-301K, ApH=pH-6.8
22
i=O j-0
* " 2
k=O
cjjk[*e',*e/,**'e& (43)
and rn,,,,, is the total maintenance coefficient former control the intracellular reactions.
for all maintenance-associated processes, re- When writing the balance equations for intrin-
spectively: sic variables, one has to take special care of
the consistency of the model. The derivation
presented here is based on that of ROELS
(1982).
The intrinsic concentrations are defined as:
Usually, when evaluating the model parame-
ters, the values for X,,,,, and rn,,,,, will be de-
(49)
termined, and not the values Yxs, and m,,
which may be taken as “theoretical” parame-
ters. The total yield already accounts for all From Eqs. (7) and (49) it follows for the sum
products, whether they are considered in Eq. of all components of c ( t ) :
(45) or not. The product formation then de-
pends solely on the parameters Ypx and mpx. IT.C(t) =1 (50)
and therefore,
d (1 * c (t))
=o
4 Structured Models dt
Therefore, the model is only valid for cell recy- 4.2 Examples of Structured Models
cle from the same fermentor or without inflow
of biomass. By partial differentiation of Eq.
(54) one obtains: In the following, two structured models will
be presented, one for diauxic growth on two
carbon sources and another for the production
of an antibiotic.
= Yx*r(t).C,(t)+
4.2.1 Comparison of an Implicitly
"L Structured Model
Multiplying ITby each summand and substi- with a Two-Compartment Model
tuting in Eq. (49) gives:
Unstructured models cannot provide a satis-
factory simulation of processes with lag phase.
An example that was already presented in Sect.
3.4.2 is the switch from the first substrate to
+
= I T * ( Yx * r (t)) * Cx(t) the second in diauxic growth processes. To im-
prove the model behavior while avoiding struc-
+ FI(0
-* I'.c(t)'(C*I(t)-C*(t)) tured models, an explicitly time-dependent for-
VL. mal kinetic model similar to Eq. (18) can be
or after simplification by using Eqs. (50) and applied:
(51):
P(G I (0,c S z ( 0 ) =
=PI ( G I (0)+/42(Csz(t),
(57 b)
where the second summand is time-depend-
ent,
where
(60)
p ( t )= IT* (Yx' r (t)) (58)
Here tl is the starting time for growth on the
is the specific growth rate. After considering second substrate, Csz. The first two factors
Eqs. (57b) and (58), in Eq. (56) the balance can be interpreted as a time-varying maximum
equation for the intrinsic concentrations be- growth rate:
comes:
- 0.4-.+IS- ~ 1 . 5 -p
L -I,- L
:E : S I : -
7
m2
-.0.3{
I: 5,:5
0.2 : I
- 0.5 0.5-
0.f Fig. 10. Simulation with a for-
mal kinetic model of diauxic
0- 0- 0- 0 growth of Klebsiella terrigena
on glucose and maltose (legend
Time tl h 1 see Fig. 6).
(66)
25
1.5
m l
0.5
\
n--j n.
0 3 9 3- 12 15
Timetlhl
- 02
cE, and cEzand specific growth rate 0 3 6 9 12 15 18
U. Timet Ih I
288 9 CeNModels
Starting from the given initial conditions, antibiotic myxothiazol (MXF 16), which is
cE2is approaching its constitutive level during formed at low growth rates under oxygen limi-
the first growth phase. Afterwards, its tran- tation. Since the role of secondary metabolism
sient controls the diauxic lag phase. and its regulation as a part of the entire meta-
bolism is, qualitatively, still not sufficiently
clear - because of the complexity of the prob-
4.2.2 Compartment Model lem - the model must be based on numerous
simplifications and reasonable assumptions.
for Antibiotic Production The structure of the cell model shown in
Fig. 14 contains three compartments for the
The production of secondary metabolites main pathways. The solid lines describe mass
such as vitamins, pigments, or antibiotics is flows and the dashed lines the effect of regula-
closely related to regulatory effects of the me- tion. In catabolism, the substrate is decom-
tabolism as a response to impaired growth posed into the collective metabolite, cM,from
conditions. For modelling such effects one can which the further cell components are synthe-
use structured models. Besides growth, they sized. Anabolism is for the production of mac-
can also describe the regulatory state of the romolecules and cellular material such as en-
cells and the pathways of metabolism leading zymes. Hence, output of this block must also
to products. This section presents a simple include enzyme 1. This catalyzes the substrate
mathematical model for the production of the decomposition process, and thereby controls
Structured Models on the Cellular Level 289
Anabolism
sumption about coupled kinetics by extrapola- The formation rate of the key enzymes is con-
tion from experiments on single substrates. trolled by the cybernetic variables ui and by
For formulation of the model, the microbial substrate kinetics:
growth on each substrate is represented by a
reaction:
-%(1 + Yxs,).C,
cx+csi (73)
where Eiis the representative pool of key en- A constitutive formation rate rEi,Oappears in
zymes for the catabolism of substrate i. The Eq. (78) besides the inducible formation rate.
enzymes have to be formed by the optimal This is due to mathematical requirements, but
strategy given later. The specific growth rate is also in agreement with experimental studies
on substrate i, pi(Csi), is a product of three showing that there is not complete repression
factors, the intrinsic concentration of the key of enzymes. Therefore, the adaption process
enzyme, cEi, a kinetic expression tGi,chosen starts with a minimum speed that is clearly
from Tab. 1, and an activity controlling varia- greater than zero. Now the balance equations
ble uGi: of the model for a chemostat become:
tional conditions have to be introduced to ob- The optimization strategy is a local one, since
tain a predictive model. These are derived in J only depends on the actual reaction rates. As
two ways. in the cybernetic models, the optimization
The first is to consider inherently rate-limit- strategy followed by the microorganisms is as-
ing steps of the metabolism, such as kinetics of sumed to be the maximization of the specific
substrate uptake systems or a maximum bio- growth rate:
synthetic capacity for macromolecules and cell
material. Rate-limiting steps are of special im-
portance for the modelling of microbial adap-
tion due to regulation of key enzymes by in- The metabolic coordinator is functionally
duction and repression. Besides this slow regu- equivalent to the control variables v ( t ) in Eq.
lation, there is a fast regulation by enzyme in- (74).
hibition. This means the turnover of a regu- To complete the modelling, the inherently
lated pathway may be lower than its maximum rate-limiting steps in Eq. (89) have to be speci-
turnover, which is determined by the concen- fied. If ri(t) is a substrate uptake step, then
tration of the key enzyme. Therefore, addi- rmmJt) can be chosen as any of the kinetics in
tional conditions for the actual reaction rates r Tab. 1 or 3. Also in this approach, there are
have to be formulated as inequalities for its no coupled kinetics independent of the com-
elements: plexity of the metabolism.
If ri(t)is a reaction controlled by a key en-
rmin,i(Osri(Osrmax,i(t) (89) zyme which is subject to induction or repres-
sion, then the dynamics of its regulation must
where rminand rmaxare the possible rates of the be given. In the cybernetic models this is done
inherently rate-limiting steps. When the equal by the balance equations (71) together with the
sign is valid, the reaction ri is not inhibited. cybernetic variables, ui. In this approach, a
Expressions for the minimum and maximum strategy is used which is local to the regulated
rates are given later. Usually rmi,(t) equals pathway. The metabolism is said to be in the
zero. Only if there is a reaction against the di- adapted state if none of the reactions ri(t)
rection defined as mathematically positive, which are controlled by key enzymes is rate-
may rmi,( t ) contain non-zero negative ele- limiting. This means that the concentration of
ments. the key enzyme has to be increased by further
The second way to derive more conditions induction, as long as the coordinator J fully
and finally to solve Eq. (88) uniquely is to in- uses the capacity of the pathway up to the ac-
troduce an optimality criterion, the metabolic tual maximum rate. If the actual rate allocated
coordinator J , for the immediate metabolic to the pathway is not rate limiting, then the
regulation by inhibition and activation, which key enzyme is inhibited. In this case the con-
has been left unspecified until now: centration of the enzyme should be reduced by
stronger repression to economize resources.
r(0
J ( r ( t ) ) Eqs. ( 8 8 ) and (89) 'Optimum (90) These conditions can be met by a system of the
form:
The meaning of Eq. (90) is that the metabolic
coordinator optimally determines the actual
rates of all pathways under the constraints giv- (93)
en by the structure of the metabolism, Eq.
(88), and the inherently rate-limiting steps, Eq.
(89). N-A4 reactions ri will always equal the
corresponding elements of rminor rmaX. A reac-
tion ri(t) of r ( t ) is said to be rate-limiting if where cEj is the intrinsic concentration of the
one of the following conditions holds: key enzyme for the pathway j , and roj is a low
constitutive level of enzyme synthesis. This
system can be viewed as a tracking controller,
a metabolic regulator for the concentration of
294 9 CeNModels
the key enzyme, which is regulated according As an example, the modelling approach is ap-
to actual requirements rj(t). Metabolism is plied to the growth simulation of baker's
viewed as a system of feed-back control loops, yeast. Yeast has a wide choice of metabolic
pathways to respond to environmental condi-
tions. Depending on the substrate and oxygen
supply, great changes in the stoichiometry of
I Rate limiting steps 1 1- growth can be observed, accompanied by dy-
Substrate4 u e t a w UCI U4I * , ,hl ,r3ta
a , c ,,
namic metabolic regulation. The model con-
Internal
a
Substrate-m uE
limitation $or'in'
Metabolic Substrate uptake
Product formation
p
siders the glucose uptake rate r,, the oxygen
uptake rate r,, the ethanol reaction rate (up-
-
toichiometric -Actual rates take rEc, or production rEp),and the carbon
Activity of - o f regulated
regulated I dioxide production rate rc. Intracellular reac-
--U
oat hwavs
pathways tions included in the model are: the rate of gly-
Metabolic
regulator 1 rltl colysis r,, of acetate production rAc, the repre-
rminlfl
rmaxlt)
sentative rate for the citric acid cycle rTc, the
! rate of gluconeogenesis rG, and the specific
rate of biosynthesis p = rx. The following stoi-
chiometric model can be derived from the stoi-
Fig. 16. Block diagram of the metabolic regulator chiometry of the metabolic pathways, together
model: metabolism as a system of control loops. with the balances for ATP and NADH:
rG
r0
rS
rAC
0
TTC
0
rEP
0 (97)
rE C
mATP
rS
0
rx
rC
as shown in Fig. 16, each consisting of a meta- Although this model can cover most of the
bolic controller, Eqs. (93) and (94), for exactly complex growth phenomena of yeast, it con-
one reaction. Additional conditions for the tains only eight independent stoichiometric pa-
constants can be derived to obtain the behav- rameters, all of which have a biological mean-
ior of the metabolic regulator described below. ing. In Eq. (97)' the maintenance requirements
If the pathway is not rate-limiting, Eq. (93) of the microorganisms are modelled anal-
ogously to Eq. (27) as a fraction mATp of the
has to be stable for all p ( t ) to keep rmanaxJ(t)
close to rj(t),and therefore: total energy consumption rATp due to growth
processes:
K1 <O (95)
45-
30-
-
c .
Fig. 17. Simulation with a meta- L
bolic regulator model of chemostat m
I
45-
30-
15-
OA
aoii
Ib)
nance concept Eq. (97) is a mechanistic union tured models only fail when intracellular dy-
of HERBERT'S (Eq. 20)) and PIRT'Smodels namics must be considered. Then the concepts
0%. (23)). of structured models can provide a better ap-
proximation to cellular metabolism.
Different ways of building structured mod-
els were presented. For many processes, the
5 Conclusion descriptive power of all kinds of structured
models may be quite similar, as shown by
This chapter has been a rush through a few some examples. Structured models must still
aspects of the modelling of the system cell in a be based on oversimplification of reality, and
biotechnological process. Unstructured, for- one should be careful when assigning model
mal kinetic models with their extremely simpli- variables to certain unique substances in the
fied view of the living cell can often provide an cell. A good point of view for model building
adequate description of growth dynamics, is to take the model for what it is: a vague pic-
even under non-stationary operation of tem- ture of the input-output behavior of a cell
perature or other process variables. Unstruc- population.
References 291
Compartment models as a special type of EDWARDS,V. H. (1970), The influence of high sub-
structured models have been successfully de- strate concentrations on microbial kinetics, Bio-
veloped for many processes. But it seems not tech. Bioeng. 12, 679.
ESENER,A. A., ROELS, J. A., KOSSEN, N. W. F.
to be useful to include more than four com- (1980), The influence of temperature on the
partments. Otherwise the number of paramet- maximum specific growth rate of Klebsiella pneu-
ers and of possible interactions between com- moniae, Biotech. Bioeng. 23, 1401.
partments increases rapidly. In contrast to FASOL,K. H., J ~ R G LM. , P. (1980), Principles of
compartment models, cybernetic models and model building and identification, Automatica
metabolic regulator models emphasize the dy- 16, p. 505, London: Pergamon Press.
namics of metabolic regulation. But at the FRAME,K. K., Hu, W. S. (1988), A model for den-
same time they completely neglect the mecha- sity-dependent growth of anchorage-dependent
nisms of metabolic regulation and substitute a mammalian cells, Biotech. Bioeng. 32, 1061.
very general optimization strategy. These FREDRICKSON, A. G., MAGEE,R. D., TSUCHIYA,
kinds of models are suprisingly successful in H. M. (1979), Mathematical models for fermen-
predicting growth on multi-substrates as well tation processes, Adv. Appl. Microbiol. 13, 419.
as changes in the product spectrum of primary GADEN,E. L. (1959), Fermentation process kinet-
metabolites. ics, J. Biochem. Microbiol. Technol. Eng. 1,
413.
HAN,K., LEVENSPIEL, 0. (1988), Extended Monod-
equation for substrate, product and cell inhibi-
tion, Biotech. Bioeng. 32, 430.
6 References HERBERT,D. (1959), in: Recent Progress in Micro-
biology (TENVALL,D., Ed.), Stockholm: Alm-
quist & Wiksell.
BELLGARDT, K. H., KUHLMANN, W., MEYER,H. IERUSALIMSKY, N. D. (1967), Bottle-necks in meta-
D. (1983), Deterministic growth model of Sac- bolism as growth rate controlling factors, in:
charomyces cerevisiae, parameter identification Microbial Physiology and Continuous Culture,
and simulation, in: Proc. 1st IFAC Workshop on 3rd International Symposium (POWELL,E. O.,
Modelling and Control of Biotechnological Proc- Ed.), p. 23, London: H. M. S. 0.
esses, Helsinki, 1982, New York: Pergamon LAM, J. C., OLLIS,D. (1981), Kinetics of multipro-
Press. duct fermentations, Biotech. Bioeng. 23, 1517.
BELLGARDT, K. H., HOPF, N., LUTTMANN, R., LEHMANN,J., BERTHE-CORTI, L., STEVEN,W.,
DECKWER,W. D. (1988), A new approach for NOTHNAGEL,J., PIEHL, G. W., GERTH, K.
structured groyth models, in: Preprints OJ (1979), Verfahren zur Vermehrung von Myxococ-
Fourth Int. Conf, on Computer Applications in cus fulvus DSM 1368, Ger. Patent Appl. No.
Fermentation Technology: Modelling and Con- 2924868.
trol of Biotechnological Processes, Cambridge, MASON,T. J., MILLES,N. F. (1976), Growth kinet-
UK, Chichester (UK): Ellis Horwood Ltd. Pub- ics of yeast grown on glucose of hexadecan, Bio-
lishers. tech. Bioeng. 18, 1337.
CHEN,B. J., LIM, H. C., TSAO, G. T. (1976), A MONOD,J. (1942), Recherches sur la croissance des
model for bacterial growth on methanol, Bio- cultures bacteriennes, Paris: Herrmann et Cie.
tech. Bioeng. 18, 1629. MOSER,A. (1958), The dynamics of bacterialpopu-
CHU, W. B., CONSTANTINIDES, A. (1988), Model- lations maintained in the chemostat, Publication
ing, optimization and computer control of the 614, Washington, DC: The Carnegie Institution.
cephalosporin C fermentation process, Biotech. MOSER,A. (1981), BioprozeJtechnik, Wien-New
Bioeng. 32, 277. York: Springer Verlag.
CONTOIS, D. E. (1959), Kinetics of bacterial PIRT,S. J. (1965), The maintenance energy of bac-
growth: Relationship between population density teria in growing cultures, Proc. R. SOC. Ser. B
and specific growth rate of continuous culture, J. 163, 224.
Gen. Microbiol. 21, 40. PIRT, S. J. (1987), The energetics of microbes at
CUI, Q., LAWSON,G. J. (1982), Study on models of slow growth rates: Maintenance energy and dor-
single populations: An expansion of the logistic mant organisms, J. Ferment. Technol. 65(2),
and exponential equations, J. Theor. Biol. 98, 173.
645. POWEL,E. 0. (1967), Proc. Microbial Physiology
DIXON,M., WEBB,E. C. (1967), Enzymes, 2nd. and Continuous Culture, Third Int. Symposium
Ed., London: Longman. (POWEL,E. O., Ed.), p. 34, London: H. M. S. 0.
298 9 Cell Models
MATTHIASREUSS
Stuttgart, Federal Republic of Germany
RAKESHBAJPAI
Columbia, Missouri 65203, U.S.A.
1 Introduction 301
2 Characterization of Mass and Energy Distributions in Stirred Tanks 304
2.1 Stimulus-Response Methods Based upon Tanks-in-Series Models 304
2.2 Flow-Follower Techniques 307
2.2.1 Measurements and Data Analysis 308
2.2.2 Estimation of the Mean Circulation Time from Simple Flow Models 310
2.2.3 Influence of Aeration 311
2.3 Tracer Responses from Circulation-Time Distribution 313
2.4 Circulation-Time Distributions for Two-Impeller Systems 3 15
2.4.1 Measurements Using Flow-Follower Techniques 315
2.4.2 Measurements Using Tracer Responses 319
2.5 Multi-Compartment Models for Coupling Oxygen Transfer, Mixing and Kinetics 320
2.6 Turbulence Models 327
2.6.1 Governing Equation for Two-Phase Flow 328
3 Applications to Microbial Systems 330
3.1 Oxygen Transfer to Molds 331
3.1.1 Problem of Energy Distribution 331
3.1.2 Mass and Energy Distribution 333
3.1.3 Multiple-Impeller Systems 336
3.1.4 The Role of Micromixing 338
3.2 Substrate Distribution in Baker’s Yeast Fermentation 340
3.2.1 Simulations of Fed-Batch Operations Based upon a Fixed Schedule for
Sugar Feeding 340
3.2.2 Influence of Mixing on Computer-Controlled Feeding 341
4 Conclusions 344
5 References 344
300 10 Stirred Tank Models
& energy dissipation rate exclusive and, therefore, do not allow an exact
&P volume fraction of particle replication of environmental similarity at any
fl effectiveness factor two different scales (AIBAet al., 1973; Kos-
P specific growth rate of cells SEN, 1985; KOSSEN et al., 1985; OLDSHUE,
Peff effective viscosity 1983; SWEEREet al., 1987). Under such cir-
PI parameters of log-normal distribu- cumstances, the behavior of microorganisms
tion in different fermentors remains uncertain. As
Pt turbulent viscosity a matter of fact, the use of volumetric proper-
V specific product formation rate, ties as scale-up criteria demands a guarantee of
dynamic viscosity uniformity of properties throughout the sys-
e mean residence time; mean of the tem, something that may be impossible even in
circulation time distribution a small reactor.
Ornix mixing time
w backmixing parameter, Eq. (54)
D2 variance of the circulation time Tab. 1. Common Criteria for Scale-up of Stirred
distribution Bioreactors
0; normalized variance of the circula-
tion time distribution Volumetric oxygen transfer coefficient k,a
@ Thiele modulus Volumetric power input P/ v
w general modulus Volumetric gas flow VG/ v
@P pumping capacity Impeller tip speed n di
Agitation speed n
Terminal mixing time 6,
Transport Processes
1
a) Gas-liquid mass transfer tOT = ~
5.5-11.2
kL a
V/ni,p
b) Circulation time tc = - 12.3
1.5 nd:
V
c) Gas residence time tc= (1 -aG)- 20.6
VG
VP CP
d) Heat transfer fHT =- 330-650
hA
Conversion Processes
CL
a) Oxygen consumption to, = - 0.7-16
rgy
SO
b) Substrate utilization t,, = - 5 . 5 . lo4
r y
1
c) Biomass growth tG =- 1.2.104
Pmax
P Cp ATcoolinp
d) Heat production fHP = 350
THM +THS
a Typical time values for gluconic acid fermentation in a production-scale
reactor
tor in order to analyze the quality of mixing models, and modern fluid dynamic models
and other transport phenomena such as mass based on the numerical solution of turbulent
transfer between the phases causing gradients flow equations. With the aid of two simple
in the concentrations of various substrates and examples it will be demonstrated how the in-
products. Because of the interaction between teractions of mass and energy distributions in-
the biotic and abiotic phases of the system via fluence the process outcome at different scales
flow of energy and material, the structures of of operation as well as suggest modifications
the cellular kinetics and the structures for the in the operation and design of bioreactors.
physical transport of momentum, mass, and Aerobic growth of baker's yeast and oxygen
heat should not be considered alone. Both of transfer in viscous broths will be used as exam-
these, individually and jointly, affect the final ples.
outcome of the process. Let us follow the discussion with some qual-
Despite this strong interdependence, this itative considerations of distributions in stirred
chapter will concentrate on structured model- bioreactors in the light of scale-up problems. It
ling of the abiotic phases. The discussion will is now well known (GUNKEL and WEBER,
focus on problems of the mechanically agi- 1975; NAGATA,1975; OKAMOTO et al., 1981)
tated bioreactor. An attempt is made to criti- that energy introduced into such a system with
cally discuss the state of the art in describing the help of an agitator is dissipated mainly in
the distribution of mass and energy, thereby the vicinity of the agitator, i.e., a non-uniform
confronting the classical approach based on distribution of energy takes place. For small
recirculation time distributions, compartment reactors, the region of intense energy distribu-
Introduction 303
mass and energy distributions may have seri- ing conditions. However, this does not allow a
ous implications for a number of biological quantitative evaluation of the spatial variation
systems. It may be added, however, that of property values in the vessel and, therefore,
whereas the discrepancies of mass distribution cannot be used to account for the effect of
affect the reaction rate directly by influencing non-homogeneous distributions of substrate
the concentration levels, the problems of en- concentrations upon the microbial metabolism
ergy distributions will also play a role in the fermentors. Moreover, this concept ap-
either through the gas-liquid/liquid-solid mass pears to suggest that the goal should be
transfer or through the diffusion in morpho- to achieve a uniformity of concentration
logical forms, or by affecting cellular meta- throughout the vessel, which may not be eco-
bolic capabilities. In any case, a thorough un- nomical as the scale of operation increases. If
derstanding of the problems of mixing and of a region of intense turbulence exists in the ves-
energy distribution and their interactions with sel, such as the impeller zone in agitated reac-
the kinetic processes is required. tors or the gas distribution zone in an air-lift
reactor, the contents of the vessel may be cir-
culated through it at a high enough frequency
so that the concentration of the target reac-
tant(s) is kept within bounds along the circula-
2 Characterization tion path. This approach involving circulation
time and its distribution manifests itself in re-
of Mass and Energy cycle models. However, the number of papers
dealing with such models for batch and semi-
Distributions in Stirred continuous operation is limited. Earlier impor-
tant contributions were made by HOLMESet
Tanks al. (1964), KHANG and LEVENSPIEL(1976),
MANN and CROSBY(1973), MANN et al.
(1974) and VONCKENet al. (1964). The experi-
The quantitative characterization of mixing mental methods used for the measurement of
has been a subject of active research in chemi- circulation time and distribution can be cate-
cal engineering, and a number of publications gorized either as stimulus-response techniques
and review papers have appeared in the past. or as flow-follower techniques.
For continuous flow reactors, the technique of
residence-time distributions is generally em-
ployed. Based upon the observed residence- 2.1 Stimulus-Response Methods
time distribution, different models of mixing
are proposed, and these are then used to evalu- Based upon Tanks-in-Series Models
ate the effect of mixing upon the performance
of the reactor (LEVENSPIEL,1972; NAUMANN, In these methods pulse injection of a tracer
1981; SMITH, 1981; WEN and FAN, 1975). is made near stirrer tips and the response is
Since most of the microbial reactors are oper- measured nearby. Thus, injection and meas-
ated in a batch or fed-batch (semicontinuous) urement are made in a well-defined region of
mode, the well-established and widely used the tank in a small, well-mixed volume. Using,
residence-time distribution techniques will not for instance, ionic tracers and conductivity
be discussed here. cells located in the form of a loop around the
Mixing in batch reactors is typically charac- impeller (Fig. 3), the experimental observa-
terized with the help of a terminal mixing time, tions may be interpreted with the aid of var-
defined as the time required to achieve a given ious recycle models. As a simple representa-
degree of homogeneity of concentrations. This tion of the recirculation flow, KHANG and
parameter has often been used to design con- LEVENSPIEL(1976) suggested a tank-in-series
trol loops for operating variables such as tem- model illustrated in Fig. 4. The impulse re-
perature, pH, etc., and a number of correla- sponse for such a recycle system takes the
tions exist relating mixing time to the operat- form:
Characterization of Mass and Energy Distributions in Stirred Tanks 305
\
Ring electrode
where B = 0; N odd
B = 1; N even
M = (N- 1)/2; N odd
K
I I
M = N/2 - 1; N even I I I
I I I
which can be reduced for large times and 2 d N
Q 1 to the approximation
'COS ($t + $)
This is an equation of a wave with decreasing
amplitude
A 2 2 exp (- NB
2x2
t)
I I
I
4 ----
w-a Q-aI
w u
MI
- 1
I- - - - - - - - - - - - - - - - - - - - - - -I
Average residence time 8
-
Vessel number N = 92
62
Variance of f i t ) Fig. 4. Recycle model for a batch stir-
red reactor.
306 10 Stirred Tank Models
np cycles approaches a log-normal distribution tank, (Fig. 5 ) the amplitude decay rate con-
as np-+03, or stant KA in Eq. ( 5 ) can be obtained by measur-
ing the peak and valley values A from the sig-
nal amplitudes and the times of their appear-
ance. Taking logarithms of both sides of
Eq. (9,one gets
for large np
0;
In A = In 2 - KAt, where KA = 2 n 2 - (6)
Since the variance of the tank-in-series model
is given by
e
Thus the slope of the 1nA vs. t plot gives the
a2 1 KA value. From measurements in two tanks
O;=BL=N with several different impellers, KHANG and
LEVENSPIEL(1976) suggested a dimensionless
Eq. (2) may be written for an arbitrary single- equation for turbine impellers:
pass distribution in the form
n d:
for Re = -> 2+103 (7)
'COS
("8"
-t+2nn;
1 (4)
V
2 i s 6
Time lime i s )
Fig. 5. An example of the impulse response in a Fig. 6 . Measured impulse response for the asymme-
reactor with symmetrical impeller position. trical impeller position shown in system 3 of Fig. 7 .
Characterization of Mass and Energy Distributions in Stirred Tanks 307
When trying to apply this simple concept to partly overcome by considering that the system
the characterization of the circulation flow in a now consists of two different flow regions.
stirred bioreactor, we are first confronted with The overall system response may thus be pre-
the problem that in such vessels the position of dicted from a superposition of the recircula-
the impeller@)is usually not symmetrical. For tion time distributions in the two different re-
an aerated stirred bioreactor, the impeller has gions similar to the way suggested by MANN
to be placed near the bottom. If the liquid and CROSBY(1973) in their theoretical consid-
height is still equal to the tank diameter, this erations.
may result in a geometrical configuration illus- Fig. 7 illustrates how the asymmetrical flow
trated in system 3 of Fig. 7. Applying the regions of system 3 may be replaced by two
symmetrical systems. The overall pulse re-
sponse can then be calculated from
System 3
7li'-
using Eq. (4) and ( 5 ) for the calculation of
tracer response in the two symmetrical subsys-
tems 1 and 2.
System 1 System 2 An example of the comparisons between
Fig. 7. Replacement of an asymmetrical impeller po- measured and predicted results is shown in
sition by superposition of two symmetrical sys- Fig. 8. Though the agreement between the
tems. measured and predicted results appears to be
satisfactory, it must be emphasized that a
number of uncertainties remain when applying
this strategy. First of all, it requires identifica-
method suggested by KHANG and LEVENSPIEL tion of four parameters from a single signal re-
(1976), BERKE (1980) measured tracer re- sponse curve. Secondly, an additional ex-
sponses in such configurations. A typical change of fluid between the individual recircu-
example for the results obtained is illustrated lation flows cannot be entirely excluded. Such
in Fig. 6. Obviously, the asymmetrical impeller an exchange, however, would require further
position drastically influences the response sig- parameters to be estimated from the measured
nal. Thus, it is no longer possible to treat this signal, which would finally make such a proce-
signal in the way suggested by KHANG and dure rather intractable.
LEVENSPIEL (1976). The difficulties may be
2.2 Flow-Follower Techniques
1.5
Some of the problems mentioned before
m
may be overcome by making use of flow-fol-
sn 1.0 lower methods. The advantages of these tech-
a2
niques have been well summarized by BRYANT
(1977), and his paper should be consulted for
-=
m
v)
---calculated further guidance. The analysis of the frequen-
-Ea5 cy pattern for passage of flow followers
through an active region - impeller zone in
the case of a stirred vessel - provides us with
0 the following useful information:
0 2 6 10 a) mean circulation time 8,
lime is]
b) standard deviation of circulation times &,
Fig. 8. Example of a comparison between measured and
and calculated impulse response. c) distribution of circulation times.
308 10 Stirred Tank Models
Sphere /
P - **,
0.20 1
I El0 3000
1 and
: - 1)
0’= 02(exp a
or
and
N 99.981 I I 8 1 1 1 1 1 1 I I I I I I l l 1
f2= t^ffi.At
i= 1
1 -
f ( t ) = ___
mt
1 exp (- (1nii:i)2) (14)
where I I ,I , I 1 I I , ,
1 2 3 1 5 678910 20 10 60 80100
lime (s)
O=exp p i + -
( 3 Fig. 11. Log-normal plot of circulation time data.
3 10 10 Stirred Tank Models
where q1 is the fraction of the flow that passes Thus, the characteristic flow for circulation of
through region 1 having a log-normal distribu- the fluid includes the entrainment flow. The
tionf, (MA" and CROSBY,1973). This more application of magneto-flow follower tech-
accurate procedure, however, would require niques with a coil located close to the impeller
the adjustment of five parameters, including blades results in the detection of the discharge
q l . For reasons of simplicity, the effect of geo- flow from the impeller. Under these conditions
metric asymmetry upon the distribution may it is more reasonable to estimate the corre-
be neglected. In this case, the measured data sponding mean circulation velocity from the
are approximated by a single log-normal distri- continuity equation, which takes the form
bution, which can be used in connection with
microbial kinetics.
From systematic measurements of distribu-
tions in two different vessels (100 and 3000 lit- with pumping capacity of the impeller Q,.
ers) at different ratios of liquid height/tank di- Here, it is assumed that the fluid flow through
ameter and impeller diameterhank diameter, the impeller region is equal to the correspond-
respectively, the mean circulation time could ing circulation through the cross-section of the
be correlated with the geometrical and opera- tank, characterized by the fluid velocity 8.
tional parameters in the following manner From Eq. (23) we predict
(BOELCKE,1983; REUSS and BRAMMER, 1985;
REUSS,1988):
tr= C;ndi (2)'
v, = 0.85 x n di (2)7/6
Fig. 12. Circulation paths for a single impeller
(asymmetrical position).
Characterization of Mass and Energy Distributions in Stirred Tanks 311
The circulation paths are predicted in a similar agreement in the exponent for (IUD,). The
manner as suggested by MCMANAMEY(1980). theoretical value of the exponent for (DT/di)
The lengths of the two circulation paths illus- is, however, higher than that observed in the
trated in Fig. 12 are given by experiments. This deviation is caused by the
uncertainties in the estimation of the charac-
L1 =DT+2H-2.5di (25) teristic circulation velocity.
and
[I + E - O . 5 5D]T0 :
e=
n d? 2.2.3 Influence of Aeration
As shown in Fig. 13, the term in brackets in Only a few published papers deal with the
Eq. (28) can be approximated by a simple mixing characteristics in aerated stirred tanks.
power function, resulting in the following cor- The reported results for the terminal mixing
relation: time under aerated conditions are somewhat
controversial. BLAKEBROUGHand SAMBA-
H- 0 . 5 -
[1+%
Ddi
T 1 = MURTHY (1966) as well as EINSELE and FINN
(1980) observed higher mixing times with in-
creasing aeration rates. In contrast, PACA et
al. (1976) predicted shorter mixing times. Jo-
SHI et al. (1982), measuring terminal mixing
times in various liquids, reported a slight in-
This correlation holds in a region 0 . 2 1 fluence of aeration rate:
di/DT50.5 and 1 s H / D T s 2 . The final equa-
tion for the mean circulation time 0 is, there-
fore, given by
BRYANTand SADEGHZADEH (1979) were the
first to thoroughly investigate the mechanisms
of the mixing process in aerated liquids with
the aid of the radio flow-follower technique.
A comparison between this equation and the This method provides the necessary informa-
empirical correlation from the experimental tion on how the mean circulation time and the
observations, Eq. (20), shows an excellent variance of the distribution are affected by
312 I 0 Stirred Tank Models
As can be seen from Fig. 14, the variance of tween the different fluid elements may range
the distribution a2 also increases with increas- from complete micromixing to complete segre-
ing air flow rates below the flooding point. gation. The residence time distribution in this
When calculating the terminal mixing time by zone, called a macromixer, is the circulation
making use of Eqs. (3a), (6), and (8) time distribution in the reactor. This results in
a two-environment model (Fig. 16) similar to
that proposed by MANNINGand coworkers
(1965). The different cases of complete, par-
tial, or zero segregation may be simulated by
using a tanks-in-series approach or a Monte-
Carlo simulation approach. For non-reactive
tracers and for first-order reactions, all de-
one observes that this value is only weakly in- grees of segregation give the same results.
fluenced by the aeration rate.
Increasing age of
volume elements
-
Vmacro
Contribution of
Circulation time elements of the
distribution Volume of elements circulating
stream to the
micromixer
OATEN
OATEN-1
aA TEN-?
These contributions are denoted by dark areas of each element in the macromixer as in the
in Fig. 17. As a result, volumes of the elements following equation:
decrease. The contribution of the Nth element
is exactly equal to its volume; in other words, N
this element leaves the macromixer. At this EiCiQAt
i= 1
moment, the ages of all the elements are ad- Cexit = (33)
vanced, i.e., the j t h element is termed the
(j+1)th element. The newly entered element of
2 EiQAt
i= 1
volume Q A t becomes the first element. The
composition of the exit stream is determined where E is the fraction of the circulation times
by the weighted average of the contributions between ( i - 1 ) A t and i A t . Ei is related to the
Characterization of Mass and Energy Distributions in Stirred Tanks 3 15
L I
The volume elements in the macromixer are as-
sumed to be completely segregated, and the Fig. 19. Material balance around the micromixer in
progress in each of these with age can be calcu- the case of impulse tracer input.
lated with the help of known initial conditions
from the time each entered the macromixer.
Since each element has a unique age, reactions time unless the variance is insensitive to the
in each proceed also to a unique extent, and measurements. The results of MIDDLETON
each thus makes a different contribution to the (1979) show this not to be the case.
recycle stream. The circulation time distribu- The different degrees of segregation in the
tion and the volumes of the different elements macromixer must also be considered, if reac-
in it are shown in Fig. 18. For further details tions are also simultaneously taking place.
the original paper by BAJPAI and REUSS Specific examples of such cases will be present-
(1982a) is recommended. For intermediate de- ed later in this chapter.
grees of segregation, methods based upon
coalescence and redispersion can be used
(CURL,1963; BAJPAIet al., 1977).
The volume of the micromixer is assumed to
be negligible compared to that of the macro-
mixer. Hence, the micromixer acts as an ideal
mixer of the recirculating stream with any in-
coming stream. This situation is shown in Fig.
19 for tracer inputs into the reactor. A materi-
al balance around the micromixer results in the
following expression for the concentration of
tracer in the new element:
.-s
L I
(35) I
c
0 ,'
g o 1 2 3 I 5
c
u
0 Oimcnrionlcts time, t i 0
Typical simulations of tracer responses in the
impeller region for log-normal circulation time Fig. 20. Computer prediction of the dynamics of
distributions (corresponding to symmetrically dispersion of inert tracer throughout reactor vol-
placed impellers) using the discrete simulation ume.
procedure are shown in Fig. 20. The predicted
oscillating concentrations of tracer in the im-
peller region are similar to those experimental- 2.4 Circulation-Time Distributions
ly observed by others (KHANGand LEVEN- for Two-Impeller Systems
SPIEL, 1976; MIDDLETON, 1979). Similar re-
sults were predicted by BRYANT(1977) using a
convolution procedure. 2.4.1 Measurements Using
Mixing time, defined as the time required Flow-Follower Techniques
for the oscillations to die down to a suitably
low level, can be seen from these figures to de- The circulation flow in a multiple impeller
pend not only upon the mean circulation time, system may be envisaged as a superposition of
8, but also upon the variance, u2.This contra- the circulation flow of the individual impellers
dicts the popular suggestion that the mixing and an exchange flow between the impeller
time is a fixed multiple of the mean circulation zones. The problem of the quantitative analy-
316 10 Stirred Tank Models
Entire distribution
I ).
Macromixer {i G
0
Mjcro-
mixer
0
I ii,and iiz
Macromixer
I 0
Mlcro-
mixer
0
Exchange distribution
I 43
Fig. 21. Two-environment model for mixing in a
stirred bioreactor with two impellers.
r
n
H
Exchange
volume V,,
Fig. 24. Impeller configuration and exchange flow in the two-impeller system.
L, 17
-=-- - 0.7
L2 24
t
0.5 H: Dl
a value slightly higher than the measured value
of 0.55.
In order to estimate the exchange coeffi-
cients P from the measured circulation time
distributions, it is assumed that only those
fluid elements contribute to the exchange flow
which belong to the exchange volume illus-
trated in Fig. 24b. For the geometrical config-
uration shown in Fig. 22a, this volume is given 0.5 H=Dl
by
6
V,,=-K i=l,2 (42)
7
The exchange coefficients are then calculated
from
factor of 1.3, predicted for the effects of the scale of operation. However, sometimes it may
difference in the lower and upper circulation not be possible to employ flow-follower tech-
(Tab. 3), results in an overall ratio of niques to establish the parameters of circula-
tion, particularly in large-scale industrial proc-
62 esses involving multiple impellers. Specifically,
-= 1.6 the requirements of sterility and presence of
81
filaments in broth may rule out the introduc-
This value is in excellent agreement with the tion of non-sterilizable flow followers and the
experimentally observed differences between aerials or coils in the production reactor. In
the mean circulation times in the upper and such cases, tracer measurement methods can
lower impeller regions presented in Tab. 3. still be used as exemplified by the work of
From the simple calculation presented above, JANSEN et al. (1978), who measured mixing
it may be concluded that reasonable estimates characteristics with working volumes up to
for the mean circulation times can be predicted 120 m3. During operation of the bioreactor,
for different impeller configurations by calcu- calculated quantities of radioactive tracer
lating the mean lengths of the individual circu- (technetium isotope) were introduced at the
lation paths. This feature should lead to rea- top, and mixing was followed by measuring
sonable estimates for the mean circulation the tracer concentration at the bottom. Under
times. these circumstances, a question arises as to
whether the circulation and exchange parame-
ters can be reliably estimated from such data.
2.4.2 Measurements Using For a two-impeller system, simulations of
tracer responses were conducted using the pre-
Tracer Responses viously described micro-macro mixer model.
For tracer injection in the upper impeller re-
The concept of using circulation time and its gion, typical tracer concentration profiles in
distribution to analyze mixing phenomena ap- the two turbulent zones are shown in Fig. 26.
pears to be very sound and helpful in explain- Using discretized tracer concentrations pre-
ing many observed phenomena related to the dicted at the lower impeller, the parameters 19,
320 10 Stirred Tank Models
-n
.-
Fig. 27. Results of parameter esti-
mation from a simulated signal
0 5 10 15 20 2; response in the lower impeller re-
lime, f(s] gion.
o
', and p were then back-calculated using a
modified simplex optimization algorithm
(Nelder-Mead method). The results of such an
exercise are shown in Fig. 27 where the (so-
called) experimental data are plotted as dis-
crete points, and the continuous curve belongs
to the predicted profile with optimized param-
eters. That the crucial parameters 6' and p are
estimated to within 10% of the actual values
shows that the parameters of the two-compart- Fig. 28. Two-region mixing model (SINCLAIRand
ment model are sensitive enough to allow an BROWN,1970).
Characterization of Mass and Energy Distributions in Stirred Tanks 321
4
U I I
0
I I I
f
10
Dilution r a t e , D l h-’I
Fig. 29. Exit cell concentration and productivity as a
function of the dilution rate at different interchange
rates for the two-region model of SINCLAIRand
BROWN(1970). (9
Fig. 30. Five-compartment model (OOSTERHUIS
and
The two-compartment model of SINCLAIR KOSSEN,1984).
and BROWN(1970)
the equation for coalescing systems (VAN’T The multi-turbine model of BADER(1987a, b)
RIET, 1979):
The model, schematically shown in Fig. 32,
Pc separates the reactor into a series of mixing
kLa=0.0323 (E) 0.4
@
I------
Liquid
m
f Gat
0.5
Y;‘ “2
0.200 12.71
0.195 19.77
0.190 26.61
0.185 33.90
0.160 10.96
Fig. 31. Comparison between computed and meas-
ured dissolved oxygen profiles (dissolved oxygen
and KOSSEN,
tension in 070 saturation) (OOSTERHUIS Fig. 33. Model prediction of dissolved oxygen pro-
1 984). files at different uptake rates (BADER,1987b).
Characterization of Mass and Energy Distributions in Stirred Tanks 323
Upper loop 9
Bottom loop 9
Fig. 34. Extension of the model of KHANG and LEV-
ENSPIEL (1976) for the case of two agitators (BAJ-
PAI and SOHN,1987).
C
100 1
-----
------ ----------
Impeller 1
Fig. 37. Dissolved oxygen profiles in a 40 m3 fermentor (1.5 kW/m3 power input) (SINGHet al.,
1987).
-(l-O)fCL,,,+,+ffCL,,,_,
2
+
liquid c h x l a t i o n
mass transfer
(54)
rea;tion
-
0% 3’10
Dissolved oxygen
tant first steps regarding numerical computa-
tions of multi-dimensional multi-phase flow
which show the direction for future research in
this field. In the following, only an introduc-
Fig. 42. Computed distribution of oxygen profiles in tion to this topic is attempted. The interested
a 30 m3 tank. Oxygen consumption is predicted with reader must be referred to the extensive special
the aid of Monod kinetics. literature.
328 10 Stirred Tank Models
48 with
CD = -
R ej
PL
Rej = 2 R Uj-
PL
k2
Cct=CpPL-
&
The quantities k and E are taken to obey the The values for the different constants are sum-
following transport equations marized in Tab. 5 .
Numerical solutions of the system of cou-
pled balance equations have been presented
C, CL c, 6k 6,
1.44 1.92 0.09 1.oo 1.30
330 10 Stirred Tank Models
for one- and two-phase flow. HARVEYand agitated and aerated vessels is only now begin-
GREAVES(1982a, b) predicted turbulent sin- ning. The task is not an easy one, and it seems
gle-phase flow in an agitated vessel by applica- too early to speculate on coupling the turbu-
tion of the k--E model. The computations, lent transport equations with the microbial
however, rely on a rather unrealistic boundary reaction model which must be structured in or-
condition for the turbulence parameters in the der to account for the behavior in a contin-
impeller region. The adopted approach as- uously changing environment. Nevertheless,
sumes that the gradients of mean velocity com- results from models for turbulent two-phase
ponents determine the level of turbulence in flow, as far as the distribution of kinetic ener-
the vicinity of the impeller. This inadequate gy and dissipation in the tank is concerned,
oversimplification of the flow situation in the could be used immediately to improve confi-
impeller region has been removed in the work dence in the multi-phase compartment models
of PLACEK et al. (1986). These authors extend described in the previous section. This strategy
the k--E model through an incorporation of would result in a simpler model structure for
different turbulent energy scales. Thus, a mul- the abiotic phases with distributed parameters
tiple-scale model is created in which the energy predicted from the two-phase turbulent model.
spectrum of turbulence in the vessel is divided This kind of model reduction is presently a
into the large-scale vortices produced by the prerequisite to couple more complex metabolic
impeller, the intermediate or transfer eddies, models to the dynamic response of microor-
and the dissipation eddies. Furthermore, the ganisms to changing environmental condi-
boundary conditions in the impeller region rely tions.
on a discharge flow model suggested by PLA-
CEK and TAVLARIDES (1985), which includes
the trailing vortices at the impeller blades ob-
served by VAN’T RIET and SMITH(1975).
ISSA and GOSMAN(1981) tried to predict 3 Applications
the three-dimensional turbulent two-phase
flow in an agitated and aerated vessel. Again,
to Microbial Systems
the most critical point seems to be the bound-
ary condition in the impeller region. The em- Through the application of regime analysis
pirically described flow in this region makes (SWEEREet al., 1987 and Tab. 2), and also
the predicted results uncertain. With the aid of from the concept of relaxation times (ROELS,
a superposition of oxygen transfer from the 1983), it becomes clear that all the processes
gas bubbles into the liquid as well as a Monod- with time constants for reactions comparable
like expression for oxygen consumption, or smaller than those for mixing have the po-
TRAGHARD(1988) presented the numerical so- tential for interactions between mass and ener-
lution of the equations of motion together gy distribution and kinetics. Aerobic proc-
with the k - E turbulence model for the two- esses, in which oxygen must be continuously
phase flow. As mentioned by the author, the delivered to the liquid phase via aeration (in-
treatment of the gas bubbles as well as the troduction of mass) and agitation (energy) to
flow conditions in the impeller region must be meet large oxygen demands of microbial cells,
considered as weaknesses in this model. The are one category of such systems. A second
most difficult problem and also the most im- class of potentially interesting microbial proc-
portant task for the future is the comprehen- esses are the ones in which manipulations of
sive mathematical treatment of the disruption the extracellular environment are used to di-
and coalescence of the gas bubbles, including rect metabolism in a specific direction (GRIOT
the existence of a size distribution. The other et al., 1986). In such cases, the demands of the
aspect, hitherto not thoroughly investigated, is critical nutrients change during the process.
the analysis of the relative motion of the gas Their supply, therefore, must also be changed
bubbles in the turbulent flow. appropriately. At smaller scales of operation
It may be concluded that the investigations in which mixing is normally not a problem (for
of the two-phase modelling of turbulence in exceptions see HANSFORDand HUMPHREY,
Applications to Microbial Systems 331
1966; KNOEPFEL,1972; EINSELEet al., 1978), high cell densities in which oxygen consump-
purely kinetic considerations govern the proc- tion rates will be high. For fermentations using
ess. As the scale of operation increases, mixing molds existing in pellet or filamentous form,
patterns must also be considered. This whole diffusional limitations are known to exist even
spectrum of problems has been an area of vig- at lower cell densities.
orous research and developmental activity in Diffusion of oxygen in pellets has been ex-
biochemical engineering. tensively investigated (YANO et al., 1961; Yo-
Applications of models in some simple situ- SHIDA, 1976; AIBAet al., 1971; KOBAYASHI et
ations will be discussed in the next sections to al., 1973; ATKINSON,1974; METZ, 1976; MIU-
demonstrate quantitative investigations of sev- RA, 1976; REUSS,1976; VAN SUIJDAM, 1980).
eral important operational and scale effects. Various asymptotic and numerical solutions to
the governing equation
3.1 Oxygen Transfer to Molds
Based upon a Sherwood number of 2 for
diffusion from bulk to a spherical particle hav- involving molecular diffusion of oxygen and
ing no relative motion, CALDERBANK (1967) its consumption under pseudo-steady-state
has shown that for unicellular microorganisms conditions have been presented in the litera-
as single cells (i.e., not as flocs or films), no ture. This analysis was extended by REUSS
external diffusional limitations for the uptake (1976) to include the various external transport
of nutrients should exist. When, however, they resistances in which the boundary condition is
do form flocs (clusters of single cells) or films, written as
diffusion of nutrients in these may become
limiting under suitable circumstances (low
bulk concentrations, large uptake rates, large
particle diameter, etc.). ATKINSON(1974) has
covered these phenomena in detail. For con- p, here, is the overall mass transfer coeffi-
sideration of these effects, the text of BAILEY cient
and OLLIS(1986) can be also used.
0.8
0.6
-
002
QD”llX
Ok
0.2
Fig. 44. Effect of exter-
nal and internal mass
0- I
transfer limitations on
lo-‘ ioo 10’ 102 103 effective oxygen con-
external mass transfer sumption of spherical
Bi = pellets.
~~f~ internal mass transfer
For filamentous morphology, few quantita- where q is the effectiveness factor. The effec-
tive data are available dealing with transport tiveness factor is related to diffusion and reac-
resistances. Also the physical nature of the sus- tion parameters as follows:
pension is somewhat unclear (METZ et al.,
1979). Several experimental observations,
however, show changes in the critical dissolved
oxygen concentration for growth as well as for
product formation during the course of the and
’(
fermentation (STEELand MAXON,1962, 1966;
WANGand FEWKES,1977; Fox, 1978). Parti-
cularly, STEELand MAXON(1966) as well as v=--- -- 1) for y r 1 (80)
WANGand FEWKES(1977) found a strong in-
fluence of the impeller to tank-diameter ratio. where 0 is known as the Thiele modulus, de-
The latter authors postulated that the mass fined as
transfer of oxygen from bulk liquid to the my-
celial surface was the controlling factor.
This problem has been quantitatively ana-
lyzed by REUSSet al. (1982) with the experi-
mental data for oxygen uptake in Aspergillus and w as a “general modulus” related to the
niger broths. The mycelial mass has been con- Thiele modulus and other indices as
sidered to consist of hypothetical spheres of di-
ameter d,,, in which biomass is uniformly dis-
tributed. Assuming that the diffusion of oxy-
gen in these hypothetical spheres is the critical
transport phenomenon, the known solutions
of Eq. (75) could be used. ATKINSON(1974)
has provided a pseudo-analytical solution of
this problem as
a is a geometrical parameter whose value for
spherical particles was suggested by ATKINSON
(1974) as 1.16.
Applications to Microbial Systems 333
In analogy to turbulent eddies, the sizes of 1966; WANGand FEWKES,1977; RIZZI, 1982)
these diffusion elements are governed by a bal- suggest that for the same energy input smaller
ance of shearing forces and the forces neces- impellers are more effective than larger ones in
sary to break mycelial filaments. Through supplying oxygen to filamentous organisms.
comparisons between suitably designed oxygen
uptake measurements and the Eqs. (78)
through (82), the sizes of the elements were 3.1.2 Mass and Energy Distribution
found to pertain to the inertial subrange of
turbulent eddies. In the absence of coalescence STEEL and MAXON (1966), studying the
of the spherical elements, their sizes are gov- scale-up of the novobiocin fermentation, ob-
erned by the maximum shearing forces en- served that the pronounced influence of the
countered in the vessel, and the following ex-
pression for d, can be obtained:
~,,,=0.5E (2) 3
‘1
11
0.5
I
1 2
E=
I
(Wlkgl
5
i
10
obtained from measurements of local energy
dissipation in stirred reactors having turbine
impellers (LIEPEet al., 1971). Accordingly, Fig. 45. Thiele modulus as a function of the energy
dissipation rate for filamentous morphology (data
from batch fermentation with Aspergillus niger).
1.o
The constant K is related to the strength of
mycelial filaments.
BERKE (1980) and RIZZI (1982) measured 0.8
the oxygen uptake kinetics at different power
input levels and impeller to tank diameter ra- 0.6
tios. Thiele modulus values calculated from
these experiments are plotted in Fig. 45, show- 0.1
ing the validity of Eq. (85). The oxygen uptake
data obtained with different values of (di/DT) 0.2
are shown in Fig. 46 as a function of power
input. Also plotted are the results of simula-
0
tions using Eqs. (78) through (85) with 0 2 6 e 10
K=8.8. The close agreement confirms F (Wlkg)
the proposed influence of geometric parame- Fig. 46. Influence of mean energy dissipation rate
ters upon uptake kinetics through their effect and ratio of impeller-tank diameters on the effective
upon local energy dissipation in the impeller oxygen consumption rate of Aspergillus niger grow-
region. A large number of experimental data ing with filamentous morphology (symbols: meas-
(BERKE, 1980; STEEL and MAXON, 1962, ured data; solid lines: predictions).
334 10 Stirred Tank Models
\
20 1 Fermentor
Here is the effectiveness factor, account-
ing for diffusional limitations in diffusion ele-
ments of size dp. q and dp are related to oper- -
ating and environmental conditions by Eqs. 15 000 1 Fermentor
(78) through (83) and Eq. (85). The constant
can be determined by measuring the oxygen
uptake kinetics (Qo, vs. CLin a small ferment- 0.2 ar 0.6 aa 0.2 OA 0.6 aa
er) through changes in power input and using di /Dl di / D l
an optimization procedure to minimize differ- Fig. 50. Comparison of experimental observations
ences in Qo, observed and Qo, calculated with (STEEL and MAXON, 1966) and model predictions
Eq. (78). for oxygen supply in a fermentation broth of Strep-
For a given circulation time distribution, the tomyces niveus.
concentration of oxygen in the exit stream
from the macromixer is given by
m N
diameter ratio (di/DT)at fixed power input ( E )
CL= 1 CL(t)f(t)dt= 2 CL(ti)Ei (87) at any scale. These simulation results show a
0 i= 1
behavior qualitatively similar to that observed
Loading of elements in the micromixer can be by STEELand MAXON(1966), Fig. 50. In these
described by taking the oxygen balance around simulations, 8 values were calculated from a
it (Fig. 49) as correlation given by MIDDLETON(1979) and
kLa values using a correlation by REUSS
VkL a(Cf - Ct) = Q(CO, - EL) = P G @" -y ) (88) (1983). No attempt was made to obtain a
quantitative fit due to the diversity of sources
Here Q is the circulation rate ( = V/8 ), and Cf of parameter values. The qualitative agree-
is the solubility of dissolved oxygen. This may ment suggests that the pronounced effects of
be related to y and y" by using an appropriate the scale of operation may be explained by a
assumption concerning gas residence time-dis- coupled effect of circulation and diffusion. Al-
tribution in the reactor. If the gas phase is con- though the smaller impeller in the large tank
sidered well-mixed, then still produces smaller diffusion elements due to
a higher maximum energy dissipation rate,
these elements stay away from the impeller
Cf 2 , I . (89) (loading zone) for a longer time due to reduced
H 'mpe"er
pumping capacity. As a result, diffusion and
Eqs. (86) through (89) can be iteratively solved circulation effects are antagonistic in nature
to predict the influence of the impeller to tank and tend to annul each other as the (di/DT)ra-
tio is changed. At the same time, the speed of
agitation in the larger vessel is low compared
ri,. Y from macromixer to that in a smaller vessel (constant B), result-
ing in lower kLa values. This causes a net re-
duction in oxygen transfer to cells in the large
vessels.
For low-viscosity fermentation broths, the
assumption of no mass transfer in the mac-
l o macromixer romixer may not be correct. In this case, Eq.
A Q ,C! (86) will be modified to include an oxygen
V supply term involving a gas-liquid mass trans-
fer coefficient in the macromixer. As sug-
%,Y"
gested by OOSTERHUIS (1984) and OOSTER-
Fig. 49. Oxygen balance around the micromixer. HUIS and KOSSEN (1984), the flow phenomena
336 10 Stirred Tank Models
in the regions away from the impeller resemble pellers to be independent of each other and to
those in bubble columns, and one could use have similar gas-liquid mass transfer coeffi-
gas-liquid mass transfer correlations for bub- cients.
ble columns. A similar proposal was made by Considering a simple form of exchange, a
ANDREW(1982), SINGH et al. (1986, 1987, micro-macromixing framework for two impel-
1988) and RAGOT and REIJSS (1990). lers of Fig. 5 1 may be used. For the sake of
simplicity, the two impellers may be consid-
ered identical and to possess the same recircu-
3.1.3 Multiple-Impeller Systems lation flow as well as circulation time distribu-
tions. These assumptions may be somewhat
Even though multiple impellers are com- drastic in the light of data presented earlier
monly used in laboratory and industrial bio- (see Sect. 2.4). As a result of the equal circula-
reactors, precious little work dealing with tion rate, however, the exchange coefficient,
them has been published, at least in the open p, will also be the same for both impellers. If
literature. With regard to mixing and mass we used the superscripts "upper" and "lower"
transfer, it is clear that when the circulation to distinguish between concentrations at the
loops become too large, one should introduce two impellers (after exchange) these will be
more impellers on the shaft. Some quantitative given by
information as to when exactly it should be
done and how can be obtained through the ap-
plication of the concept here proposed for
mass and energy distribution. The information
concerning circulation time distributions in
each impeller zone, exchanges between the im- where CL denotes the exit stream concentra-
pellers, and power input and oxygen transfer tions as per Eq. (87). The balances at the mic-
at each impeller will be of crucial importance. romixer can be taken in the manner shown by
Because of little information available con- Fig. 5 1 as
cerning the last two, we may assume the im-
In this case,
*
lower
lo w e r -
Yout -Yin
upper- c~
- f I p
plower
and
T (93)
Fig. 51. Oxygen balances around micromixers for a Here Q is the circulation rate given by
two-impeller system. 1,742e).
Applications to Microbial Systems 337
Eq. (91) accounts for possible hydrostatic As a result, a net reduction in oxygen trans-
pressure effects that may be encountered in in- fer is predicted in spite of some reduction in
dustrial-scale reactors. the mean circulation time. RIZZI (1982) con-
Eqs. (90) through (93) were solved with Eqs. ducted several experiments in which Aspergil-
(78) through (82), and the calculated oxygen lus niger broths from a 300 liter fermenter
uptake rates in filamentous broths (complete were transferred into two identical 80 liter ves-
segregation in the macromixer) are plotted sels having one and two impellers, respective-
against power input levels using a single im- ly. A typical result comparing the two systems
peller in the same configuration in Fig. 52. is shown in Fig. 53, which confirms the same
trend as the theoretical predictions.
1.01
o.e -
0.6 -
3.1.4 The Role of Micromixing possible within any two maximally-mixed stir-
red tanks, whereas in “maximum mixedness”
While dealing with problems involving mix- molecules having the same life expectancy
ing and microbial reaction kinetics, mixing at from any part of the entire system should be
the cellular and even molecular level in certain able to mix infinitely fast. However, due to the
cases must also be considered. In the previous recycle nature of systems considered here, it is
sections it was assumed that fluid and diffu- most convenient to use the tanks-in-series con-
sion elements in their recirculation paths (in figuration to represent the extreme of mixed-
the macromixer) are in a state of complete seg- ness. All the tanks were considered to be maxi-
regation. BAJPAI and REUSS (1982b) have in- mally mixed and had equal volume. For the ith
vestigated the extreme cases of segregation in tank in the series, the governing equation for
the macromixer for the kinetics of uptake of oxygen uptake is
oxyen by filamentous microorganisms. The
reaction kinetics for this system are described
by Eqs. (78) through (82). The sizes of the dif-
fusion elements depend upon the power input,
E , and the ratio of impeller diameter to tank
diameter, di/DT, as in Eq. (85). The schema-
tics of simulations are presented in Fig. 54. CLois the concentration of dissolved oxygen in
Maximum mixedness was simulated using a the micromixer (Ct in Eq. (88)), and ,C, is
stirred-tanks-in-series configuration. To be rig- the concentration in the exit stream of the
orous, this representation corresponds to the macromixer (C, in Eq. (87)). For micromixer
case of “sequential mixedness” (WEN and balance, Eq. (88) for a single-impeller system
FAN, 1975), because molecular diffusion is not holds. For a pseudo-steady state of dissolved
1.0-
o.e -
Y/ segregation
0.6 -
-QQ,
QY
06 -
0 10 20 30
Cay ( % saturalion)
Fig. 54. Two-environment model for mixing with Fig. 55. Predicted oxygen uptake kinetics for differ-
schematic representation of the two extremes of ent reactor circulation time distributions corre-
micromixing. sponding to a mean circulation time (0) of 10 s.
Applications to Microbial Systems 339
Fh----
romixing, it is worthwhile to apply the estab-
’.or /-----
&/-*
For zero-order kinetics, such an analysis leads
to Fig. 56, where the results for a PFR and a
completely segregated CSTR are presented.
For a completely mixed CSTR, the solution is
trivial, i.e., the maximum rate holds at all
0.8 - CaV>O.Comparison of Figs. 55 and 56 sug-
gests that kinetic behavior in the circulation is
dominated by a zero-order reaction. This is
I ,/’
understandable due to the low KM value for
the oxygen uptake ( = 1.28 pmol/L), although
it is corrupted by the first-order diffusive proc-
ess. For maximum mixedness, both zero- and
first-order kinetics contribute.
The results presented in Fig. 55 reveal that
reactors with circulation time distributions re-
sembling those of a PFR are better for natural-
ly segregated systems such as highly viscous
0
; 13.9 mmol /ILhl non-Newtonian fermentation broths. Also, the
uncertainties of mixedness or the extent of seg-
0’ I I I I regation are far less important in such reac-
20 10 60 80 tors. This points to the desirability of achiev-
KM C, (%saturation) ing narrow distributions of draft tubes or of
Fig. 56. Predicted kinetics for a zero-order reaction using loop reactors. Such designs (HINES,
in a continuous stirred tank (----) and plug flow 1978; BLENKE,1979, 1985; FAUSTand SITTIG,
configurations (-) for two different mean resi- 1980) offer advantages for minimizing dead
dence times. spaces, increasing reliability of scale-up, and
340 10 Stirred Tank Models
also providing operational benefits for the vis- in conjunction with the micro-macro-mixer
cous non-Newtonian broths. model as described in Sect. 2.3 (BAJPAI and
REUSS, 1982a). For studying the effect of
varying glucose concentration during circula-
3.2 Substrate Distribution tion, it is only necessary to consider the micro-
in Baker’s Yeast Fermentation bial reaction in the vicinity of the critical sugar
concentration. Thus, a rather simple unstruc-
The second application of the concept of tured empirical model was chosen for the pur-
coupling of mixing and microbial reactions pose of this study. It involves Monod kinetics
presented in this chapter is concerned with for growth
problems of uniform distribution of the energy
and carbon source in large-scale operations. 1 dX S
Thus, referring to Tab. 2, we are now consid- p=--= Pmax - (97)
ering a class of processes in which the time X dt Ks+S
constants of substrate consumption (lsc=
S,,/rY) for zero-order reactions are of the and also a production rate of ethanol above
same order of magnitude as the circulation the critical sugar concentration Scrit
times in the bioreactor. The most impressive
and interesting example for this problem is the
continuous SCP production on methanol as for S 5 Scrit
carbon and energy source in the ICI loop reac-
tor at a scale of 2100m3 (SENIORand WIND-
NASS,1980; SCOTT,1983). Here the backmix-
ing of the substrate originally resulted in a de-
creased biomass yield due to high local metha- Substrate consumption in the individual ele-
nol concentrations. This problem was finally ments for the macromixer is given by
solved by multiple introduction of the sub-
strate. Many other industrially important
processes operated in a fed batch mode may be
considered because of their low substrate con-
centrations in the reactor. Obviously, baker’s
yeast production belongs to this category of The model parameters were estimated from
processes. In order to prevent ethanol produc- VON MEYENBURG’S data for continuous cul-
tion and, thus, yield reduction beyond a crit- ture of Saccharomyces cerevisiae (VON
ical sugar concentration under aerobic condi- MEYENBURG, 1969) and presented in the origi-
tions (Crabtree effect), sugar must be fed to nal paper (BAJPAIand REUSS, 1982a). By inte-
the batch processes. One of the scale-up prob- grating the balance equations for a short time
lems is then to get a uniform distribution of period A t in each of the volume elements pres-
the concentrated sugar feed, because the re- ent in the macromixer, the concentrations of
laxation time of the Crabtree effect is known the individual contributions to the recycle
to be extremely short (EINSELEet al., 1978). stream to the micromixer can be calculated.
These concentrations are then averaged in the
micromixer according to
3.2.1 Simulations of Fed-Batch N
Operations Based upon a Fixed Sa,= 2 SjEj
j=1
Schedule for Sugar Feeding N
X,, = 2 XjEj
The effects of mixing dynamics upon the j= 1
yield and productivity of this process have N
been theoretically investigated by using the cir- Pa,= 2 PjEj
culation model for flow of the biosuspension j= 1
Applications to Microbial Systems 341
in which Ej is again the fraction of the circula- its for the system. Due to the very high sugar
tion times between (j-l)At and j A t (Eq. concentration in the feed (So= 350 kg/m3), the
(34)). The calculation for the next time interval effect of circulation is also significant and re-
is then repeated after affecting a balance sults in a decrease in the critical growth rate at
across the micromixer. Here it is assumed that which ethanol appears in the system. These
the concentrated sugar solution is fed to the predictions are in qualitative agreement with
impeller region. The calculated composition of those observed in industrial practice, where the
the newly formed elements, which are the ini- exponential feeding program for sugar at the
tial conditions for the first element in the mac- industrial scale usually differs from the opti-
romixer for the next time interval, are then mal conditions predicted in the laboratory.
given by:
F=Fo[1 - ~ ~ ( Q c o , - ~ Q ” Q o , ) l (102)
1
0 2 1 6 8
Time I h l
-
where the specific growth rate p and the bio- Fig. 59. Effect of the feeding point on growth of
mass concentration X can be estimated on-line Saccharomyces cerevisiae in computer-controlled
by making use of the elemental balances and fed-batch cultivations (10 L scale).
the continuously measured oxygen uptake and
COz evolution rates. Fig. 58 summarizes the - - for simulation
experimentally observed growth curves by ap- Qco1 ‘0, o f control
plying this strategy to three different scales of
operation (10, 1000, and 2000 liter working
t
volumes). It must be emphasized that the same
set point (RQ= 1.05) was chosen for the ex- Qcol Qco, Qco,
4 4 4
periments at the different scales. As expected
from the model simulations presented in the
previous section for the open-loop feed-for-
ward strategy (Fig. 57), the productivity at the
larger scale is reduced because of larger circu-
lation times. The observed increase in growth
rate after making a shift in the speed of agita-
tion in the 1000 liter fermentor (Fig. 5 8 ) is a
further proof of this hypothesis. Since oxygen Recirculation
was always present in excess, increased oxygen Fig. 60. Simplified model structure for recirculation
transfer at higher speeds of agitation can be flow in an agitated batch reactor.
Applications to Microbial Systems 343
The idea behind this model reduction is the with the recirculation stream
long-term objective of these investigations.
One of the problems of interest will be the sim-
ulation of the system dynamics for designing
control strategies for large-scale operations.
This application requires the non-stationary p and v represent the kinetics for growth, and
solution of the model, including the gas-liquid ethanol formation rates are then calculated
mass transfer. Obviously, the basic structure from
of the model should be simplified for this com-
plex task. VIP1 + V2P2 + V3P3 (107a)
The simplified model is shown in Fig. 60, Pto t a l =
Vtotal
where the entire vessel is divided into three
CSTRs in series. The first CSTR (V1), which and
accounts for a small fraction of the volume, is
used to model the conditions in the vicinity of V]Vl+ v2v2+ v3v3
the injection ring of the concentrated sugar Vtotal = I, (107b)
Vtotal
feed. It is assumed that the volume of this ves-
sel is equal to the product of the cross-section The entire oxygen consumption rate is then
of the reaction times and the width of the im- predicted with
peller blades. A similar approach has been sug-
gested by MARINIet al. (1984) for studying the
effects of mixing on the outcome of a poly-
merization reaction. The two other vessels ac-
count for the conditions in the largest part of and the C 0 2 evolution rate with
the reaction volume. The number of two has
been estimated from the measured recircula- X
tion time distributions (Sect. 2.2. l), according QCO, = RQsp Qo2+ Vtotal - (109)
YCO,/P
to
where Yx/o,and Yc0,,, are the yield coeffi-
cients in kg biomass per mol O2 and mol COZ
per kg ethanol, respectively.
Assuming a constant biomass concentration For simulation of the entire growth curve,
for a short time period of the process, the sub- the system is considered in a pseudo-steady
strate balance equations in the three vessels state for A t = 5 min process time. Thus, the
can be written in the following form
-- F X X
s1-p1 -- v1-
Vtotal yx/s YP/S
dt
o! I I I I I I , I I
dS3
-- --
Q+F F 0 2 I 6 8 10
(SZ-S3) -- s 3 - Time (hl
dt V3 Vtotal
Fig. 61. Simulations of computer-controlled fed-
X X
-p3 -
yx/s
- v3 -
YP/S
(105) batch cultivations at different mean circulation
times 8.
344 10 Stirred Tank Models
system of Eq. (105) is iteratively solved with in the batch and fed-batch mode, however,
d S / d t =O. The three substrate concentrations defy such a treatment due to a lack of precise
S1, S,, and S3 are then used to predict the aver- quantitative description of their hydrodynam-
age growth and ethanol production rate, Eqs. ics. The capability of characterizing fluid
(107a), (107b), oxygen consumption and movement in stirred bioreactors with the help
carbon dioxide evolution rates, Eqs. (108), of circulation time distributions permits such a
(log), and finally the respiration coefficient coupling. These distributions can be measured
R Q= Qco,/Qo,. With the aid of a search al- by using flow-followers or by injection of non-
gorithm, the feeding rate Fj(ti) is calculated reaction tracers. The latter method appears to
afterwards, which minimizes the deviation be particularly suitable for viscous non-New-
between actual R Q and the set point tonian fermentation broths. With existing cor-
(RQsp= 1.05). Finally from relations for single-impeller systems, this
measurement method can be utilized for pre-
dX XF, cise estimation of mass exchange between im-
- - - Ptotal -
dt
~
Vtotal,
pellers in a multi-impeller system. In viscous
fermentation broths, distribution of energy in-
and troduced into the fluid also plays an important
role.
In order to understand the influence of the
geometry and scale of operation, the distribu-
biomass concentrations and reaction volumes tions of mass and energy should be coupled
for the next 5 min time interval are calculated. with kinetic processes in bioreactors.
The results of the simulations using the circu-
lation times calculated from Eq. (20), and tak-
ing into account the effect of aeration (Fig.
15), are presented in Fig. 61. As observed in
the experiments, the increase in circulation 5 References
time results in a corresponding reduction of
growth rate. Due to the reduced recirculation
AIBA, S., KOBAYSHI, K. (1971), Comments on
stream, the concentrated sugar solution is not oxygen transfer into mycelial pellets, Biotechnol.
diluted as much as for shorter circulation Bioeng. 13, 583-588.
times, and local spots of high concentrations AIBA,S . , HUMPHREY, A. E., MILLIS,N. F. (1973),
will appear in the first vessel of the scheme in Biochemical Engineering, 2nd Ed., Tokyo: To-
Fig. 60. The control strategy, therefore, re- kyo University Press.
duces the feeding rate as a result of which sug- ANDREW,S. P. S . (1982), Gas-liquid mass transfer
ar concentrations in the larger part of the ves- in microbiological reactors, Trans. Znst. Chem.
sel are below the critical value to match the Eng. 60, 3-13.
ATKINSON, B. (1974), Biochemical Reactors, Lon-
critical growth rate.
don: Pion, Ltd.
BADER,F. G. (1987a), Modelling mass transfer and
agitator performance in multiturbine fermentors,
Biotechnol. Bioeng. 30, 37-51.
BADER,F. G. (1987b), Improvements in multi-tur-
4 Conclusions bine mass transfer models, in: Biotechnology
Processes, Scale-up andMixing (Ho, C. S . , OLD-
Problems related to the geometry and the SHUE, J. Y., Eds.), pp. 96-106, New York:
scale of reactors may arise wherever transport American Institute of Chemical Engineers.
processes are present along with the kinetics of BAILEY,J. E., OLLIS, D. F. (1986), Biochemical
Engineering Fundamentals, 2nd Ed., New York:
the reaction system. The concept of coupling McGraw Hill.
transport phenomena and reaction kinetics is BAJPAI,R. K., REUSS, M. (1982a), Coupling of
commonly used to predict the observed reactor mixing and microbial kinetics for evaluation of
performances for situations involving defined the performance of bioreactors, Can. J. Chem.
hydrodynamics. Stirred bioreactors operated Eng. 60, 384-392.
References 345
BAJPAI, R. K., REUSS,M. (1982b), Considerations CUTTER,L. A. (1966), Flow and turbulence in a
of macromixing and micromixing in semi-batch stirred tank, AZChE J. 12, 35-45.
stirred tank bioreactors, in: Chemical Reaction EINSELE,A., FINN,R. K. (1980), Influence of gas
Engineering (WEI, J., GEORGAKIS, C., Eds.), pp. flow rates and gas hold up on blending efficiency
555-565, Washington, D. C.: Boston ACS Sym- in stirred tanks, Znd. Eng. Chem. Proc. Des.
posium Series 196. Dev. 19, 600-603.
BAJPAI,R. K., SOHN,P. U. (1987), Stage models EINSELE,A., RISTROK,D. L., HUMPHREY, A. E.
for mixing in stirred bioreactors, in: Biotechnolo- (1978), Mixing times and glucose uptake meas-
gy Processes: Scale-up and Mixing (Ho, C. S . , ured with a fluorometer, Biotechnol. Bioeng. 20,
OLDSHUE,J. Y., Eds.), pp. 13-21, New York: 1487-1492.
AIChE Publ. ELLUL,I. R., ISSA, R. I., LOONEY,M. K. (1985)
BAJPAI, R. K., RAMKRISHNA, D., PROKOP,A. Numerical computation of multi-dimensional
(1977), Effect of drop interactions on the per- multi-phase flow, in: Numerical Methods in
formance of hydrocarbon fermentors, Biotech- Laminar and Turbulent Flow, Proc. 4th Int.
nol. Bioeng. 19, 1761-1772. Conf., Swansea, pp. 1218-1231.
BERKE,W. (1980), EinfluB der Diffusion im Mycel FAUST,U., SITTIG,W. (1980), Methanol as carbon
auf die Sauerstoffkinetik von Aspergillus niger, source for biomass production in a loop reactor,
Diploma Thesis, Technische Universitat, Berlin. Adv. Biotechnol. Eng. 17, 63-99.
BIRYOKOV, V. V. (1983), Computer control and op- Fox, R. I. (1978), The applicability of published
timization of microbial metabolite production, scale-up criteria to commercial fermentation
in: Modelling and Control of Biochemical Proc- processes, First Eur. Congr. Biotechnology, Zn-
esses (HALME,A., Ed.), pp. 135-144, Helsinki: terlaken, Preprints Part 1, pp. 1/80-1/83.
IFAC-Publication, Pergamon Press. GRIOT,M., MOES, J., HEINZLE,E., DUNN,I. J.,
BLAKEBROUGH, N., SAMBAMURTHY, K. (1966), BOURNE,J. R. (1986), A microbial culture for
Mass transfer and mixing rates in fermentation the measurement of macro and micro mixing
vessels, Biotechnol. Bioeng. 8, 25-42. phenomena in biological reactors, Proc. Conf.
BLENKE,H. (1979), Loop reactors, Adv. Biochem. on Bioreactor Fluid Dynamics, Cambridge, Pa-
Eng. 13, 121-214. per 16, pp. 203-216.
BLENKE,H. (1985), Biochemical loop reactors, in: GUNKEL,A. A . , WEBER,M. E. (1975), Flow phe-
Biotechnology (REHM,H.-J., REED, G., Eds.), nomena in stirred tanks, AZChE J. 21, 931-949.
Vol. 2, pp. 465-569, Weinheim-Deerfield Beach/ HANSFORD,G. S., HUMPHREY, A. E. (1966), The
Florida-Basel: VCH. effect of equipment scale and degree of mixing
BOELCKE,C. (1983), Experimentelle Untersuchun- on continuous fermentation yield at low dilution
gen zur Zirkulationsverteilung der Fliissigphase rates, Biotechnol. Bioeng. 8, 85-96.
in Riihrreaktoren, Diploma Thesis, Technische HARLOW,F. H., AMSDEN,A. A. (1975), Numerical
Universitat, Berlin. calculation of multiphase fluid flow, J. Comput.
BOLZERN,O., BOURNE,J. R. (1983), Mixing and PhyS. 17, 19-52.
fast chemical reaction - IV. Extension of the HARVEY,P . S., GREAVES,M. (1982a), Turbulent
reaction zone, Chem. Eng. Sci. 38, 999-1003. flow in an agitated vessel. Part I: A predictive
BRYANT,J. (1977), The characterization of mixing model, Trans. Znst. Chem. Eng. 60, 195-200.
in fermenters, Adv. Biochem. Eng. 5, 101-123. HARVEY,P. S., GREAVES,M. (1982b), Turbulent
BRYANT,J., SADEGHZADEH (1979), Circulation flow in an agitated vessel. Part 11: Numerical so-
rates in stirred and aerated tanks, in: Proc. 3rd lution and model predictions, Trans. Znst. Chem.
Eur. Conf. on Mixing, 4-6th April, York, Vol. I, Eng. 60, 201-210.
pp. 325-336. HEIJNEN,J. J., VAN’T RIET, K. (1982), Mass trans-
CALDERBANK, P. H. (1967), Mass transfer in fer- fer, mixing, and heat transfer phenomena in low-
mentation equipment, in: Biochemical and Bio- viscous bubble column reactors, Proc. ZVth Eur.
logical Engineering Sciences (BLAKEBROUGH, Conf. on Mixing, 27-29th April, Noordwijker-
N., Ed.), Vol. 1, New York: Academic Press. hout.
COSTES,J., COUDEREC, J. P. (1982), Pumping ca- HENZLER,H.-J. (1982), Verfahrenstechnische Aus-
pacity and turbulence intensity in baffled stirred legungsgrundlagen fiir Riihrbehalter als Fermen-
tanks; influence of the size of the pilot unit, ter, Chem. Zng. Tech. 54, 461-476.
Proc. 4th Eur. Conf, on Mixing, 28-29th April, HINES,D. A. (1978), The large scale pressure cycle
Noordwijkerhout, Paper B2, pp. 25-34. fermenter configuration, Proc. First Eur. Congr.
CURL,R. L. (1963), Dispersed phase mixing: I. The- on Biotechnology, Survey Lectures, DECHEMA-
ory and effects in simple reactors, AZChE J. 9, Monographien, Vol. 82, Nr. 1693-1703, pp. 55-
175-181. 64, Weinheim: Verlag Chemie.
346 10 Stirred Tank Models
HOLMES,D. B., VONCKEN,R. M., DEKKER,J. A. (1965), Model simulation of stirred tank reactors,
(1964), Fluid flow in turbine-stirred, baffled AZChE J. 11, 723-727.
tanks - I. Chem. Eng. Sci. 19, 201-208. MARINI,L., GEORGAKIS, C. (1984), Low-density
ISSA,R. I., GOSMAN,A. D. (1981), The computa- polyethylene vessel reactors. Part I: Steady
tion of three-dimensional turbulent two-phase state and dynamic modelling, AZChE J. 30, 401-
flows in mixer vessels, in: Numerical Methods in 408.
Laminar and Turbulent Flow, pp. 829-839, MCMANAMEY, W. J. (1980), A circulation model
Swansea: Prineridge Press. for batch mixing in agitated, baffled vessels,
JANSEN,H., SLOT,S., GORTLER,H. (1978), Deter- Trans. Znst. Chem. Eng. 58, 271-276.
mination of mixing times in large-scale fermen- MERSMANN, A ., EINENKEL, W.-D., KAEPPEL,M.
ters using radioactive isotopes, in: Preprints First (1979, Auslegung und Maktabsvergrdfierung
Eur. Congr. on Biotechnology Part 11, pp. 80- von Ruhrapparaten, Chem. Zng. Tech. 47, 953-
83, Poster Papers, Frankfurt: DECHEMA. 964.
JOSHI, J. B . , PANDIT,A. B., SHARMA,M. M. METZ,B. (1976), From Pulp to Pellet, Ph. D. The-
(1982), Mechanically agitated gas-liquid reactors, sis, University of Delft.
Chem. Eng. Sci. 37, 813-844. METZ,B., KOSSEN, N. W. F., VAN SUIJDAM,J. C.
JURY, W., SCHNEIDER,G., MOSER, A. (1988), (1979), The rheology of mould suspensions, Adv.
Modelling approach to industrial bioreactors,
Biochem. Eng. 11, 103-156.
Proc. 6th Eur. Conf. on Mixing, 24-26th May, VON MEYENBURG, K. (1969), Katabolit-Repression
Pavia, pp. 451-456. und der Sporungszyklus von Saccharomyces cere-
KHANG,S. J., LEVENSPIEL,0. (1976), New scale- visiae, Dissertation 4279, ETH Zurich.
up and design method for stirred agitated batch MIDDLETON, J. C. (1979), Measurement of circula-
mixing vessels, Chem. Eng. Sci. 31, 569-577. tion within large mixing vessels, Paper A2, 3rd
KNOEPFEL, H. P. (1972), Zum Crabtree-Effekt bei Eur. Conf. on Mixing, 4-6th April, York, Vol. I,
Saccharomyces cerevisiae and Candida tropicalis, pp. 15-36.
Dissertation 4906, ETH Zurich. MIURA,Y. (1976), Transfer of oxygen and scale-up
KOBAYASHI, T., VAN DEDEM,G., MOO-YOUNG, in submerged aerobic fermentation, Adv. Bio-
M. (1973), Oxygen transfer into mycelial pellets, chem. Eng. 4, 3-40.
Biotechnol. Bioeng. 15, 27-45. MOECKEL,H. 0. (1980), Die Verteilung der Ortli-
KOSSEN,N. W. F. (1985), Bioreactors: Consolida- chen Energiedissipation in einem Rtihrwerk,
tion and innovations, Review/Preview Lecture Chem. Tech. (Leipzig) 32, 127-129.
presented at the 3rd Eur. Congr. on Biotechnolo- VAN DER MOLEN,K., VAN MAANEN,H. R. E.
gy, Miinchen, Vol. IV, pp. 257-282, Weinheim- (1978), Laser-Doppler measurements of the tur-
Deerfield Beach/Florida-Basel: VCH Verlagsge- bulent flow in stirred vessels to establish scaling
sellschaft . rules, Chem. Eng. Sci. 33, 1161-1168.
KOSSEN,N. W. F., OOSTERHUIS, N. M. G. (1985), MUKATAKA,S., KATAOKA,H., TAKAHASHI,J.
Modelling and scaling up of bioreactors, in: Bio- (1976), Measurement of mean circulation time of
technology (REHM,H.-J., REED,G., Eds.), Vol. a mycelial suspension in a stirred tank, Kagaku
2, pp. 572-605, Weinheim-Deerfield Beach/Flo- Kogaku Ronbunshu 2, 628-630.
rida-Basel: VCH Verlagsgesellschaft. MUKATAKA,S., KATAOKA,H., TAKAHASHI,J.
LAI, K. Y. M., SALCUDEAN, M. (1987), Computer (1980), Effects of vessel size and rheological
analysis of multi-dimensional, turbulent, properties of suspensions on the distribution of
buoyancy-induced, two-phase flows in gas-agi- circulation times in stirred vessels, J. Ferment.
tated-liquid reactors, Comput. Fluids 15, 28 1- Technol. 58, 155-161.
295. MUKATAKA,S., KATAOKA,H., TAKAHASHI,J.
LEVENSPIEL, 0. (1972), Chemical Reaction Engi- (1981a) Circulation times of various fluids in stir-
neering, London: John Wiley & Sons. red fermentors, in: Advances in Biotechnology
LIEPE,F., MOECKEL,H.-O., WINKLER,H. (1971), (MOO-YOUNG,M., Ed.), Vol. I, pp. 523-528,
Untersuchungen in Ruhrmaschinen, Chem. Tech. Toronto: Pergamon Press.
(Leipzig) 23, 231-237. MUKATAKA,S., KATAOKA, H., TAKAHASHI,J.
MANN,U., CROSBY,E. J. (1973), Cycle time distri- (1981b), Circulation time and degree of fluid ex-
bution in circulating systems, Chem. Eng. Sci. change between upper and lower circulation re-
28, 623-627. gions in a stirred vessel with a dual impeller, J.
MANN, U., CROSBY,E. J., RUBINOVITCH, M. Ferment. Technol. 59, 303-307.
(1974), Number of cycles distribution in circulat- NAGATA,S. (1975), Mixing Principles and Applica-
ing systems, Chem. Eng. Sci. 29, 761-765. tions, New York: John Wiley & Sons.
MANNING,F. S., WOLF, D., KEAIRNS,D. L. NALLASAMY, M. (1987), Turbulence models and
References 347
their applications to the prediction of the internal (JOHNSON,A., Ed.), pp. 119-124, Oxford-New
flows: A Review, Comput. Fluids 15, 151-194. York-Toronto-Sydney-Paris-Frankfurt: Perga-
NAUMANN,E. B. (1981), Continuous mixing, mon Press.
Chem. Eng. Commun. 8, 53-127. REUSS,M., BAJPAI,R. K., BERKE,W. (1982), Ef-
OLDSHUE,J. Y. (1983), Fluid Mixing Technology, fective oxygen consumption in fermentation
New York: McGraw Hill. broths with filamentous organisms, J. Chem.
OKAMOTO,Y., NISHIKAWA, M., HASHIMOTO, K. Technol. Biotechnol. 32, 81-91.
(1981), Energy dissipation rate distribution in RIESMEIER, B. (1984), EinfluB der Begasung auf das
mixing vessels and its effects on liquid-liquid Rezirkulationsverhalten der Fliissigphase in ge-
dispersion and solid-liquid mass transfer, Znt. riihrten Bioreaktoren, Diploma Thesis, Tech-
Chem. Eng. 21, 88-94. nische Universitlt, Berlin.
OOSTERHUIS, N. M. G. (1984), Scale-up of Bioreac- VAN’T RIET, K. (1979), Review of measuring meth-
tors. Ph. D . Thesis, University of Delft. ods and results in non-viscous gas-liquid mass
OOSTERHUIS, N. M. G., KOSSEN,N. W. F. (1984), transfer in stirred vessels, Ind. Eng. Chem. Proc.
Dissolved oxygen concentration profiles in a pro- Des. Dev. 18, 367-375.
duction-scale bioreactor, Biotechnol. Bioeng. 26, VAN’T RIET, K., SMITH,J. M. (1975), The trailing
546-550. vortex system produced by Rushton turbine agi-
PACA, J., ETTLER,P., GREGR,V. (1976), Hydro- tators, Chem. Eng. Sci. 30, 1093-1105.
dynamic behaviour and oxygen transfer rate in a RIZZI, M. (1982), EinfluB des Energieeintrages auf
pilot plant fermenter. I. Influence of viscosity, J. die Sauerstoffaufnahme von Aspergillus niger in
Appl. Chem. Biotechnol. 26, 309-317. geriihrten Bioreaktoren, Diploma Thesis, Tech-
PANDIT,A. B., JOSHI,J. B. (1983), Mixing in me- nische Universitat, Berlin.
chanically agitated gas-liquid contactors, bubble ROELS,J. A. (1983), Energetics and Kinetics in Bio-
columns and modified bubble columns, Chem. technology, Amsterdam: Elsevier Biomedical
Eng. Sci. 38, 1189-1215. Press.
PARKER,D. S . , KAUFMANN, W. J., JENKINS,D. SCOTT,R. (1983), Design and evaluation of experi-
(1972), Floc breakup in turbulent flocculation ments to provide scale-up information for the
processes, J. Sun. Eng. Div. SAl, 79-99. ICI single cell protein process, in: BIOTECH 83,
PLACEK,J., TAVLARIDES, L. L. (1985), Turbulent Northwood, UK: Online Publ. Ltd.
flow in stirred tanks. Part I: Turbulent flow in SENIOR,P. J., WINDNASS, G. (1980), The ICI sin-
the turbine impeller region, AZChE J. 31, 1113- gle cell protein process, in: Biotechnology, a Hid-
1120. den Past, a Shining Future, 13th Int. TNO
PLACEK,I., TAVLARIDES, L. L., SMITH,G. W., Conj., The Hague.
FORT,I. (1986), Turbulent flow in stirred tanks. SINCLAIR,C. G., BROWN,D. E. (1970), Effect of
Part 11: A two scale model of turbulence, AZChE incomplete mixing on the analysis of the static
J. 32, 1771-1786. behaviour of continuous cultures, Biotechnol.
RAGOT,F., REUSS,M. (1990), A multi-phase com- Bioeng. 12, 1001-1017.
partment model for stirred bioreactors incorpo- SINGH,V . , HENSLER,W., FUCHS,R. (1986), On-
rating mass transfer and mixing, Poster, Paper line determination of mixing parameters in fer-
in: 2nd Int. Symp. on Biochemical Engineering, mentors using pH transient, in: Bioreactor Fluid
5-7th March, Stuttgart. Dynamics, 15-17th April, Cambridge, Paper 18,
REUSS, M. (1976), Bioverfahrenstechnik - Sauer- pp. 231-256.
stofftransport in Bioreaktoren, VDZ-Fortschr. SINGH, V., FUCHS, R., CONSTANTINIDES, A.
Verfahrenstech. 14, 551-563. (1987), A new method for fermentor scale-up in-
REUSS, M. (1983), Mathematical models for cou- corporating both mixing and mass transfer ef-
pled oxygen transfer and microbial kinetics in fects - I. Theoretical basis, in: Biotechnology
bioreactors, in: Modelling and Control of Bio- Processes, Scale-up and Mixing (Ho, C. S., OLD-
technical Processes (HALME,A., Ed.), pp. 33- SHUE,J. Y., Eds.), New York: American Insti-
45, Oxford-New York-Toronto-Sydney-Paris- tute of Chemical Engineers, pp. 200-214.
Frankfurt: Pergamon Press. SINGH, V., FUCHS, R., CONSTANTINIDES, A.
REUSS,M. (1988), Influence of mechanical stress on (1988), Use of mass transfer and mixing correla-
the growth of Rhizopus nigricans in stirred bio- tion for the modelling of oxygen transfer in stir-
reactors, Chem. Eng. Technol. 11, 178-187. red tank fermentors, in: Bioreactor Fluid Dy-
REUSS,M., BRAMMER, U. (1985), Influence of sub- namics, Proc. 2nd Int. Conj., 21-23th Sept.,
strate distribution on productivities in computer Cranfield, UK, pp. 95-115.
controlled baker’s yeast production, in: Model- SMITH,J. M. (1981), Chemical Engineering Kine-
ling and Control of Biotechnological Processes tics, 3rd Ed., New York: McGraw Hill.
348 10 Stirred Tank Models
STEEL,R., MAXON,W. D. (1962), Some effects of VONCKEN,R. M., HOLMES,D. B., DEN HARTOG,
turbine size on novobiocin fermentations, Bio- H. W. (1964), Fluid flow in turbine-stirred, baf-
technol. Bioeng. 4, 231-240. fled tanks - 11. Dispersion during circulation,
STEEL,R., MAXON,W. D. (1966), Dissolved oxy- Chem. Eng. Sci. 19, 209-213.
gen measurements in pilot- and production scale WANG,D. I. C., FEWKES,R. C. J. (1977), Effect of
novobiocin fermentations, Biotechnol. Bioeng. 8, operating and geometric parameters on the be-
97-1 08. haviour of non-Newtonian, mycelial antibiotic
VAN SUIJDAM, J. C. (1980), Mycelial Pellet Suspen- fermentation, Dev. Ind. Microbiol. 18, 39-56.
sion, Biotechnological Aspects, Ph. D. Thesis, WANG,H. Y., COONEY,C. L., WANG, D. I. C.
University of Delft. (1979), Computer control of baker’s yeast pro-
SWEERE,A. P. J., LUYBEN, K. C. A. M., KOSSEN, duction, Biotechnol. Bioeng. 21, 975-995.
N. W. F. (1987), Regime analysis and scale WEN, C. Y., FAN, L. T. (1975), Models for Flow
down: Tools to investigate the performance of Systems and Chemical Reactors, New York:
bioreactors, Enzyme Microb. Technol. 9, 386- Marcel Dekker.
398. YANO, T., KODAMA,T., YAMADA,K. (1961), Fun-
TRAGHARD,C. (1988), A hydrodynamic model for damental studies on the aerobic fermentation.
the simulation of an aerated agitated fed-batch Oxygen transfer within a mold pellet, Agric.
fermentor, Proc. 2nd Znt. Conf. on Bioreactor Biol. Chem. 25, 580-584.
Fluid Dynamics, 21-23th Sept., Cambridge UK, YOSHIDA, T. (1967), Studies on submerged culture
pp. 117-134. of Basidiomycetes. The oxygen transfer within
UHL, V. W., GRAY,J. B. (1966), Mixing, Vols. 1 pellets of Leutinus edodos, J. Ferment. Technol.
and 2, New York: Academic Press. 45, 1119-1129.
11 Tower Reactor Models
JOSE C. MERCHUK
Beer Sheva, Israel
1 Introduction 351
2 Flow Modelling 352
2.1 Flow Configuration 352
2.1.1 Flow Regime Transitions 353
2.1.1.1 Homogeneous Bubbly Flow-Slug Flow 353
2.1.1.2 Slug Flow-Dispersed Bubble Flow 353
2.1.1.3 Slug Flow-Churn-Turbulent Flow 354
2.2 Bubble Column Flow Models 354
2.3 Air Lift Reactor Flow 359
2.3.1 The Drift Flux Model 359
2.3.2 Air Lift Reactor Flow Models 360
2.4 The Axial Dispersion Model 364
2.5 Summary of Flow Models 367
3 Mass Transfer Modelling 368
3.1 Prediction of Mass Transfer Rates 369
3.2 Structured Models for Mass Transfer in Tower Reactors 370
4 Process Models in Tower Reactors 372
4.1 Bubble Column Modelling 372
4.2 Air Lift Models 374
5 Operation Policies 378
6 Closing Remarks 379
7 References 379
350 11 Tower Reactor Models
List of Symbols kl
L
first-order kinetic constant (s-')
ungassed liquid height (m)
equivalent lkngth (m)
equivalent length of the bottom section
cross-sectional area (m2) (diameters)
interfacial area per unit volume (m-') mixing length (m)
biomass concentration (kg mU3) productivity ratio, Eq. (1 10)
dimensionless biomass concentration ratio of gas to liquid flow rate, Eq.
(B/ Yx sf) (1 11)
Bodenstein number, Eq. (60) molar flow rate (mol s-')
concentration of component B total number of cells
(mol L-') flow index in power law model; param-
liquid phase concentration in stage j eter of gas holdup radial profile, Eq.
distribution parameter (17)
constant in the friction factor correla- inlet molar flow rate (mol s - I )
tion pressure (Pa)
saturation concentration of oxygen in Peclet number (-), Eq. (59)
the liquid phase (mol L-') radial point of inversion of liquid ve-
diffusivity coefficient (m2 s), Eq. (79) locity
tower diameter (m) gas flow rate (m3 s - I )
hydraulic diameter (m) liquid flow rate (m3s - I )
liquid phase dispersion coefficient constant in Eq. (6)
(m s - ~ ) radius (m); recycle ratio (-); ideal gas
Sauter diameter (m) constant
energy dissipation in the liquid (W) dimensionless stream function, Eq. (1 1)
friction factor (-) dimensionless reaction group, Eq. (109)
constant in Eq. (87) Reynolds number, Eq. (84)
fraction of reversed flow gas in the respiration rate of microorganisms
downcomer ( - ) ( ~ o ~ O ~ L - ~ S - ~ )
gravitational acceleration (ms - 2 ) radial distance (m); fraction of en-
height (m) trained gas in Eq. (98)
Henry's constant (Pa m3 mol-') substrate concentration (kg m-')
effective width, in Eq. (52) (m) dimensionless concentration of sub-
superficial gas velocity (m s - I ) strate ( S / S , )
superficial liquid velocity (m s - I ) ratio of mean residence times in the
drift velocity (m s-'), Eq. (24), Eq. perfectly mixed and plug-flow zones,
(25) Eq. (65)
friction coefficient; consistency index in Schmidt number, Eq. (83)
power law model (Pa s") Sherwood number, kLaD2/D
loss coefficient for change in velocity Stanton number for the gas phase, Eq.
direction ( - ) (102) (-1
loss coefficient for change in velocity Stanton number for the liquid phase,
magnitude ( - ) Eq. (103) (-1
dimensionless group ( K M Sf) temperature (K)
constant in Eq. (49, Eq. (70) time (s)
dimensionless group (KM/Si) batch time (s)
Michaelis-Menten constant (mol L - I ) exposure time (s)
coalescence factor, Eq. (89) velocity (m s - I)
mass transfer coefficient (m s - I ) slip velocity (m s - I )
initial mass transfer coefficient (m s - '), bubble terminal velocity (m s - I )
Eq. (88) velocity of slug (m s -')
local value of kL volume (m3)
Biotechnology Second, Completely Revised Edition
Edited by H.-J. Rehm and G.Reed in cooperation with
A. Puhler and P. Stadler
copyright@WILEY-VCH Verlag GmbH, D-69469 Weinheim (Federal Republic of Germany). 2001
Introduction 35 1
2.1.1 Flow Regime Transitions are moving on oscillatory paths, they will col-
lide frequently before this fraction is attained.
BARNEA and TAITEL(1986) have reviewed If it is assumed that the minimal distance be-
the criteria for predicting the transitions from tween bubbles that allows them freedom of
one flow configuration to the other. movement is half their radius, then a maximal
gas fraction of 25% is obtained.
Defining the superficial gas and liquid velo-
2.1.1.1 Homogeneous Bubbly cities as:
Flow-Slug Flow
This transition can be depicted as a series of
states starting with small bubbles generated at
low rates, which behave as rigid spheres rising
rectilinearly. But as gas flow rate increases, so The relative velocity of the gas phase is:
does the bubble diameter. Above a certain di-
ameter, which is 0.002-0.003 m in air/water
systems at atmospheric pressure, spherical
caps appear and with them zigzag motion. In the range of bubble size of interest, the rela-
Considerable turbulence develops, with ran- tive velocity U,,depends only very slightly on
dom collisions between bubbles and conse- the bubble diameter. Taking HARMATHY’S
quent coalescence. Larger Taylor-type bubbles (1960) expression
may appear, rising much faster than the small-
er ones. As they take over those smaller bub-
bles in their path, they increase in volume until (4)
they finally bridge the whole cross-section of
the tower (TAYLOR,1954). However, this oc- and 9 = 0.25, Eqs. (1) to (4) give
curs, as can be seen in Fig. 1, only for a tower
diameter smaller than 0.10-0.15 m. For larger
diameters, slug flow as defined above is not
observed. Instead, an increase in superficial
gas velocity leads directly to the churn-turbu- which becomes the limit between bubbly flow
lent regime. This configuration offers a small- and slug flow on the ( J L - J G ) plane. This limit
er gas-liquid interfacial area than the bubble depends on liquid properties only, since for the
flow, but the mass transfer coefficient kLis en- situations of interest in tower reactor design pG
hanced because of the turbulence, balancing at is usually negligible, and geometric parameters
least partially the loss of interfacial area. do not appear in Eq. (5).
BARNEAand TAITEL(1986) have proposed
a series of criteria for characterizing the transi-
tion between the flow configurations in two- 2.1.1.2 Slug Flow-Dispersed Bubble
phase flow. For the case of the transition from Flow
bubble to slug flow, it is accepted that a slug
will form when a very high gas holdup is at- The authors then define a boundary between
tained. GRIFFITHand SNYDER(1964) found slug flow and dispersed bubble flow, a flow
this to happen when the gas holdup 9 is configuration where small bubbles are dis-
around 0.25 to 0.30. In this situation it is as- persed into a very turbulent liquid. The condi-
sumed that the bubbles are so close to one an- tion is:
other that the frequency of collision and coal-
escence increases sharply. BARNEAand TAI-
TEL consider the maximum allowable packing
of the bubbles. The maximum fraction of rigid
spheres packed in a cubic lattice is 0.52. But
since the bubbles are not perfect spheres and
* u : ( ~ - ’ ) ’=0.725
~
(3
+ 4.15 - o’5
354 I1 Tower Reactor Models
and is based on considering a balance between (9,(6), and (7) can be seen in Fig. 2 . This fig-
surface tension and forces due to turbulent ure is for a 5.1 cm diameter column. As the
fluctuations. The transition from slug flow to diameter increases to sizes representative of
dispersed bubble flow would occur when the bench scale or pilot plant size, the slug flow re-
turbulent breakup process is able to disinte- gime disappears and instead becomes a diffuse
grate the Taylor bubble into bubbles small and uncertain area between bubble flow and
enough to remain spheric. However, the high churn-turbulent flow. Most of the cases of in-
liquid velocities required for that process ex- terest for chemical and biochemical tower
ceed the usual operational range of tower reac- reactor design fall either into this area or into
tors, and, therefore, this boundary is of little the bubble flow region.
relevance for our purposes.
2.2 Bubble Column Flow Models
2.1.1.3 Slug Flow-
Churn-Turbulent Flow In the case of bubble column operation,
where there is little or no net liquid flow, the
For the transition between slug flow and bubbly flow configuration covers two sub-re-
churn-turbulent flow, the expressions given by gions.
BARNEAand TAITEL (1986) can be written At low gas superficial velocity, the bubbles
as: generated in the gas sparger rise without much
(T-
JL = 0.825 1) J G - 0 . 2 2 w
coalescence. Bubble breakup is also negligible
in this region, and the initial bubble diameter
(7) is maintained, changing only because of hy-
drostatic changes or interchange of gas with
In Eq. (7), the diameter of the tower, 0, the liquid phase. Under these conditions, most
plays an important role. As D increases, the of the mass transfer takes place near the sparg-
superficial liquid velocity, JL, required for er (ZHAOet al., 1988). As superficial gas vel-
transforming a slug flow into churn-turbulent ocity increases, bubble size, bubble oscilla-
flow for a given superficial gas velocity, JG, tions, and liquid turbulence increase, coales-
decreases. The three boundaries given by Eqs. cence and breakup appear in a certain meas-
ure, but the general behavior does not change
very much. The rate of the gas holdup increase
in this region is rapid and nearly linear. But as
the superficial gas velocity increases further,
the so-called recirculation sub-regime is at-
tained. Under these conditions, a central as-
cending and oscillating liquid path develops
with recirculation cells at both sides of it, as
can be seen in Fig. 3. These patterns can be
easily seen in two-dimensional bubble col-
umns. MERCHUK (1986) compared the two-di-
mensional flow patterns in a bubble column
and an airlift reactor.
When operating in the bubble column mode
I and at low gas velocities, bubbles ascend in
0.0014
0.01 0.1 1.0 10 100 straight lines, especially in the lower half of
JG Irnlsl the column. In the top half of the column,
Fig. 2. Flow pattern transitions for a 5.1 cm diame- they begin to oscillate (Fig. 3A), apparently
ter tube, upward air-water flow. 0.1 MPa, 25 "C; 7 due to the increase in bubble diameter caused
bubble or dispersed bubble flow; 0 slug flow; o by the decrease of the hydrostatic pressure and
churn flow. A annular flow E, Eq. (5); G, Eq. (6); by some degree of coalescence. However, as
H, Eq. (7) (from BARNEAand TAITEL,1986). the gas velocity increases, the oscillations of
Flow Modelling 355
Air
Ilft
6 00
0 a ov 0.3 0
P
05
~~
0 0
a a &a I 1.0 A A
I I Porous v V
1 I
IU
10-2 10-1 100 Fig. 3. Recirculation patterns in a bub-
JE i m h l ble column (from MERCHUK,1986).
the bubbles increase in amplitude, and a mean bubbling regime and the liquid recirculating
stream is generated characterized by high local regime. They exist in bubble columns having a
velocity and larger bubbles. Recirculation cells sparger hole diameter of less than about 0.01
with small bubbles that oscillate in accordance m (WALLIS,1969). One of the characteristics
with the mainstream line appear at the sides of of the transition regime is an increase in the
the column (Fig. 3B). With a further increase gas holdup. A critical superficial velocity has
in gas velocity, the frequency of oscillations in- been defined (SAKATAand MIYAUCHI,1980;
creases sharply while the amplitude remains MARUYAMA et al. 1981) as the superficial gas
constant or increases slightly. This leads to a velocity at maximum holdup, and this is sup-
sharp increase in turbulence. The recirculation posed to be the point at which incipient regular
flow rate in the side cells increases at the ex- recirculation begins.
pense of the central path, which, under these The modelling of the recirculation regime is
conditions, consists of the largest bubbles very important for prediction of the mixing
present in the system (Fig. 3C). A clear down- characteristics of bubble columns.
stream pattern develops, bringing the smallest One of the more widely accepted models for
bubbles to the bottom of the column. At even the hydrodynamics of bubble columns is the
larger values of the superficial gas velocity, the circulation cell model by JOSHI and SHARMA
relative importance of the main central path (1979). It is based on the concepts of stream
increases. The amplitude of the oscillations de- function and of vorticity, and on their rela-
creases, and most of the gas passes rapidly tionship as given by LAMB(1932):
through the center. The loops at the sides be-
come even more turbulent, and the down-
stream of recirculating liquid is clearly visible
at both sides of the column.
Flow configurations B and C correspond to This model assumes axial symmetry, and
the “transition regime” between the uniform the height of each circulation cell is obtained
356 I1 Tower Reactor Models
by means of a criterion of minimum vorticity and yo is the maximum value of the stream
and an energy balance. It was found that the function. This function is calculated under the
minimum vorticity appeared when the height assumption that all the kinetic energy asso-
of the circulation cell was equal to the tower ciated with the downward flow of the liquid is
diameter, for any values of the diameter. The dissipated in the turbulence. JOSHI et al.
number of circulation cells is therefore equal (1986) give relationships for the energy dissipa-
to the total dispersion height divided by the di- tion by the liquid, E, obtained from energy
ameter (Fig. 4). The model predicts the follow- balances, in which the kinetic energy related to
ing axial and radial components for the veloci- the inflowing gas, the energy associated with
dissipation at the column wall, and the energy
used for bubble breakup are neglected. The in-
tegration of Eq. (8) to give Eqs. (9) and (10) is
facilitated by the fact that, as shown by LAMB
(1932), the vorticity divided by the radial dis-
tance from the axis depends on the stream
function only:
where k2 is a constant.
This model does not agree with observations
on two-dimensional bubble columns (MER-
CHUK, 1986), especially because of the as-
sumed axisymmetry. Also, the axial compo-
nent at the wall, Eq. (9), would change direc-
tion in passing from one cell to another be-
cause of the function sinus. This contradicts
the widely accepted fact that the flow at the
wall is downwards all along the tower in the
recirculation regime. This has been criticized
by VAN DER AKKERand RIETEMA(1982). How-
ever, JOSHIand SHARMA(1979) have shown
Fig. 4. Multiple circulation cells in a bubble column that their model allows a satisfactory predic-
(JOSHIand SHARMA,1979). tion of the axial velocity and the axial disper-
sion coefficient. Later on, JOSHI(1980) modif-
ied the model in order to take into account in-
ty as a function of the tower diameter, r, and teraction between adjacent cells. JOSHI et al.
the axial position, z: (1986) have recently reviewed the achievements
of this model when applied to bubble columns,
yo 1 dRi gas-liquid-solid sparged reactors, and gas-sol-
u,= - - -sin(nz*)
(D/2)2 r* d r *
(9) id fluidized beds. A model of similar charac-
teristics was applied by VAN DER AKKERand
RIETEMA(1982) to liquid-liquid systems.
u,=--2 = yo 5 ! cos(rcz*> The recirculation model has lately been ex-
DH r* tended to a three-dimensional version which
assumes cylindrical eddies piled up transverse-
where U, and U, are the axial and radial com- ly to each other in the column as depicted in
ponents of the liquid velocity, respectively, R 1 Fig. 5 (ZEHNER, 1986a). This arrangement
is the dimensionless stream function: solves the problem of interference between ed-
dies if a downward path along the wall is ac-
cepted. ZEHNERderives a n equation for the
Flow Modelling 357
lowing equation:
dU
-v, - = (1 -q)G;ii;
dr
where ii; and ii; are the fluctuating components
of the liquid velocity in the r and z direc-
tions.
The profile of the gas holdup along the ra-
dius is assumed to be given by:
]
"
1 d dP
- - - (t)= - + (1 - p ) p L g
r dr dz
The shear stress 7 is then related to the ve-
locity through the molecular and turbulent vis-
cosity by:
-0.24 , , , , , ,
7= -(VM+Vt)pL
0- 0 0.2 0.4
cpl-I
0.6 0.8 1.0
This point of zero velocity has been used by 1968). They use published expressions for the
other investigators as the main parameter, in- axial shear stress across the diameter of the
stead of the turbulent viscosity used by UEYA- tower (LEVY,1960), for the radial distribution
MA and MIYAUCHI(WALTERand BLANCH, of the gas holdup (BANKOFF,1960) and for the
1983; YANG et al., 1986). In this way, a clear mixing length (SCHLICHTING,1968), and fi-
advantage is obtained since flow inversion can nally obtain velocity profiles that fit the exper-
be measured directly, in contrast to turbulent imental data obtained by HILLS(1974).
viscosity, which has to be obtained by model The model by CLARKet al. (1987) has no
fitting. analytical solution so that the velocity distribu-
The solution to the equation of motion, Eq. tion must be obtained by numerical integra-
(14), with the holdup distribution given by Eq. tion. The solution is a function of the shear at
(17), has been obtained by YANGet al. (1986) the wall. By integration of the velocity profile
with the boundary conditions of symmetry in over the cross-sectional area, the net liquid
the center of the tower, and zero liquid veloci- flow rate is obtained:
ty at a point where (0 = r / R = Q (flow inversion
R
point). The correct expression is:
QL = j 2 n r U ( r ) d r (19)
r, 0
UL
-- - l+Q-'p'+
UL,, Since QL is usually known, or measurable,
the wall shear corresponding to the system is
2-Q-'
+ [Qfl- 6) ('"*'-
Q"P*> (18) obtained.
While the recirculation regime is the usual
flow state in tower reactors in batch operation
or with low liquid velocities, visual observa-
This is a very simple expression depending on tions show clearly that air lift operation leads
two parameters: n, which represents the pro- to much higher liquid velocities. VERLAANet
file of gas holdup, and Q which therefore en- al. (1988) extended the obsevations of MER-
compasses the influence of geometry and liq- CHUK and STEIN(1981a) and gave a criterion
uid properties. It is tempting to consider that, for the transition between bubble column (re-
since gas holdup depends on the same varia- circulation regime) and air lift flow. When the
bles, it should be possible to correlate Q with (0 liquid circulation velocity, as calculated from
to obtain a single parameter model. However, JOSHI and SHARMA'Smodel (1979),
at the present time, not enough independently
obtained data are available. UcL = 3.0 [D(JG - 4 U,)]1'3 (20)
Eq. (18) is similar to the result given by
WALTER and BLANCH (1983). YANG et al. is much higher than the liquid superficial ve-
(1986) go further and extend their model to co- locity:
current columns by simply adding the superfi-
cial liquid velocity to UL. This is, however, an
oversimplification. Net liquid velocity induces
UCL * JL (21)
changes both of the gas holdup and the flow the recirculation regime predominates and the
pattern. It follows that both n and Q, the pa- above described models are applicable. On the
rameters of the model, are in fact functions of other hand. if
U,. This is not apparent in the model by
YANG et al. (1986), and such a function is
neither given nor suggested.
CLARKet al. (1987) presented a model for the recirculation has a minor role in the fluid
turbulent circulation in bubble columns that dynamics of the system, and other types of
predicts the radial distribution of the axial liq- models, as described in the next section,
uid velocity in tall columns with fully develop- should be used.
ed turbulent flow. Their analysis is based on The criteria proposed by VERLAANet al.
the mixing length theory (SCHLICHTING, (1986) can be evaluated with the data in Fig. 7.
Flow Modelling 359
The second term in the right-hand side of latter from easily measurable variables. In
Eq. (29) depends on both the distribution of 4 their analysis, the energy input is the potential
and on the local drift flux of the gas. For the energy of the gas bubbles once inside the reac-
case when interaction between bubbles is not tor, thus neglecting the kinetic energy asso-
negligible, ZUBERand HENCH(1962) give the ciated with gas injection. This energy input is
following expression: equal to the sum of the energy dissipation due
to the downward movement of entrained bub-
JG=
1.53 [r]
agAp
-4)3’2
1’4
(1
bles in the downcomer, the energy converted
into kinetic energy of the liquid, the energy
losses due to the change of direction and mag-
which is valid for the bubbly churn-turbulent nitude of liquid velocity at the bottom of the
regime. After integration, reactor, and the energy losses due to friction
of the gas-liquid mixture on the walls. The au-
thors present the final expression of the bal-
ance as follows:
a=(K+ML) + -
dr
(33) (35)
dd
practice has established the need for specific and M encompasses the friction factors for the
models for the corresponding fluid dynamics, two-phase flow in the riser, fr, and downcom-
especially for liquid velocity and mixing. er, fd:
Two main methods have been used for the
modelling of the two-phase flow in air lift
reactors: energy balances and momentum bal-
ances.
CHAKRAVARTY et al. (1974) used the energy LEE et al. (1986) used a slightly different
balance approach in order to obtain a link be- analysis to predict the liquid velocity in air lift
tween superficial gas velocity, holdup, and liq- reactors. They obtained the following expres-
uid velocity, which allows an evaluation of the sion from the energy balance:
Flow Modelling 361
This equation has the particularity that the where i can be riser, downcomer, and bottom
gas flow rate, the main and many times the section. In the case of the latter, L is replaced
only manipulable variable in the operation, is by L,, the equivalent length.
not present directly, but exerts its influence The pressure drop balance gives:
through the gas holdup. CHISTIet al. (1988)
show that most of the published data on liquid LPLg(+r - 4 d = ZApi (47)
velocity for the different types of air lift reac-
tors can be satisfactorily correlated by Eq. (44) which, assuming the same frictional coefficient
by choosing adequate values for the friction for all sections, leads to the following relation-
coefficients in each case. Only one coefficient ship between holdup, velocities, and geometric
needs to be adjusted, since the authors assume characteristics of the air lift reactor:
that KT, the friction coefficient at the top of
the loop, is negligible in concentric tube reac-
tors, and that in external loop reactors KT can
be taken as equal to KB, the friction coefficient
for the bottom of the loop.
The authors present an empirical correlation
of KB for their fitting:
(45)
where I is the equivalent length of the bottom
section in terms of the hydraulic diameter, R is
the recycle ratio,
where AB is the minimal cross-section at the
bottom of the airlift reactor. R = QGd/QGr (49)
Although the fitting of the data is accepta-
ble, it seems that the method of adapting KB /3 is the “flowing volumetric concentration”
must be improved in order to use it for scale- defined by ZUBERand FINDLAY
(1965),
up and design.
CHISTIand MOO-YOUNG(1988) further ex-
tended this model in order to predict the liquid
circulation in air lift reactors operating with
pseudoplastic fluids such as mold suspensions. applied in this case to the riser, and C is the
This is a very important improvement, since ratio between cross-sectional areas of riser and
many commercially useful fermentations in- downcomer,
volve such non-Newtonian liquids.
The other technique, used by several re-
searchers to predict liquid velocity, is the
C=-A
Ad
momentum balance in the air lift reactor. This
method has been used by BLENKE(1979), HSU MERCHUKand STEIN (198 1a) considered
and DUDUKOVIC (1980), KUBOTA et al. (1978), the special case of their experimental system,
BELLO (1981), and KOIDE et al. (1984). where no gas was recirculated, and obtained a
MERCHUKand STEIN(1981a) presented a justification for their experimental finding that
simple model for the prediction of the liquid the liquid velocity was proportional to the su-
velocity as a function of the gas input in an air perficial gas velocity to the power of 0.4.
lift reactor. They assumed that the pressure KUBOTAet al. (1978) have used a similar ap-
drop at the bottom of their external loop reac- proach to the analysis of ICI’s “deep shaft”
tor could be expressed as a continuation of the reactor. This is one of the largest air lift reac-
downcomer, using an equivalent length L,. tors ever built. It consists of a long, vertical
The frictional pressure drop in the different shaft more than 100 meters deep. KUBOTA et
sections of the loop can be defined as: al. (1978) presented a model for liquid circula-
tion in this reactor, which has the characteris-
tic that under steady-state operation the air is
injected not at the bottom, but in the down-
Flow Modelling 363
2.4 The Axial Dispersion Model from which the dispersion coefficient, D,, can
be obtained if the response of the system to a
pulse of tracer is measured. The results are
A complete model of a tower reactor would usually presented as the Peclet number:
require velocity vectors at each point of the
system, for both gas and liquid phases. This Pe = U, H/D, (59)
information, as seen in the discussion above, is
not always available with a high degree of cer- or as the Bodenstein number:
tainty because of the complicated nature of the
fluid dynamics in two-phase flow. Further-
more, the solution of this type of model de-
mands a high degree of sophistication and In the case of the net liquid flow in the col-
computational effort. For this reason, it is cus- umn, Eq. (56) must be considered. This case
tomary to use models for the residence-time was solved by VAN DER LAAN(1958) following
distribution and mixing which allow the con- the original treatment by WEHNERand WIL-
centration of all the mixing characteristics HELM (1956). The solution can be expressed as
within a few parameters. a relationship between the variance of the tem-
The axial dispersion model has the advan- poral change of the concentration of a tracer
tage of having a single parameter and is widely after a pulse injection and the Peclet number:
accepted for the representation of tower reac-
tors.
This model is based on a visualization of the
2
t =-
'
Pe
+-'
Pe2
[ 1 - exp ( -Pe)]
mixing process in the tower reactor as a ran-
dom, diffusion-like eddy movement superim- In the case of airlift reactors, the closed sys-
posed on a plug flow. The axial dispersion tem boundary conditions which lead to Eq.
coefficient D, is the only parameter in the for- (59) are no longer valid.
mulation, The case of a loop reactor with open bound-
ary was dealt with by VONCKENet al. (1964).
ac
-= 0,-+
a2c -
U-
ac The mathematical procedure was modified by
at az2 az MURAKAMIet al. (1982), and a simplified ap-
proximation was offered. But this solution re-
where C is the concentration of a tracer. The fers to a system which has a constant value of
boundary conditions depend on the specific
type of tower reactor.
For a batch bubble column operation, the
net average velocity is nil, and Eq. ( 5 5 ) is re-
3
duced to:
ac a2c
-=DD,- (57)
at az2
Riser Down-
This is a closed system as far as the liquid is TI 4
concerned. Using the appropriate boundary
+
and initial conditions, OHKIand INOUE(1970)
gave the following solution:
QU"o c2
-C_ - R
ce Qtf
co n27c D, t
1 + 2 n2
=l [(COS?.~) exp(- H2 )](58) Fig. 10. Model for a two-section loop reactor (WAR-
NECKE et al., 1985).
Flow Modelling 365
the dispersion coefficient over the whole vol- and the mean time is:
ume. This is not true for airlift reactors be-
cause of the different hydrodynamic character-
istics explained in Sect. 2.3.2.
WARNECKE et al. (1985) presented a model
showing the importance of considering each WARNECKE et al. (1985) showed that this
section as a separate unit, as well as the inter- model can reproduce the behavior of an airlift
action between the sections (Fig. 10). The reactor, assuming that R1 and R2 are repre-
model is based on two sections R1 and R2 (riser sented by reactors ranging from plug flow to
and downcomer) with liquid flow rates QL1 perfect mixing (Fig. 11).
and QL2,with a recycle ratio: A model completely opposed to the above
was presented by MERCHUKand YUNGER
(1990) for an air lift reactor. Instead of a riser
and a downcomer directly interconnected, and
and with mean residence times and variances each with a different dispersion coefficient, as
t l ,T ~ 0, : and a
:. Thus, this model covers the assumed by WARNECKE et al. (1985), they as-
case of net liquid flow through the reactor sumed that both riser and downcomer behave
(continuous operation). In this case, the over- as plug-flow sections, and all the mixing oc-
all variance is: curs solely in the gas separator, presented as a
perfectly mixed stage. The response of the sys-
tem to a pulse input is given by:
-- exp(ns-t/T2) (65)
I -5=3
---i;i/
150.
late the liquid circulation velocity in bubble where U: is the averaged fluctuation velocity
column reactors: of the slurry and I, is the mixing length for
mass transfer. Assuming spherical bubbles of
UCL = 3 1 [H(u, - E U,)]
1 l3 (66) uniform diameter in a cylindrical tower, an
overall energy balance and appropriate simpli-
Eq. (66) can be used in the relation pro- fications allow the derivation of the following
posed by JOSHIand SHARMA(1979) for calcu- expression:
lating the dispersion coefficient in the liquid
phase:
JG
(5)
(y)] 1/2
fi
R
UL
.JL= 12x:r(l-4) -dr u ~ j( m / S I D = 0.2m
0 x: R 2 A 0.21
o 0.34
where the local velocity UL is given by: 0.48
V 0.68
0.12 e 0.87
0.10 e
5, = ~
PL
(1 1.63)2 uLw ' '
uLw (75)
2.5 Summary of Flow Models
The two-phase flow in bubble columns is a
Their final equation for the axial dispersion very complex phenomenon, and exact descrip-
coefficient is: tion is still beyond our knowledge. But the use
368 11 Tower Reactor Models
of sensible assumptions has permitted the for- Whatever the method used, the basis of the
mulation of several mathematical models technique is the comparison of a measured
which allow acceptable approximations for variable with the value (or values) predicted by
both the liquid velocity and the mixing charac- a mathematical model of the process. It has al-
teristics of the system. In particular, the axial ready been pointed out that the choice of the
dispersion model is one of the most accepted model is very important and a poor assump-
models for the representation of mixing, and tion of gas or liquid phase flow characteristics
models have been published for the different may lead to errors and deviations from the
types of tower reactors. true values (KEITELand ONKEN,1981; MER-
CHUK and SIEGEL, 1988).
All published data on mass transfer in tower
reactors are based on fitting very simple fluid
dynamic models to experimental values of con-
centration, measured either in steady or un-
3 Mass Transfer Modelling steady state operation. These models assume
either plug flow or complete mixing of the two
The determination of mass transfer coeffi- phases present.
cients in laboratory or industrial equipment The most common assumption is complete
can be based either on steady-state or on un- mixing (CM) in the liquid phase and plug flow
steady-state mass transfer experiments. For (PF) in the gas phase. This approach is justi-
gas-liquid systems, these experiments can be fied by considering the relaxation times of the
either absorption or desorption studies. processes involved (ROELS, 1983).
Steady-state experiments are associated The ratio defined by DECKWER(1986),
either with the measurement of small differ-
ences in the concentration of continuous
phases or, in the case of gas absorption in 4 = liquid phase mixing time -
-
mass transfer time
batch systems, with the provision of a sink for
the absorbed gas by means of a chemical reac-
tion. Such chemical reactions imply the pres- (77)
l/kLa
sence of reactive solutes in the liquid that may
greatly influence the physical characteristics of may be taken as a sufficient indicator. A low
the system. Therefore, most of the experimen- value of this ratio implies that mixing is much
tal results are largely restricted to the system faster than mass transfer, and therefore the
being studied. This is due to the sensitivity of concentration is homogeneous throughout the
gas dispersion characteristics, especially of in- reactor. In such a case, the CM assumption
terfacial areas and gas holdup, to variations in seems to be justified, and therefore the values
density, viscosity, ionic strength, surface ten- of the mass transfer coefficient obtained from
sion, etc. Since the hydrodynamics of the sys- fitting the experimental data to the model are
tem are so heavily dependent on gas holdup, reliable and valid for scale-up. But problems
the entire reactor performance is affected by may arise when applying these data to the
changes in the physico-chemical properties of modelling of an absorption with chemical
the liquid phase. reaction, where the mass transfer rate is in-
Unsteady-state methods for determining creased because of the modification of the pro-
mass transfer coefficients, k,u, are compli- files of the absorbed component near the gas-
cated by the possibility of distortion of the liquid interphase (DANCKWERTS,1970). The
measured values by the probe response dynam- theory of absorption with chemical reaction al-
ics. Probe response dynamics become an in- lows the prediction of a newly increased coeffi-
fluential factor in k , calculations
~ when the cient, on the basis of the kLu without chemical
value for k,u is approximately equal to or reaction, the kinetics and the diffusion con-
greater than the inverse of the value for the stants. But the problem that is usually not ad-
probe response time. These factors complicate dressed is that under the chemical reaction re-
the modelling. gime the first assumption of perfect mixing in
Mass Transfer Modelling 369
the bulk of the liquid phase may no longer be The model is based on the penetration the-
valid, since kLa in Eq. (75) is increased and ory, which predicts the following simple de-
therefore the mass transfer coefficient based pendence of the mass transfer coefficient k L
on this fluid-dynamic model may also no long- on the diffusivity and the contact time of an
er be valid. element of liquid with the gas:
Had the experimental mass transfer coeffi-
cient kLabeen obtained by the fitting of a real-
istic fluid-dynamic model, kLa would be valid (79)
under any condition irrespective of the kinetic
regime. But, as stated before, the state of the The contact time is obtained by use of Kol-
art is still far from this achievement. gomoroff's length scale, which results in a
The key for further improvements in this di- function of the energy dissipation in the tower
rection lies in recognizing that mass transfer reactor, E , finally giving:
rates depend strongly on the hydrodynamics,
and in the quite common case of sparingly sol-
uble gases, specifically on the dynamics of the
liquid phase near the gas-liquid interface. It
follows that a connection must be found be-
tween the models presented above for the flow where n is the index in the power law model.
in a tower reactor and the mass transfer rate in The interfacial area per unit volume, a, was
it. expressed by use of CALDERBANK'S (1958)
equation for bubble size and an equation given
by KAWASE and MOO-YOUNG(1986) for the
3.1 Prediction of Mass Transfer gas holdup. The volumetric mass transfer
coefficient is then given in a dimensionless
Rates form as:
N!k!2
agreement of experimental data with Eq. (81),
for both non-Newtonian and Newtonian
(n- 1) fluids. They found the experimental re-
sults presented by various researchers to fall
within f 50% of the predicted values.
While the above model is certainly encour-
aging, it should be understood that it is not the lnterphase
ideal connection between the field of flow and mass transfer
the mass transfer rate. The model by KAWASE N-2
et al. (1987), as with others previously pub-
lished, is based on the specific energy dissipa-
tion in the reactor for the fluid-dynamic evalu-
ation. However, the specific energy dissipation
is a macroscopic parameter that lumps all the
volume into a single value. It is obvious that
different geometric arrangements will give dif-
ferent distributions of flow characteristics for
the same energy input. The model is not able k liquid phase
to discriminate between them. inlet Qi
k-l
3.2 Structured Models for Mass
!
t! kI
Transfer in Tower Reactors
Several attempts have been made to estab-
lish models of a bubble column reactor which
take account of the fluid dynamic structure of
the reactor. Such structured models consider
that different areas whithin the reactor may
have distinctive mass transfer characteristics
due to the flow in the region. Fig. 16. Back flow cell model for representation of a
DECKWERet al. (1978) conducted a very tower reactor (DECKWERet al., 1980).
comprehensive study of carbon dioxide trans-
fer in two bubble columns by measuring the
profiles of the absorbed gas along the col-
umns. They found that a constant value of the
mass transfer coefficient, while giving a satis-
factory interpretation of the overall mass
transfer rate in the tower, could not predict the
concentration profile. In order to achieve this
it was assumed that kLa was highest near the
gas sparger and then diminished along the tow-
er. The following profile was used:
I
, .
0.05 di 0.1s 0.2 i
- 2
The profile can be seen in Fig. 15. This ap- where j is the cell number and k denotes the
proach permitted a satisfactory fitting of the cell where the gas is injected into the system
experimental carbon dioxide profiles. DECK- and where most of the energy is dissipated
WER et al. (1978) recommended the value of (Fig. 17).
fb=2. LUTTMANet al. (1983b) found similarly
The spatial variation of kLa has also been that a constant value of the mass transfer coef-
considered by other authors, and has been ficient for oxygen absorption could not ex-
proven experimentally by ZHAOet al. (1988). plain the dissolved oxygen concentration pro-
A model of a bubble column based on back files during yeast cultivation. They explained
flow cells, which can be seen in Fig. 16, was this phenomenon as due to a strong change in
proposed by DECKWERet al. (1980). In order the specific interfacial area caused by in-
to explain the absorbed C 0 2 profiles satisfac- creased coalescence of bubbles near the gas
torily, the authors postulated the following sparger. The mass transfer coefficient was giv-
variation of the mass transfer coefficient as a en by:
function of cell number:
k,a(z, t ) =
kL= kLB[ 1 + aexp - b ( j - k2)] (88)
kl I
where KSTis defined as the coalescence factor,
k,a is the maximum value of the mass transfer
coefficient and is valid only at the gas inlet.
From the sparger up it decreases until z = a H ,
where it reaches the constant value
kLai(t)exp( -KsT). Fig. 18 shows how consid-
I
I ) I '
eration of the axial variations of the mass
k-4 k-2 k k:2 k+4 transfer rate improves the prediction of oxy-
Fig. 17. Dependence of the mass transfer coeffi- gen concentration profiles.
cient, kL,on cell number in the back flow cell model Still another aspect of this problem was
(DECKWER et al., 1980). considered by SCHUMPE and DECKWER
(1980). In their critical analysis of the chemical concentrations, temperatures, pH, etc. Recent-
method for the determination of interfacial ar- ly, it has been suggested (MERCHUK,1990)
eas, they consider the distribution of bubble that, for biochemical reactions, the shear
size and its effect on the overall mass transfer stress has an effect on the kinetics itself, there-
rate. The same idea had been presented earlier fore adding a new dimension to the problem,
by MERCHUKand RONCO(1966), who studied since the distribution of shear over the volume
the effect of the physical mass transfer rate on of the reactor would become important.
the enhancement factor, and concluded that
the existence of areas of different kinetic re-
gimes may originate from local variations of 4.1 Bubble Column Modelling
the physical mass transfer coefficient, kLa.
In airlift reactors, the different flow charac- Process modelling in bubble column reac-
teristics in each part of the reactor (riser, tors has been reviewed very clearly by DECK-
downcomer, and gas separator) indicate that a WER (1986). The classification proposed for
structured model should be used for the sys- the models is based on the assumptions of the
tem. Since the dependence of the mass transfer flow of the gas and the liquid phases. Fig. 19
coefficient on the fluid dynamics is well estab- shows the categories considered by DECKWER.
lished, it is obvious that the minimum desir-
able analysis of such a system would involve a
different kLa for each region. This argument
has not yet been tested experimentally, and all
of the mass transfer coefficients published are Gas phase
in fact based on the assumption of complete
mixing in the liquid phase.
MERCHUKand SIEGEL (1988) proposed a Liquid
method for the simultaneous determination of
the mass transfer rates in the three regions of PF X
an airlift reactor by measuring the transient I I I I
uid phase; thus, only the cases marked with an where the Stanton number for the gas phase at
X in Fig. 19 have been analyzed by DECKWER cell j is:
(1986).
The general procedure presented is as fol- StGj= ( k ~ j / N( 6) 6 p j / d s ) ( H / u ~ ~* )
lows. First, a balance is written for the main (RT/He) (92)
reactant, A, in the gas phase. For a completely
mixed liquid phase, this balance can be inte- and a is the coefficient for the linear variation
grated directly to give the axial profile of A in of hydrostatic pressure along the column:
the gas. This allows the calculation of the gas
absorbed per unit volume or the specific ab- a = - p g ( l -~)H/(NP,) (93)
sorption rate. This is a function of the concen-
tration of A in the bulk of the liquid. A bal- The function pj describes the spatial distribu-
ance for A over the liquid film allows the solu- tion of the gas holdup in the cell j , and N is the
tion of the problem, depending on the kinetic total number of cells, identical for both gas
regime. For this DECKWER(1986) used the so- and liquid phase.
lutions given by CHAUDARIand RAMACHAN- In the liquid phase, the balance for the case
DRAN (1980) based on the film model. In the of cocurrent flow gives the following:
case of spatial variation of concentrations in
the liquid phase, the procedure must be modi- (l+y)Cj-,-(1+2y+St,)Cj+yCj+,+
fied to take this into account.
Axial dispersion models for bubble columns
+ Stb q =o (94)
will not be discussed in this chapter. Interested where y is the back flow ratio, C, is the con-
readers are referred to DECKWER’Sexcellent centration of the considered component in the
work. liquid phase, stage j , and the two Stanton
An alternative way of presenting the mixing numbers are given by:
in a reactor is the use of series of completely
mixed cells. It is known that a series of com- StLj = ( k , / N )(64 p J N ) ( H / U L ) / d s (95)
pletely mixed cells approaches plug flow when
the number of cells is very large, assuming uni- +
Stb = StLjPB (1 aj)/He (96)
directional connections between successive
cells. A more complex behavior can be mod- Similarly, equations were derived for coun-
elled on the basis of this scheme, as in the back tercurrent gas-liquid flow. As explained in the
flow cell model proposed by DECKWERet al. mass transfer section, a variation of the mass
(1980), which can be seen in Fig. 16. An equal transfer coefficient was assumed by DECKWER
number of cells is assumed for gas and liquid, et al. (1980) in order to account for an in-
which are interconnected at each level via in- creased mass transfer rate near the gas sparger.
terfacial mass transfer. This does not imply The authors reported satisfactory fitting of
equal mixing characteristics in both phases, data on absorption of COz, where both the to-
because while the gas-side cells are connected tal mass transfer rate and the concentration
in such a way that convective mass transfer is profiles were monitored.
only possible upwards, the liquid-side cells Sometimes, the chemical process is of such
also allow back flow. A mass balance for the a nature that very simple hydrodynamic mod-
absorbing component A around the cell j in els can combine successfully with the kinetics.
the gas phase gives: This is the case with the model presented by
374 I1 Tower Reactor Models
20, they considered two separate gas phases: Fig. 21 shows the effect of tower height on
an entrained gas phase flowing downward, the dissolved oxygen profiles for several gas
flow rates, as predicted by the model for an
nj-' Y,-,-nj Y,-nT q- oxygen consumption rate of 0.5 mg L - ' s P 2
- k, aj [(Pj/He)Y,- Cj](1 - 4j) Vi= 0 (99) 02.It can be seen that an increase in tower
height would be beneficial at higher gas flow
and a gas phase flowing upward rates. The predicted dissolved oxygen profiles
are more homogeneous at lower tower heights.
njcl yJ!+ 1 +nTJ y.-n!
J J
yJ! - Perfectly mixed cells in series have also been
k Laj [(Pj/He)Yj - Cj](1 - 4j)V ,= 0 ( 100) used to model chemical processes in air lift
tower reactors. A rather elaborate model was
The mass balance for the liquid phase is simi- presented by WACHI and MORIKAWA(1986)
lar to Eq. (98), except that two mass transfer for the chlorination of ethylene in a boiling
terms appear, one for each gas phase. bubble column reactor. The fluid dynamic de-
The molar flow between phases can be ob- scription of the reactor is similar to the model
tained assuming that a fraction r from the inlet by DECKWERet al. (1980) (Fig. 16), but with
gas flow rate is entrained into the downcomer, the addition of a loop connecting the first and
and remembering that a fractionfr will reverse the last liquid phase cells (Fig. 22). This fol-
direction due to coalescence: lows the model proposed by MIYAUCHIand
VERMEULEN(1963). In addition, heat transfer
between gas and liquid phases is considered
9.0-
.
E
8.0-
7.0-
.-..0
2 6.0-
"
W
5.0-
c
W
52 4.0-
0
U
g 3.0-
0 Fig. 21. Profiles of oxygen
2 2.0 300 t h i n concentration predicted by
the model of Ho et al.
1.0- (1977) for several gas flow
rates and two different
0' ,
heights, and an oxygen con-
1 2 3 4 5 H 6 7 8 9 10 sumption rate of 0.5 mg
Stage number L-'s-' 0 2 .
The disengagement top section is considered along with the variation in gas holdup due to
a single perfectly mixed stage for both phases. the absorption of gaseous species, vaporiza-
The mass transfer coefficient was assumed to tion of solvent, and axial variations in both
vary due to changes in holdup in the down- temperature and pressure.
comer, but was regarded as constant all along The process consists of the addition of chlo-
the riser due to the balancing effects of gas ab- rine to ethylene to give 1,2-dichloroethane,
sorption and expansion because of pressure and simultaneous absorption and reaction of
changes. both gases is considered via both mass and
376 I 1 Tower Reactor Models
hI"1
4
1
- Pln I
UL l m l s l
'\\
a 0
4
0 1.0
D 1.9
TI /+I1
C'kll P( ;+I 1
IitaIUL aUi
\
0'
2
0
10-41
1b-j 2 4 6 81b-* 2 1 6 8Ib-l
Excess ethylene /Feed
Fig. 22. Backflow cell model for chlorination of Fig. 23. Predicted and experimental conversions of
ethylene in a boiling reactor (from WACHIand Mo- chlorine and ethylene in a boiling bubble column
RIKAWA, 1986). reactor (from WACHIand MORIKAWA, 1980).
liquid
t z
- -- 2
0 0 Fig. 24. Distributed parame-
ter model for an airlift tower
reactor for single-cell protein
B u l k flow Axial dispersion Bulk flow Axial dispersion production (MERCHUK et al.,
1980).
heat energy balances. The model was able to protein production in an airlift reactor. Con-
represent the conversions of ethylene and chlo- tinuous operation was considered. The oxygen
rine satisfactorily, as shown in Fig. 23. consumption was represented by a Monod-
A distributed parameter gas model has been type expression for biomass production and
used (MERCHUKet al., 1980; MERCHUKand substrate and oxygen consumption. Fig. 24
STEIN, 1981b) for the modelling of single-cell shows a mass balance in a differential of the
Process Models in Tower Reactors 377
height dz. Although equations were presented The reaction group RA is given by:
for axially dispersed flow of both phases, cal-
culations were done assuming plug flow.
The differential equations representing the
riser are as follows for the oxygen concentra- The gas separator was represented by a well-
tion in the gas phase: mixed stage, both in the gas and in the liquid
dY1 [
(1 - y , Yi)' (1 + a - a q )
phase. The gas holdup was related to the su-
perficial gas velocity via a momentum balance
--
du
- -StG
(1 - Y') l+CY
Y1 -Y2
]
(102)
as given in Eq. (47). Simulations were carried
out by integration of the system for the special
case of no gas entrainment in the downcomer.
and for the oxygen dissolved in the liquid This would be the most critical operation
phase: mode because of the danger of anoxia in the
downcomer.
Fig. 25 shows the profiles of oxygen concen-
tration along the airlift reactor. At q = 0.5, the
a = gp, ( 1 - 9 ) H / 2 P , (104) 25
-F
St, = L kLa/ JL CI
(106)
15
In Eq. (103), BA and SA are the dimension-
less concentrations of biomass and substrate.
The axial distribution of these concentrations
is given by:
10.
5. t
+ I)$( 9
Fig. 25. Concentration profiles in the gas and liquid
phases of an airlift reactor during single-cell protein
production in continuous operation for two tower
heights: -H = 12 m;---- H = 4 m (MERCHUK
and STEIN,1981b).
378 11 Tower Reactor Models
M = [(QLf C B f f c h - QL I CBLdfach)/fcyl/
5 Operation Policies
9r 9r 7 References
Fig. 27. Ratio of cyclic-operated to batch-operated
reactor, M , as a function of the fractional conver-
sion q,, for two values of the Stanton number St and BANKOFF,S. G. (1960), A variable density single
two values of m. - V,i"/ v,,, = 0.1 ; - -- - Vrni"/ fluid model for two-phase flow with particular
V,,,=O.S. reference to steam-water flow, ASME J. Heat
Trans. 82, 265-272.
BANERJEE,S., SCOTT,D. S, RHODES,E. (1970),
Studies on cocurrent gas liquid flow in helically
coiled tubes. 11. Theory and experiments with
reacting gas in the liquid phase, which is true and without chemical reaction, Can. J. Chem.
only for a slow reaction regime. Mathematical Eng. 48, 542-551.
expressions for the enhancement factor for the BARNEA,D., TAITEL,Y. (1986), Flow pattern tran-
different operating conditions were obtained sition in two-phase gas-liquid flow, in: Encyclo-
as a function of the Hatta number pedia of Fluid Mechanics (CHEREMISINOFF, N.
P., Ed.), Chap. 16, pp. 103-491, Houston: Gulf
Publ. Co.
BELLO,R. A. (1981), Ph. D. Thesis, University of
Waterloo, Ontario, Canada.
and the volume-to-interfacial area ratio, PI, BLENKE,H. (1979), Loop reactors, A h . Biochem.
Eng. 13, 121-215.
CALDERBANK, P. H. (1958), Physical rate processes
in industrial fermentation, Trans. Znst. Chem.
Several factors affect the processes in a Eng. 36, 443-449.
rather complicated way, which makes it diffi- CHAKRAVARTY, M., SINGH,H. D., BARUAH,J.
cult to predict the effect of changes in the pa- N., IYENGAR, M. S. (1974), Liquid velocity in a
rameters of the system and which stresses the gas-lift column, Indian Chem. Eng. 16, 17-22.
CHAUDARI, R. V. RAMACHANDRAN, P. A. (1980),
importance of model simulations. Three-phase slurry reactors, AZChE J. 26, 177-
HALLAILEand MERCHUK(1986) presented 201.
the required equations for the calculations of CHISTI,M. Y., MOO-YOUNG, M. (1988), Prediction
the productivity in a gas-liquid process under of liquid circulation velocity in airlift reactors
various operational conditions that allow the with biological media, J. Chem. Tech. Biotech-
selection of the optimal operation mode. nol. 42, 21 1-219.
380 I 1 Tower Reactor Models
CHISTI,M. Y., HALARD, B., MOO-YOUNG,M. JOSHI, J. B., PANDIT,A. B., RAO, R. (1986), in:
(1988), Liquid circulation in air-lift reactors, Encyclopedia of Fluid Mechanics (CHEREMISI-
Chem. Eng. Sci. 43, 451-457. NOFF, N. P., Ed.), Chap. 16, Houston: Gulf
CLARK,M., JONES,A. G. (1987), On the prediction Publ. Co.
of liquid circulation in a draft-tube column, KALISCHEWSKY, K., SCHOGERL,K. (1978), Deter-
Chem. Eng. Sci. 42, 378-378. mination of mass transfer coefficients of absorp-
CLARK,N. N., ATKINSON,C. M., FLEMMER,R. L. tion of oxygen and carbon dioxide in different
C. (1987), Turbulent circulation in bubble col- culture media, Ger. Chem. Eng. 1, 140-149.
umns, AZChE J. 33, 515-518. KAWASE,Y., MOO-YOUNG,M. (1986), Influence of
DANCKWERTS, P. V. (1970), Gas-Liquid Reactions, non-Newtonian flow behavior on mass transfer
New York: McGraw Hill. in bubble columns with and without draft tubes,
DECKWER,W. D. (1986), Design and simulation of Chem. Eng. Commun. 40, 67-83.
bubble column reactors, in: Chemical Reactor KAWASE, Y., HALARD, B., MOO-YOUNG, M.
Design and Technology (DE LASA, H. I., Ed.), (1987), Theoretical prediction of volumetric mass
pp. 41 1-461, NATO ASZ Series, Dordrecht-Bos- transfer coefficients in bubble columns for New-
ton-Lancester: Martinus Nijhoff. tonian and non-Newtonian fluids, Chem. Eng.
DECKWER,W. D., ADLER,I., ZAIDI,A. (1978), A Sci. 42, 1609-1617.
comprehensive study on COz interface mass KEITEL,G., ONKEN,U. (1981), Errors in determi-
transfer in vertical cocurrent and countercurrent nations of mass transfer in gas liquid dispersions,
gas-liquid flow, Can. J. Chem. Eng. 56, 43-55. Chem. Eng. Sci., 36, 1927-1981.
DECKWER,W. D., HALLENSLEBEN, J., POPOVIC, KOIDE,K., IWAMOTO,S., TAKASAKA, Y., MATSU-
M. (1980), Exclusion of gas sparger influence on RA, S., TAKASASHI, E., KIMURA,M., KUBOTA,
mass transfer in bubble columns, Can. J. Eng. H. (1984), Liquid circulation, gas holdup and
58, 190-197. pressure drop in a bubble column with draught
FREEDMAN,W., DAVIDSON,J. F. (1969), Holdup tube, J. Chem. Eng. Jpn. 17, 611-618.
and liquid circulation in bubble columns, Trans. KOIDE,K., KIMURA,M., NITTA, H., KAWABATA,
Znst. Chem. Eng. 47, T251-T262. H. (1988), Liquid circulation in bubble column
FORTESCUE,G. E., PEARSON, F. R. A. (1967), with draught tube, J. Chem. Eng. Jpn. 21, 393-
Chem. Eng. Sci. 22, 1163. 399.
GRIFFITH,P., SNYDER,C. A. (1964), MZT Report KUBOTA,H., HOSONO,Y., FUJIE, K. (1978), Char-
5003-29 (TID 20947). acteristic evaluation of ICI air-lift type deep shaft
HALLAILE,M., MERCHUK, J. C. (1986), Operation reactor, J. Chem. Eng. Jpn. 11, 319-325.
policies for a gas-liquid stirred tank reactor, LAMB, H . (1932), Hydrodynamics, Cambridge
Chem. Eng. Commun. 46, 179-196. (UK): Cambridge University Press.
HARMATHY,T. Z. (1960), Velocity of large drops LAMONT,J. C., SCOTT,D. S. (1970), An eddy cell
and bubbles in media of infinite or restricted ex- model of mass transfer into the surface of a tur-
tent, AZChE J. 6 , 281-288. bulent fluid, AZChE J. 16, 513-519.
HENSTOCK,W. H., HANRATTY,T. J. (1979), Gas LEE, C. H., GLASGOW,L. A., ERICKSON,L. E.,
absorption by a liquid layer flowing on the wall PATEL,S. A. (1986), Liquid circulation in airlift
of a pipe, AZChE J. 25, 122-131. fermentors, 8th AZChE Annual Meeting, Miami
HILLS,J. H. (1974), Radial non-uniformity of ve- Beach, November 2-7, paper 8.d.
locity and voidage in a bubble column, Trans. LEVICH,V. G. (1962), Physicochemical Hydrody-
Znst. Chem. Eng. 52, 1-9. namics, Engelwood Cliffs, N. J.: Prentice Hall.
Ho, C. S., ERICKSON,L. E., FAN, L. T. (1977), LEVY,S. (1960), Steam slip - Theoretical prediction
Modeling and simulation of oxygen transfer in from momentum model, Trans. ASME J. Hear
airlift fermentors, Biotechnol. Bioeng. 19, 1503- Transfer 82, 113-124.
1522. LUND, M. M., SEAGRAVE,R. C. (1971), Optimal
HSU, Y. C., DUDUKOVIC,M. P. (1980), Chem. operation of a variable volume stirred tank reac-
Eng. Sci. 35, 135. tor, AZChE J. 17, 30-37.
JONES,A. G. (1985), Liquid circulation in a draft- LUTTMAN,R., BUCHHOLZ, H., ZAKREZEWSKI, W.,
tube bubble column, Chem. Eng. Sci. 40, 449-462. SCHOGERL,K. (1982), Identification and mass-
JOSHI,J. B. (1980), Axial mixing in multiphase con- transfer parameters and process simulation of
tactors - A unified correlation, Trans. Znst. SCP production in airlift tower reactions with an
Chem. Eng. 58, 155-165. external loop, Biotechnol. Bioeng. 24, 817-835.
JOSHI,J. B., SHARMA,M . M. (1979), A circulation LUTTMAN,R., THOMA,M., BUCHHOLZ, H., SCHO-
cell model for bubble columns, Trans. Znst. GERL,K. (1983a), Model development, parame-
Chem. E m . 57. 244-251. ter identification, and simulation of the SCP pro-
References 381
duction process in air lift tower bioreactors with ROELS,J. A. (1983), Energetics and Kinetics in Bio-
external loop - I. Generalized distributed param- technology, Amsterdam: Elsevier Biomedical
eter model, Comput. Chem. Eng. 7 , 43-50. Press.
LUTTMAN,R., THOMA,M., BUCHHOLZ,H., SCHU- SAKATA,M., MIYAUCHI,T. (1980), Liquid recircu-
GERL, K. (1983 b), Model development, parame- lation in bubble columns, Kagaku Kogaku Ron-
ter identification, and simulation of SCP produc- bunshu 6 , 428.
tion processes in air lift tower reactors with exter- SCHLICHTING, H. (1968), Boundary Layer Theory,
nal loop - 11. Extended culture modelling and New York: McGraw Hill.
quasi-steady-state parameter identification, Com- SCHUMPE,A., DECKWER,W. D. (1980), Analysis
put. Chem. Eng. 7 , 51-63. of chemical methods for determination of inter-
MARUYAMA,T., YOSHIDA, S., MIZUSHINA,T. facial areas in gas-in-liquid dispersions with non-
(1981), The flow transition in a bubble column, uniform bubble sizes, Chem. Eng. Sci. 35, 2221-
J. Chem. Eng. Jpn. 14, 352-357. 2233.
MATSUMOTO,T., HIKADA,N., KAMIMURA,H., TANG, W. T., WISECARVER, K., FAN, L. (1987),
TSUCHIYA,M., SHIMIZU,T., MOROOKA,S. Dynamics of a draft tube gas-liquid-solid fluid-
(1988), Turbulent mixing-length model for axial ized bed bioreactor for phenol degradation,
turbulent diffusion of liquid in three-phase fluid- Chem. Eng. Sci. 42, 2123-2134.
ized bed, J. Chem. Eng. Jpn. 21, 256-261. TAYLOR,G. I. (1954), The dispersion of matter in
MERCHUK,J. C. (1986), Gas hold-up and liquid ve- turbulent flow in a pipe, Proc. R . SOC.London
locity in a two-dimensional air lift reactor, A223, 446-468.
Chem. Eng. Sci. 41, 11-16. UEYAMA,K., MIYAUCHI,T. (1976), Recirculation
MERCHUK,J. C. (1990), Shear effects in suspended in bubble columns, Kagaku Kogaku Ronbunshu
cells, Adv. Biochem. Eng. Biotechnol., submit- 2, 595-560.
ted. UEYAMA,K., MIYAUCHI,T (1977), Behavior of
MERCHUK,J. C., RONCO,J. J. (1966), Transfert de bubbles and liquid in a bubble column, Kagaku
masse avec reaction chimique simultanke, Genie Kogaku Ronbunshu 3, 19-27.
Chim. 96, 71-79. UEYAMA,K., MIYAUCHI,T. (1979), Properties of
MERCHUK,J. C., SIEGEL,M. H. (1988), Air-lift recirculating turbulent two phase flow in gas
reactors in chemical and biological technology, J. bubble columns, AZChE J. 25, 258-266.
Chem. Tech. Biotechnol. 41, 105-120. VAN DER AKKER,H. E. A., RIETEMA,K. (1979),
MERCHUK,J. C., STEIN,Y. (1981a), Local hold-up Detection of multiple generation patterns gener-
and liquid velocity in air-lift reactors, AZChE J. ated by the application of a positive temperature
27, 377-388. gradient, Trans. Znst. Chem. Eng. 57, 147-155.
MERCHUK,J. C., STEIN,Y. (1981 b), A distribution VAN DER AKKER,H. E. A.,RIETEMA,K. (1982),
parameter model for an air-lift fermentor. Ef- Correspondence on the paper by JOSHI-SHARMA,
fects of pressure, Biotechnol. Bioeng. 23, 1309- Trans. Znst. Chem. Eng. 60, 255-256.
1324. VAN DER LAAN, E. T (1958), Notes on the diffu-
MERCHUK,J. C., YUNGER,R. (1990), Chem. Eng. sion-type model for the longitudinal mixing in
Sci., in press. flow, Chem. Eng. Sci. 7, 187-191.
MERCHUK,J. C., STEIN,Y., MATELES,R. (1980), VERLAAN, P., TRAMPER, J., VAN’T RIET, K.,
A distributed parameter model for an air-lift fer- LUYBEN,K. C. A. M. (1986), A hydrodynamic
mentor, Biotechnol. Bioeng. 22, 1189-121 1. model for an airlift-loop bioreactor with external
MIYAUCHI,T., VERMEULEN, T. (1963), Diffusion loop, Chem. Eng. J. 33, B43-B53.R.
and back-flow models for two phase axial disper- VERLAAN,P., Vos, J. C., VAN’T RIET, K. (1988),
sion, Znd. Eng. Chem. Fund. 2, 304-310. From bubble column to airlift-loop reactor: Hy-
MIYAUCHI, T., FURUSAKI,S., MOROOKA,S., IKE- drodynamics and axial dispersion of the transi-
DA, Y. (1981), Transport phenomena and reac- tion flow regime, in: Bioreactor Fluid Dynamics
tion in fluidized catalyst beds, Adv. Chem. Eng. ZZ (KING, R., Ed.), pp. 259-275. London: Else-
11, 275-448. vier.
MURAKAMI,Y., HIROSHE,T., ONO, S . , NISHIJI- VONCKEN,R. M., HOLMES,D. B., DEN HARTOG,
MA, T. (1982), Mixing properties in loop reactor, H. W. (1964), Fluid flow in turbine stirred baf-
J. Chem. Eng. Jpn. 15, 121-125. fled tanks, Chem. Eng. Sci. 19, 209-213.
NIKLIN, D. J. (1962), Two-phase bubble flow, WACHI,S., MORIKAWA, H . (1986), Chlorination of
Chem. Eng. Sci. 17, 693-702. ethylene in a boiling bubble column reactor, J.
OHKI,Y., INOE, H. (1970), Longitudinal mixing of Chem. Eng. Jpn. 20, 238-245.
the liquid phase in bubble columns, Chem. Eng. WACHI,S., MORIKAWA,H., UEYAMA,K. (1987),
Sci. 25, 1-16. Gas holdup and axial dispersion in gas liquid
382 11 Tower Reactor Models
concurrent bubble column, J. Chem. Eng. Jpn. YANG, Z., RUSTEMEYER, V., BUCHHOLZ, R., ONK-
20, 309-316. EN, U. (1986), Profile of liquid flow in bubble
WALLIS,G. B. (1969), One-Dimensional Two-Phase columns, Chem. Eng. Commun. 49, 51-67.
Flow. New York: McGraw Hill. ZEHNER,P. (1986a), Momentum, heat and mass
WALTER,J. F., BLANCH,H. W. (1983), Liquid cir- transfer in bubble columns. Part 1 . Flow model
culation patterns and their effect on gas holdup of the bubble column and liquid velocities, Znt.
and axial mixing in bubble columns, Chem Eng. Chem. Eng. 26, 22-29.
Commun. 25, 243-262. ZEHNER,P. (1986b), Momentum, mass and heat
WARNECKE, H. J., PRUSS, J., LANGEMANN, H. transfer in bubble columns. Part 2. Axial blend-
(1985), On a mathematical model for loop reac- ing and heat transfer, Znt. Chem. Eng. 26, 29-
tors - I. Residence time distributions, moments 35.
and eigenvalues, Chem. Eng. Sci. 40, 2321- ZHAO,M., NIRANJAN, K., DAVIDSON, J. F. (1988),
2326. Mass transfer in short bubble columns, in: Bio-
WEHNER,J. F., WILHELM,R. H. (1956), Boundary reactor Fluid Dynamics ZZ (KING, R., Ed.), pp.
conditions of flow reactor, Chem. Eng. Sci. 6, 291-308, London: Elsevier.
89-93. ZUBER,N., HENCH, J. (1962), Steady State and
WISWANATHAN, K. (1986), Flow patterns in bubble Transient Void Fraction of Bubbling Systems and
columns, in: Encyclopedia of Fluid Mechanics Their Operating Limits, General Electric Co. Re-
(CHEREMISINOFF, N. P., Ed.). Chap. 38, pp. port, 626, Lido, N. J.
1180-1215. Houston: Gulf Publ. Co. ZUBER,M., FINDLAY,J. A. (1965), Average volu-
YAMAGOSHI,T. (1969), B. S. Thesis, Dept. of metric concentration in two-phase flow systems,
Chem. Eng., University of Tokyo, Japan. Trans. ASME J. Heat Transfer 87c, 453-468.
12 Process Models:
Optimization of Yeast Production
- A Case Study
KARL-HEINZBELLGARDT
Hannover, Federal Republic of Germany
JINGQIYUAN
Shanghai, People’s Republic of China
1 Introduction 385
2 The Baker’s Yeast Process 387
2.1 Development of the Yeast Process 387
2.2 Biological Properties of the Yeast 387
2.2.1 Metabolic Types of Yeast Growth 388
2.2.2 Regulatory Effects During Yeast Growth 388
2.2.3 The Cell Cycle of Yeast 389
2.3 Process Description 390
3 Process Model 390
3.1 Reactor Model 391
3.1.1 Liquid Phase Model 391
3.1.2 Gas Phase Model 392
3.1.3 Mass Transfer Model 393
3.2 Cell Model 394
3.2.1 Metabolic Model 394
3.2.2 Dynamic Regulation Model 395
3.2.3 Cell Cycle Model 396
3.3 Consideration of Upstream and Downstream Processing Steps in the Model 398
3.4 Verification of the Process Model 399
4 Application of the Process Model for Optimization 400
4.1 Goals for Optimization 400
4.2 Principal Strategy for Quality Optimization 401
4.3 Optimum Operation of Baker’s Yeast Production with Consideration of Product Quali-
ty 402
4.4 Influence of Operating Conditions on Optimum Operation 403
4.5 Influence of the Performance Criterion on Optimum Operation 404
5 References 406
384 12 Process Models: Optimization of Yeast Production - A Case Study
Introduction 385
volumetric ethanol reaction rate, I variable related to the inlet air flow
kg m-3 h-' L liquid phase of the reactor
volumetric oxygen uptake rate, h4 measured value
kg m-3 h-' max maximum value
volumetric substrate uptake rate, min minimum value
k g m - 3 h-' N nitrogen
volumetric growth rate of cell mass, 0 oxygen
kg m-' h-' R variable related to the substrate in-
respiratory quotient flow
-'
specific reaction rate, mol g h-' S substrate
vector of specific reaction rates, sat saturation value
mol g-' h - ' W water
total sugar concentration, kg m P 3 X cell mass
cell number doubling time, h 0 initial value at the beginning of the
discrete cycling age interval, h fermentation period
fermentation period, h
temperature of the inlet air flow, K
temperature of the liquid phase, K Superscripts
time for preparation and down-
stream processing, h T transpose of a matrix or vector
length of the process cycle, h * saturation concentration
time, h
dead time of measurements, h
duration of unbudded daughter
phase, h
duration of unbudded parent phase,
h 1 Introduction
volume of gas phase in fermentor,
m3
volume of liquid phase in fermen- A biotechnological process for the produc-
tor, m3 tion of a certain substance consists, in addition
volume of fermentor, m' to fermentation, of many elementary steps,
superficial gas velocity, m s - unit processes, and unit operations. A process
dry cell mass concentration, kg m-3 model should describe the most important
mole fraction steps of a biotechnological process. A general
cell number concentration, m - 3 process structure is shown schematically in
stoichiometric matrix Fig. 1.
state vector of the regulation model Production starts with the cleaning and ster-
vector of growth-independent specif- ilization of reactors and with preparation of
ic reaction rates, mol g-' h - ' substrate, media components, and inoculum
mean relative gas hold-up from the strain being maintained. Further-
density of water, kg m-3 more, process energy and compressed air have
specific growth rate, h - ' to be supplied. During fermentation the micro-
organisms are propagated in a series of reac-
tors. The batch or fed-batch mode of opera-
Subscripts tion is preferred for most industrial processes.
The reactors usually have an increasing vol-
C carbon dioxide ume, beginning with the smallest seed tank at a
E ethanol 10 L scale, up to the final production tank on
F value at the end of the fermentation a 10-100 m3 scale, where the largest amount of
period cell mass and product is formed. The fermen-
G gas phase of the reactor tation broth of one reactor is used as an inocu-
386 12 Process Models: Optimization of Yeast Production - A Case Study
*subslrale
R
Process energy
lum for the next stage. Between them there can plified balances may be sufficient as models
be additional processing steps such as filtra- for the calculation of the total conversion of
tion and storage. The final fermentation stage the process, e.g., yield of cell mass or product.
is followed by downstream processing steps, Greater modelling effort and a more detailed
separation of cells and media, disruption of description of the most important unit opera-
cells for release of intracellular products, pu- tions are generally required for a model-based
rification of the primary raw product, and ad- process optimization. One should be aware
ditional processing of the raw product to ob- that the result of such theoretical optimiza-
tain the final product for dispatch. The treat- tions may be significantly influenced by the
ment of waste is only indicated by one block in model accuracy, especially for the determina-
Fig. 1, but it is nevertheless an important part tion of an optimal dynamic control.
of the biotechnological process. Waste mate- To develop an accurate and complete model
rial arises from the cleaning of reactors and is not an easy task for several reasons. Some-
from downstream processing. The waste must times there are only very simplified models
be sterilized and the load of organic material available for parts of the plant, e.g., fermenta-
reduced. Depending on downstream proc- tion. Moreover, many interdependencies be-
essing it may be necessary to remove solvents tween the elementary units of the process are
or other reactants. often qualitatively and quantitatively un-
A process model should provide a coherent known. Fermentation is influenced in a com-
description of the entire process on the level of plicated manner by medium composition, sub-
plant operation. The degree of complexity of strate quality and preprocessing, and inoculum
the model or the possibility of simplified mod- preparation. The fermentation itself may in-
elling of some parts of the plant is determined fluence downstream processing by varying the
by the intended application of the model. The rheology of the fermentation broth and prod-
model may be used to answer several interest- uct properties. Model building is further com-
ing questions: What will be the output of plicated because for batch and fed-batch oper-
product per unit time for a given input of raw ation the process is never in a steady state and
material and primary energy? What are the the model must consider the process dynamics,
costs of production and of waste treatment? at least for the fermentation part. Therefore,
What is the optimum mode of operation for the modelling of the biological system is ob-
the reactors? Under what dynamic control of viously an important aspect of process models
manipulating variables is the product obtained for biotechnological systems.
with high productivity and the desired quality? To our knowledge no such comprehensive
How can the profit be maximized? Very sim- model has been published in the literature.
The Baker’s Yeast Process 387
Also in this chapter we intend only to deal technology of yeast production, accompanied
with certain aspects of an ideally complete by the development of microbiology. Follow-
process model. The baker’s yeast process will ing Pasteur’s discovery continuous aeration
be taken as an example in order to show the was applied (in England, about 1886), and the
philosophy behind process models. As an ad- method of fed-batch cultivation was intro-
ditional, often neglected aspect, a description duced (Germany and Denmark, 1915-1920) to
of the quality of the final product will be in- avoid the Crabtree effect. Further progress
cluded in the process model. The model will be was achieved by the use of pure cultures and
used to investigate strategies for optimal con- by the substitution for grain material of the
trol of the production process. Economic op- cheaper molasses in the 1920s. Since then the
timization requires a global view of intercon- process has been improved more in detail than
nected processing steps of the whole process in general layout. More recently attention has
and a detailed look at the fermentation steps. turned to automatic control of the process.
In the following sections the baker’s yeast One problem is to control the substrate supply
production process is briefly described. Then in a way that keeps the process at high produc-
biological properties of the yeast requiring tivity, but avoids loss of yield due to sugar
consideration in the model are discussed in or- over-feeding and ethanol production. In 1961
der to lead to our topic, the process model. the principle of automatic control of molasses
Based on this model the optimization goal is feed rate by the ethanol concentration in the
defined, and results of the subsequent optimi- exhaust gas was proposed by RUNGELDIER
zation are finally presented. The optimum and BRAUN. This control was first realized
control strategy is a compromise between the only in 1971 due to the lack of cheap and reli-
high productivity and yield of the fermenta- able sensors. Another method of automatic
tion part and good product quality after down- control of the molasses feed rate uses the respi-
stream processing. ratory quotient. Compared to control by etha-
nol concentration, this method requires more
exacting measuring devices and very accurate
gas-phase balancing to prevent stability prob-
lems. For further information on the history
2 The Baker’s Yeast and development of the yeast process the read-
er should refer to the literature (REED and
Process PEPPLER,1973; SKINNERet al., 1980; REED
and NAGODAWITHANA, 1990).
2.1 Development of the Yeast
Process 2.2 Biological Properties
Baker’s yeast production in submerged cul- of the Yeast
ture is one of the earliest biotechnological
processes. Its development began in the 19th Baker’s yeast Saccharomyces cerevisiae is a
century when a n increasing demand for yeast single-celled microfungus. Vegetative multipli-
in bread production could not be satisfied by cation by budding is most important for its
the limited availability of yeast from tradition- propagation in technical processes. For growth
al sources: beer and distillery processes. The in aqueous media the cells need carbon and ni-
reason for the shortage was mainly the change trogen sources, vitamins, and fatty acids. As a
to new strains for beer production, ones that eukaryotic organism the yeast has a wide
were unsuitable for baking. Initially, the bak- choice of metabolic pathways and regulatory
er’s yeast process was a copy of alcoholic fer- systems to react in a suitable manner to the
mentations, with very low productivity and availability of different substrates and oxygen.
yield - anaerobic batch cultivation on sub- Only subjects important for our modelling
strates from grains. In the second half of the goal will be mentioned here. Surveys of yeast
19th century there was great progress in the metabolism can be found in MILLSand KREBS
388 12 Process Models: Optimization of Yeast Production - A Case Study
(1986), COONEY (1981), ROSE and HARRI- yeast utilizes a complex regulatory system on
SON (1 987), and KOCKOVA-KRATOCHV~LOVA the epigenetic level by enzyme inductiodre-
(1990). pression and on the reaction level by enzyme
activation/inhibition. This regulation affects
the uptake of sugars, glycolysis, respiration,
2.2.1 Metabolic Types of Yeast glycerol metabolism, gluconeogenesis, and al-
cohol dehydrogenases. The most important ef-
Growth fects included in the model are described be-
low.
Oxygen supply is one of the main determin- The Pasteur effect, the suppression of fer-
ing factors of yeast metabolism and is the lim- mentative activity by respiration, was the ear-
iting parameter for the production process. In liest regulatory effect discovered in yeast. Its
excess of oxygen, yeast can achieve an oxida- connection with energy metabolism is well ac-
tive metabolism. The carbon source is com- cepted, but the biochemical details and their
pletely oxidized to carbon dioxide and water relation to observed growth kinetics and other
for energy production via the Krebs cycle and regulatory systems are still under discussion
the respiratory chain. These pathways can (LAGUNAS,1986). Enzymes of the respiratory
yield up to 36 mol of ATP per mol of glucose, pathways, which are repressed during fermen-
resulting in a very high yield for cell mass up tative growth, are induced under aerobic con-
to 0.5 g dry cell mass per g of sugar. Oxidative ditions. For the modelling of the Pasteur effect
growth is therefore the preferred metabolic it will be assumed that the cells always fully
type for baker’s yeast production. utilize their oxidative pathways up to maxi-
In the case of absence or lack of oxygen de- mum capacity. This is determined by the oxy-
privation, yeast changes to fermentative meta- gen supply or by the actual induction state of
bolism, and ethanol is produced as a final pro- key enzymes of the respiratory pathway. Ex-
ton acceptor for NADH. The chemical energy cess substrate is directed to fermentative meta-
comes mainly from glycolysis, with a yield of bolism.
2mol of ATP per mol of sugar. Due to this The Crabtree effect is referred to as respiro-
low efficiency compared to oxidative growth fermentative metabolism under an excess of
and due to the loss of carbon in the product oxygen and sugar. The Crabtree effect was ori-
ethanol, the cell yield is also very low. T o ob- ginally considered to be a repression of respi-
tain the same amount of cell mass as during ratory enzymes by glucose or glucose catabol-
oxidative growth much greater amounts of ites. However, the view of saturation of the
carbon source must be metabolized which is respiratory system is increasingly accepted and
accompanied by the production of a large vol- is also adopted for the model in this chapter.
ume of carbon dioxide. The result of the Crabtree effect is very similar
Yeast metabolism in the dough is also fer- to the inverse Pasteur effect. Both are very im-
mentative. The fermentative activity - the abil- portant for the control of baker’s yeast pro-
ity to produce a certain amount of C 0 2 in a duction, because an over-supply of sugar rela-
given time (usually 2 or 3 hours) - is, besides tive to a limited capacity of oxidative metabol-
storage properties and resulting flavor of the ism - either by internal saturation (Crabtree
dough, an important quality index for the effect) or by external oxygen supply (inverse
yeast product, because the produced carbon Pasteur effect) - leads to fermentative growth
dioxide raises and loosens the dough. with low yield of cell mass.
The glucose effect is the repression of up-
take systems or pathways needed for other
2.2.2 Regulatory Effects During substrates at high concentrations of glucose.
As a result of this repression glucose is the pre-
Yeast Growth ferred substrate for initial catabolism, even
when other substrates are present at high con-
For the directed coordination of its metabol- centrations. For baker’s yeast production the
ism linked to the substrate and oxygen supply, glucose effect with respect to ethanol, as a pos-
The Baker's Yeast Process 389
sible substrate for oxidative growth, is most cells (FBC)in the yeast population, as shown
important, because ethanol can be produced in Fig. 3. Since storage properties are also re-
during the fermentation. Ethanol and sugar lated to FBC (see our data in Fig. 3, where the
can be taken up in parallel; then both compete cells were stored at 3-5°C for 27 days), this
for the oxidative pathways. In the model it is single parameter can be used in a process mod-
assumed that ethanol can be used by the cells el to describe several quality aspects. A low
up to a rate determined by the remaining ca- FBC at the end of the final fermentation stage
pacity of respiration left by sugar catabolism. results in good product quality. There are oth-
The glucose effect with respect to maltose, the er parameters influencing the quality, such as
second main carbon source in molasses, need stability of the yeast cells on storage or os-
not to be considered for production strains mose-sensitivity for use of yeast in sweet
that possess a constitutive maltase system. doughs. The extension of the model to these
Such strains also have an increased fermenta- parameters remains a task for the future.
tive activity in dough. To establish the correct structure for the cell
model, including the cell-cycling model and
the metabolic model, one should know the in-
2.2.3 The Cell Cycle of Yeast teraction between metabolic processes and cell
division. Numerous investigations in this field
The cell cycle during yeast multiplication by have been summarized in WHEALS(1987). It
budding is a dynamic process of a series of or-
dered steps as shown schematically in Fig. 2:
growth of parent cells (G, phase), bud forma- B
ud
tion and bud growth (S, G2, and M phases) K m u
and separation of bud and parent cell (GT
phase). The period for the completion of the
cycle is different for parent and daughter cells.
A newly formed daughter cell needs a longer
time in the GI phase for the formation of its
first bud than for the subsequent buds. There-
fore, yeast populations are heterogeneous
mixed populations with a particular age distri-
bution of distinguishable daughter cells, par-
ent cells, and budding cells. The genealogical "P -
UI
B
u
Fig. 2. Graphical representation of the cell cycle of
age of a cell can be determined easily by count-
ing the number of bud scars on the cell sur- yeast.
face. For exponentially growing cultures these
data can be used to calculate the age distribu-
tion of a population and the duration of the
cell-cycle phases (LORD and WHEALS, 1980; . '9
THOMPSON and WHEALS,1980). The duration
of the GI phase is positively correlated with
cell-number doubling time. It is found, in con-
trast to the G, phase, that the duration of the
budded phase (S + G, + M + GT) is almost con-
stant and independent of growth conditions.
The dynamics of the cell cycle must be con-
sidered in the process model, since it is desir-
able to predict some aspects of the product 0 5 10 15 20 25 30 35 10
quality, which depends on the fermentative ac- Fraction of budding cells (%)
tivity and storage properties. TAKAMATSU et Fig. 3. Fermentative activity (0) (data of TAKAMAT-
al. (1985) have found that fermentative activi- zu et al. 1985) and storage properties (*) (own data)
ty is correlated with the fraction of budding depending on the fraction of budding cells, FBC.
390 12 Process Models: Optimization of Yeast Production - A Case Study
was found that most of the proteins, cell-wall cells are concentrated by centrifugation and
components, and RNA are synthesized at a stored in a separate tank for the pitching yeast.
constant rate during the cell cycle. The activity This storage gives additional flexibility in
of most enzymes is also not strongly modu- scheduling the final production stages, where
lated or else their concentrations are much two (stage 5 ) or four (stage 6) tanks are oper-
higher than needed for catalytic activity. The ated in parallel with some offset in inoculation
rate-limiting step of the cell cycle is obviously times. This procedure reduces the peaks in en-
the growth process, i.e., the increase of cell ergy and air consumption and leads to more
mass or volume, which is directly connected continuous downstream processing. The reac-
with the continuous processes of metabolism. tors in stage 5 and 6 are similar in size, ranging
The cells need to reach a certain mass or vol- from 60 m 3 to more than 200 m3.
ume to initiate the budding. The genetic pro- The profile of the molasses feed in these
gram for cell division can normally be exe- stages somewhat resembles the exponential-
cuted faster than the accumulation step. This linear scheme that will be discussed further in
also explains why the G,phase of the first gen- the following sections: in the first exponential
eration is longer than that of parent cells: the growth phase under an excess of oxygen a con-
daughter cells are smaller in volume. stant specific growth rate of about 0.25 h - ’ -
The following properties facilitate the mod- just below the critical value for the Crabtree
el building because they support the following effect - is set by an exponentially increasing
simplifying assumptions: substrate flow. In the second phase the growth
rate must be limited by an almost constant
0 Parent cells, daughter cells, and budding feed rate according to the maximum oxygen
cells have the same metabolic activity, transfer to the reactor in order to avoid oxygen
which is described by a unique specific limitation and ethanol production. After all
growth rate determined by substrate and substrate is fed to the production tank, the
oxygen supply. reactor is usually further aerated for about one
0 The cell-division cycle is controlled by or two hours without sugar feeding. Ethanol
the rate of increase of cell mass, the spe- that could have accumulated is thereby re-
cific growth rate. duced, and the cells are conditioned for good
0 There is no feedback from the cell-cy- storage properties. The F B C remains un-
cling model to the metabolic model. changed during this period. Therefore, a low
FBC for good quality must have already been
attained during the feed phase. In the down-
2.3 Process Description stream processing steps following fermenta-
tion, centrifugation and/or dehydration by ro-
In industry, baker’s yeast is produced in a tating vacuum filters, the final yeast product is
multi-stage reactor system. The general struc- obtained.
ture of a yeast process follows the layout in
Fig. 1 with some modifications. Fermentation
is typically carried out in a series of up to six
reactor stages. The first two are operated in
batch mode under sterile conditions, the others 3 Process Model
in fed-batch mode to avoid loss of yield. The
ethanol formed in the batch stages is used as a
substrate together with the sugar in stages 3 This section presents the dynamic process
and 4. The feed rate of the substrate molasses model that will be used to investigate several
must be adjusted accordingly in order to ob- optimum control strategies in the baker’s yeast
tain a full ethanol conversion at the end of process and to determine optimal operating
stage 4. Ethanol reduces the risk of infections; parameters. For optimization, in addition to
the accompanying disadvantage, loss of yield, productivity, yield, and profit, the quality of
can be kept low by the parallel conversion of the yeast product will be taken into account.
molasses and ethanol. After the fourth stage For simplicity, only the final production tank
Process Model 391
is considered here in detail, because it has dioxide, Cc, and liquid volume, VL,is derived
major influence on productivity and product from mass balances for the fed-batch reactor.
quality, and also because it uses a large per- The model equations
centage of the energy resources. A complete
model for the entire series of cultivations in
the multi-stage process can be built by mul-
tiplying the elementary model.
The model consists of a reactor model and a
cell model. Additional model equations are in-
troduced to consider upstream and down-
stream processing steps in a simplified form.
This is done by an additional dead time for the
whole production cycle and by calculating the
costs of the processing steps. To obtain signifi-
cant results for the description of yeast quali-
ty, the cell model must be quite complex, and
it will be explained in more detail.
According to the general structure of mod-
els for biotechnological processes shown in
Chapter 9, the presented model should de-
scribe both the reactor system, including gas
and liquid phases, and the biological system
yeast. This cell model includes a kinetic model
for the metabolic pathways, a dynamic model
for the main regulatory systems of metabol-
ism, and a model for the cell multiplication cy-
cle which permits prediction of product quali-
ty. include the accumulation in the liquid phase,
the biological reactions, and the mass trans-
port which can be divided into two parts. The
3.1 Reactor Model first term in every equation is due to the inflow
of substrate. The second term in the balances
The reactor model dynamically describes for oxygen, carbon dioxide, and ethanol
concentrations in the gas and liquid phases of comes from mass exchange with the gas phase.
the reactor as governed by initial conditions, The molasses flow rate, F, is the main manipu-
manipulating variables, and biological reac- lating variable of the reactor. It determines the
tions of the yeast cells. For simplicity it is as- increase in volume of the liquid phase and the
sumed that both phases are ideally mixed. related dilution effect for the process varia-
Therefore, the corresponding sub-models in- bles. The sugar concentration in the feed, S,,
clude only ordinary differential equations. A is an operating parameter. The reaction rates
third sub-model describes the mass exchange for cell growth, Rx, substrate and oxygen up-
between gas and liquid phases. The validity of take, Rs and Rot and ethanol and carbon
the model was verified for an industrial-scale dioxide production, RE and Rc, are given by
process, and unknown model parameters were the cell model in Sect. 3.2. The mass transfer
identified with data from production runs. rates for oxygen, carbon dioxide and ethanol,
OTR, CTR, and ETR, are determined by the
mass transfer model (Sect. 3.1.3). Water strip-
3.1.1 Liquid Phase Model off by the air flow is neglected in the mass bal-
ances of the liquid phase. The temperature is
The liquid phase model for the main compo- assumed to be constant.
nents cell mass, X , substrate molasses, S, etha-
nol, E , dissolved oxygen, C,, dissolved carbon
392 12 Process Models: Optimization of Yeast Production - A Case Study
3.1.2 Gas Phase Model The resulting outlet gas stream, FGO, is
usually not measured, but it can be calculated,
The main components of the gas phase are since the summation over all mole fractions
oxygen, carbon dioxide, nitrogen, ethanol, must equal one,
and water. They are advantageously described
by their mole fractions, xo, xc, xN, xE,and xw,
respectively. The model equations derived
from molar balances of the gas phase compo- and therefore:
nents are:
The mean Sauter diameter is ds = 0.008 m. The r = (rG,ro, rAc,rTc,rEp, rEc, rs, rx, r d T (30)
parameters of Eqs. (27) and (28) were re-esti-
mated for a 60 m 3 air-lift reactor. is the vector of specific molar reaction rates
which includes in this order the rates of glycer-
ol secretion, oxygen uptake, and acetyl-CoA
3.2 Cell Model production, the net-rate of the Krebs cycle, the
rates for ethanol production and consumption,
The cell model for the yeast Saccharomyces the rate of glycolysis, the growth rate, and the
cerevisiue determines the metabolic activity, carbon dioxide production rate. The vector of
expressed by the specific reaction rates, as a growth independent reaction rates:
Process Model 395
The metabolic model, Eqs. (29)-(31) and Eqs. are the maximum specific oxygen uptake rate,
(33) and (38), is a static one for the instanta- rOmax,and two fictitious enzymes E, and Ez.
neous adaption of cells to the actual supply of The matrix elements of F and f are non-linear
functions of model parameters and the specific
growth rate ,u((t) (BELLGARDTet al., 1986).
Tab. 1. Parameters in the Metabolic Model The model equation describing the induction
of gluconeogenesis for ethanol growth is:
Parameter Value Units
1 -
Fig. 5. Comparison of
experimental data from
LORD and WHEALS
(1980) for F B C (O),
.-c FDC (A), and FPC (*)
U of exponentially grow-
2
Y ing cultures with simu-
lation results (lines) by
the cell cycling model:
1 2 3 1 5 Dependency on the cell
Cell number doubling time T ( h ) number doubling time, T.
which define two connected cycles for parent The cell number is calculated as
cells, daughter cells, and budding cells. To
consider the smoothing effect of random fluc-
tuation of the generation time of single cells in XN= 2 d(i) + 3 PO) +
i= 1 j= 1 k= 1
b(k) (47)
the population, a moving-average filtering was
introduced in every simulation step for the The input variable, T, of the cell cycling model
cells in all cycling intervals. For exponential must be related to the specific growth rate giv-
growth the cycling process can be simulated by en by the metabolic model. Experimental data
model Eqs. (41) to (43). Fig. 5 shows a com- indicate that at relatively low growth rates,
parison of experimental data with this model. e.g., ~ ~ 0h -.' , 3 the mean cell volume of
In good agreement with the data, the simu- yeast cells undergoes only small changes. In
lated fraction of budding cells increases with this case the assumption is reasonable that cell
shorter doubling times. The fractions of cells number doubling time is equal to cell mass
in the three cycling phases are calculated by doubling time, and therefore:
summation over all age intervals. Thus, the
fraction of daughter cells is:
cells are relatively constant. However, for the and the costs for media preparation and stor-
unbudded daughter cells the cell volume was age, including the costs of water consumption
observed to vary significantly. This may re- and waste water treatment:
quire a non-uniform redistribution during the
transients of growth, a subject not further dis- PSp = K S 2 VLF (54)
cussed here.
where pMoand pso are the specific prices for
1 kg of pure molasses and media components.
The aeration rate is assumed constant. Thus,
3.3 Consideration of Upstream the costs are proportional to the fermentation
and Downstream Processing Steps period:
in the Model
PA,' FC TFPAe (55)
In this study major emphasis was put on the where pAeis the specific cost for production of
model for the fermentation part because this 1 m3 of compressed air.
governs the dynamics of the entire process. The downstream processing includes centri-
Nevertheless, for an economic optimization fugation and drying. The effort for the first is
the other parts of the process cannot be neglec- proportional to the liquid volume to be proc-
ted, although detailed modelling is not re- essed:
quired. With the optimization goal in mind it
is sufficient to estimate the approximate costs
for medium preparation, cleaning, inoculum
propagation, aeration, operation, and down- and for the latter proportional to the dried cell
stream processing. This model is presented mass:
next.
The fixed costs for operation, capital invest-
ment, and staff are proportional to the process
cycle: with the specific costs, P D D , for drying 1 kg of
yeast product. The total balance for down-
stream processing is:
where
and p E the specific maximum selling price for ethanol is produced due to the Crabtree effect.
1 kg of best quality dry yeast. The uptake of ethanol is suppressed by the glu-
cose effect until all sugar is used up. This be-
havior leads to the well-known growth-diauxy
3.4 Verification of the Process of yeast. Under fed-batch operation the Crab-
tree effect can be avoided by low feeding rates
Model of sugar at the beginning. Only after the last
step-increase of the substrate flow is there
Before the process model can be used in overfeed and ethanol production. In Fig. 6b
process optimization, it must be compared the cycling model is compared with experimen-
with experimental data to prove that it is suffi- tal data for FBC. Besides the tendency of
ciently accurate. Normally, not all model pa- FBC to increase with higher growth rates,
rameters are known in advance, so the un- some small oscillations in FBC can be ob-
known parameters must be identified from served. The latter property will be used later
suitable experiments. The full range of meta- for the optimum control strategy. The final
bolic types and growth conditions must be cov- value of FBC is about 20%.
0
Fig. 6. Combined batch and fed-
batch cultivation of baker's yeast in a
20 L reactor: Comparison of experi-
mental data (symbols) and simulation
results (lines) by the process model.
a) Variables of the reactor model:
Sugar feed rate F and concentrations
of cell mass X (o),substrate S (*),
and ethanol E (A).
b) Variables of the cell cycling mod-
el: Specific growth rate p, cell num-
=u 0 5 10 15 20 25 30 ber concentration X,, and fraction
u t (hl of budding cells FBC (v).
ered by these experiments and, therefore, the Besides this simulation a number of other
fermentation schedule may deviate from in- experiments from laboratory scale to produc-
dustrial production. As an example of model tion scale under different operating conditions
verification a single simulation by the process and feeding policies were simulated. The mod-
model of a combined batch and fed-batch ex- el was able to describe them all equally well
periment at laboratory scale is presented in with almost identical sets of parameters. This
Fig. 6. The batch-part lasts from 0 to 15 h and can be taken as a proof of the validity of the
the fed-batch part from 15 to 25 h. At the be- model, and the following optimization can be
ginning, the sugar concentration is high and expected to give significant results.
400 12 Process Models: Optimization of Yeast Production - A Case Study
The equation also considers by the factor Ks, chronize. As a result, oscillations in F B C can
the dilution effect through adding other media be observed. Because of these dynamics of the
components. cycling process, maximization of fermentative
activity of the final product under high pro-
ductivity may be realized by suitable dynamic
4.2 Principal Strategy for Quality control of the specific growth rate, ,u (DAIRA-
KU et al., 1982). The result of such a quality
Optimization optimization of a fed-batch baker's yeast culti-
vation with TF=lOh is shown in Fig. 7 . For
Prior to economic optimization with the this simulation only the cell-cycling model was
complete process model, which demands con- used. An upper limit of 0.25 h - ' was intro-
siderable computer time, the principle control duced for ,u to deal approximately with the
strategy for optimizing only the product quali- Crabtree effect. Under the conditions applied
ty will be investigated. Separate analysis of this here, the optimal control policy is in principle
parameter can assist in clarifying the results of to force synchronization of population growth
the economic optimization with its interdepen- by three short periods of increased growth
dent economic indices. rate. Pulse synchronization leads to strong os-
cillations in FBC. The cells are harvested in
the dynamic minimum of F B C at the end of
the fermentation. This minimum is much low-
0.Bl
er than the stationary value of about 20% for
- X 0.6
the averaged specific growth rate. In this way
w
productivity and quality can be decoupled to
c
2 some extent in order to optimize both.
5 0.1 For practical realization of quality control it
3
e is important to know the sensitivity of the op-
o
l
4.3 Optimum Operation of Baker’s mum strategy on the operating parameters and
economic indices is discussed in the following
Yeast Production with sections.
Consideration of Product Quality With the given economic indices the point
of optimal operation is obtained for a fer-
The optimization of the economic profit Eq. mentation period T,= 10 h, aeration rate
(60) with the process model was carried out as FG=90 m - 3 min-I, and substrate concentra-
explained in Sect. 4.1 under the constraints tion S, = 114 kg m-3. The simulation result
given in Eqs. (61) to (66). In this section the for a 60m3air-lift reactor is given in Fig. 8.
optimum flow rate profile is presented for an The time course of flow rate and fraction of
industrial production tank and a laboratory budding cells clearly resembles the pattern of
fermentor, while the dependence of the opti- the pure quality optimization, although the
reactor model and the metabolic model put The finally reached FBC of 2% is even lower
further constraints on the system. As will be than in the simulation. However, the good
analyzed in Sect. 4.5, the actual economic in- agreement of simulation and measurement
dices in the profit criterion lead to a slight pre- shows the validity of the presented model.
ference of productivity over yield. This gives
rise to some discrepancies to usual exponen-
tial-constant feeding schedules. Under this 4.4 Influence of Operating
control strategy, there is an over-feed of sugar
- even with some sugar accumulation - and Conditions on Optimum Operation
ethanol accumulation in the first half of the
fermentation. This increases the productivity The optimum mode of operation presented
greatly. The loss in yield can be kept low be- in the previous section represents only one
cause ethanol is consumed in the second half point in the multi-dimensional parameter
of the fermentation. Under the conditions space. To investigate the behavior of the opti-
shown in Fig. 8 ethanol metabolism is very ef- mal solution, its dependency on systematic
ficient because sugar is also fed. Therefore, variations of operating parameters can be
ethanol is mainly linked to the Krebs cycle for studied by simulation. Only partial results are
energy production and not to ATP-consuming presented. In Fig. 10 the profit after Eq. (60) is
gluconeogenesis. plotted over the fermentation interval and the
In Fig. 9 simulation results for a 20 L reac- aeration rate. The diagram was calculated as
tor are compared with experimental data. The follows. For every fixed pair of FG and TF op-
optimum profile of the flow rate and the liquid timization was carried out with all other pa-
phase concentrations are given in the upper rameters and with the substrate feed rate F(t),
part (a), and the variables of the cycling model and the resulting sub-optimal profit was plot-
in the lower part (b). Again, oscillations of ted. This means that the initial conditions, the
FBC are forced by pulses on the flow rate. inlet substrate concentration, and the feeding
Unlike in Figs. 7 and 8, the optimum policy policy are different for every point in the
here requires a double peak on the flow rate. FG-TF plane.
This change in the optimum scheme must be As expected, aeration rate has a strong in-
ascribed to operating parameters and process fluence on profit since available oxygen is one
dynamics determined by the reactor, which of the major limiting factors for productivity.
were not considered for optimum p-control. In principle, profit increases with increasing
al
M
aeration rate, but only to a certain extent when where the first summand is proportional to the
costs for compressed air or reduced quality be- productivity and the second is the yield. The
come dominant. For fermentation periods of total sugar consumption is:
less than ten hours profit diminishes because
of lower final yeast production. For longer fer-
mentation periods the sub-optimum is more
flat with respect to aeration rate because lower and the produced total cell mass is:
aeration rates can be compensated by a proper
feeding strategy, which reduces growth rate.
But profit decreases together with productivi-
ty. For an exact comparison of all results sum-
marized in Tab. 4, initial cell mass, X o . VLo,
amount of total sugar, rn,,,,, fermentation pe-
4.5 Influence of the Performance riod, TF, and the aeration rate were main-
tained at the same values as in Fig. 8. Profit is
Criterion on Optimum Operation always calculated by Eq. (60).
The simulation results of cases 1 to 3 are
In the previous sections optimum control of given in Fig. 11. If yield is the most important
baker’s yeast production was investigated with aspect of yeast production (cases 1 and 2), the
an economic criterion that included product exponential-linear feed profile is optimal be-
quality. The optimum policy for substrate feed cause it avoids oxygen limitation and ethanol
deviated greatly from usual flow profiles, production. The exponential increase at the
especially in the case of pure quality optimiza- beginning becomes more distinct in case 2 with
tion, as presented in Fig. 7. The optimum also slightly higher weight on productivity. But
depends on the relative value of the other eco- here also productivity is low because the
nomic indices entering the profit margin, e.g., growth rate must be maintained at relatively
productivity or the costs for substrate and en- low values to avoid loss of yield. It is interest-
ergy. For an investigation of these influences ing that in the last interval from nine to ten
on the feed schedule, the previous results in hours the flow rate is decreased in order to
Fig. 8 are contrasted with the optimum solu- convert the remaining ethanol.
tion for a modified simpler criterion: If emphasis is placed on high productivity
(case 3) the optimization results in a different
two-phase strategy. In the first phase there is
strong over-feeding with ethanol production
Tab. 4. Comparison of Yield, Productivity, and Profit for Different Optimization Criteria
and sugar accumulation, and the growth con- productivity in a similar range as before. Ob-
ditions correspond to batch operation. In the viously, quality can be improved with slight
second phase the feed rate is lowered to small additional effort and small cut-off on the oth-
values so that the ethanol can be completely er economic indices.
depleted with relatively small loss of total For completeness, pure batch fermentation
yield. This strategy is very similar to the com- data are given in case 6. This was simulated
bined batch/fed-batch scheme of DE LOFFRE, under the assumptions that the reactor is filled
proposed in 1941, which had been applied by completely at the beginning of the fermenta-
some baker's yeast factories until 1961. The re- tion and that n o ethanol or substrate inhibition
sult of economic optimization when disregard- occurs. Therefore, values for yield and pro-
ing product quality (case 4) is quite close to the ductivity are estimates for the best case, and
previous high-productivity case, case 3. This actual values would be lower. The initial sub-
indicates that with the given economic indices strate concentration
operation cost is relatively important com-
pared with the substrate price. Increase of
profit in economic optimization is mainly due
to improved product quality with yield and
406 12 Process Models: Optimization of Yeast Production - A Case Study
is chosen so that the total amount of converted viscosity bubble column reactors, Chem. Eng. J.
sugar is equal to that in the other cases. Pro- 28, 21-42.
ductivity can be compared with case 3, but the KOCKOVA-KRATOCHV~LOVA, A. (1990), Yeasts and
yield is much lower since in the first phase of Yeast-like Microorganisms, Weinheim: VCH.
LACUNAS,R. (1986), Misconceptions about the en-
diauxic growth metabolism is more fermenta- ergy metabolism of Saccharomyces cerevisiae,
tive than under fed-batch operation. The sec- Yeasts 2, 221-228.
ond diauxic growth phase is also more ineffi- LORD,P. G., WHEALS,A. E. (1980), Asymmetrical
cient because ethanol growth is not supported division in Saccharomyces cerevisiae, J. Bacteriol.
by additional sugar supply. High ethanol con- 142, 808-818.
centration lowers the yield further by a strong- REED, G., NAGODAWITHANA, W. T. (1990), Yeast
er stripping effect of the air flow. Technology, 2nd Ed., New York: Van Nostrand
Reinhold.
REED, G., PEPPLER,H. J. (1973), Yeast Technolo-
gy, Westport, Connecticut: The AVI Publishing
Company, Inc.
5 References ROSE, A. H., HARRISON, J. S. (Eds.) (1987), The
Yeasts, Vols 1 and 2. London: Academic Press.
RUNGELDIER, K., BRAUN,E. (1961), Method and
Device for Controlling and Growth of Microbial
BAILEY, J. E., OLLIS, D. F. (1983), Biochemical Cultures, U.S.Patent 3,002,894, October 3.
Engineering Fundamentals, New York: McGraw- SCHOGERL,K., LUCKE,J., OELS, U. (1977), Bub-
Hill. ble column bioreactors. Tower bioreactors with-
BELLGARDT, K. H., MEYER,H. D., KUHLMANN, out mechanical agitation, Adv. Biochem. Eng. 7,
W., SCHOGERL,K., THOMA,M. (1986), Appli- 1-84.
cation of an extended Kalman-filter for state esti- SKINNER,F. A., PASSMORE,S. M., DAVENPORT,
mation of a yeast fermentation, ZEE-Proceedings R. R. (Eds.) (1980), Biology and Activities of
133, NO. 5 , pp. 226-234. Yeast, London: Academic Press.
COONEY,C. L. (1981), Growth of microorganisms, TAKAMATSU,T., SHIOYA, S., CHIKATANI,H.
in Biotechnology (REHM, H.-J., REED, G., (1989, Comparison of simple population models
Eds.), Vol. 1, p. 76ff, Weinheim: Verlag Che- in baker’s yeast fed-batch culture, Chem. Eng.
mie. Sci. 40, 499-507.
DAIRAKU,K., IZUMOTO,E., MORIKAWA,H., THOMPSON, P. W., WHEALS,A. E. (1980), Asym-
SHIOYA,S., TAKAMATZU, T. (1982), Optimal metrical division in glucose limited chemostat
quality control of baker’s yeast fed-batch culture culture, J. Gen. Microbiol. 121, 401-409.
using population dynamics, Biotechnol. Bioeng. WHEALS,A. E. (1987), Biology of the cell cycle of
24, 2661-2674. yeast, in The Yeasts (ROSE, A. H., HARRISON,
HARTWELL, L. H., UNGER,M. W. (1977), Unequal J. S., Eds.), Vol. 1, pp. 283-390, London: Aca-
division in Saccharomyces cerevisiae and its ap- demic Press.
plication for control of cell division, J. Cell. YUAN,J. Q.,BELLGARDT, K. H., JIANG,W. S.,
Biol. 15, 422-435. DECKWER, W. D. (1991), A dynamic cell cycling
HEIJNEN,J. J . , VAN’TRIET, K. (1984), Mass trans- model for growth of baker’s yeast and its applica-
fer, mixing and heat transfer phenomena in low tion in profit optimization, Bioprocess Eng. 6 (6).
13 Aerobic Wastewater Process
Models
GRAHAME ANDREWS
Idaho Falls, Idaho 83415, U.S.A.
1 Introduction 409
1.1 What is in a Model? 409
1.1.1 Assumptions 409
1.1.2 Rate Equations 410
1.1.3 Yield Equations 41 1
1.1.4 Mass Balance Equations 415
1.2 Objectives and Advances in Modelling 416
1.2.1 Complexity versus Utility 416
1.2.2 Empirical and Mechanistic Models 417
2 Biofilm Processes 418
2.1 Biofilm Kinetics 418
2.2 Advances in Biofilm Modelling 421
2.2.1 Mathematical Simplification 421
2.2.2 Nitrification 422
2.2.3 Effects of Colloids 423
2.2.4 Steady-State Biofilm Thickness 423
2.2.5 Parameter Estimation 425
2.3 Biofilm Reactors 425
2.3.1 The Trickling Filter 426
2.3.2 Fluidized-Bed Bioreactors 427
2.3.3 Rotating Biological Contactors (RBC) 429
3 Suspended Growth Systems 430
3.1 Mass Transfer Resistance in Flocs 431
3.2 Mass Balance Equations for a Stirred Tank 431
3.3 The Contact Stabilization Process 433
3.4 Modelling Sludge Settling 435
4 References 437
408 13 Aerobic Wastewater Process Models
biomass. They are produced and consumed by ues. This exploration of the behavior predicted
eight different processes including nitrifica- under quasi-steady state, balanced growth con-
tion, denitrification, and cell decay. Actually ditions is a purely mathematical exercise that
there are nine processes, but the ninth, adsorp- should be applied to all structured models.
tion of colloidal organic matter onto the bio-
mass, is assumed to be very rapid compared to
the subsequent enzymatic hydrolysis of the ad- 1.1.3 Yield Equations
sorbed material, so there is no need to specify
a rate equation for it. This model is neatly A relationship between the yields found in
summarized in matrix form (Tab. 1) with one unstructured models can be found by perform-
column for each constituent and one row for ing element balances on Eq. (1):
each process. The rate equations for each
process are listed (as volumetric rates) down (4)
the right side of the matrix.
The form of the rate equations adopted for Ys=d, cell yield
the individual processes in a structured model carbon eauivalent of biomass Droduced
is obviously important and, to a certain extent, carbon equivalent of substrate consumed
arbitrary. The choice is usually based on a
scrutiny of the literature, experience with sim- d
ple systems, intuition about rate-determining Yo = -, oxygen yield
a
steps (as in the hydrolysis of adsorbed colloi-
dal material discussed above), and ultimately - carbon equivalent of biomass
-
the testing of the model against experimental mole O2 consumed
data. An additional, less used, test is to ana- Here P s = 4 + m - 3 n - 2 p is the number of
lyze mathematically the prediction of the set of available electrons per carbon equivalent of
rate equations under quasi-steady state, bal- substrate, a measure of the oxidation/reduc-
anced growth conditions. This can be illus- tion state of the substrate. P b is a similar quan-
trated by the simple structured model pro- tity for the organic fraction of the biomass,
posed by ANDREWS and TIEN(1977). The bio- and this analysis is useful only because Pb has
mass is divided into “active biomass” and been found to be surprisingly constant among
“stored substrate”, the latter lumping together microbial species, growth states, etc. (AN-
intracellular storage products, adsorbed colloi- D R E W ~1989).
, These quantities are related to
dal material, and some part of the extracellular COD values. PS is four times the number of
biopolymer. moles of O2 required for complete oxidation
substrate -
uptake
stored substrate -
growth
active + respiration
biomass/ products
(3)
maintenance
The rates and yields are shown in matrix of a carbon equivalent of substrate, or 1/8 the
form in Tab. 2. It can be shown (PADUKONE COD (in grams) of a carbon equivalent.
and ANDREWS,1989) that under balanced
growth conditions these equations reduce to M g biomass produced
-ys-
the basic unstructured model (Eqs. (2) and (9), 8(1 - x ) pS - YCOLJ g COD consumed ( 5 )
~
1 Aerobic growth 1
-_ 1 --1 - Y, - ixB
of heterotrophs YH YH
2 Anoxic growth 1
-_ -_1 -_
Y,
1 - ixB
of heterotrophs yrf 2.86 YH
3 Aerobic growth
of autotrophs
4 “Decay” of
heterotrophs 1-fp -1 fP
5 “Decay” of
fP
autotrophs
6 Ammonification of 1
soluble organic
nitrogen
7 “Hydrolysis” of
1 -I
entrapped organics
8 “Hydrolysis” of
entrapped organic
nitrogen
Stoichiometric
parameters:
Heterotrophic yield: Y,
Autotrophic yield: YA h
Fraction of biomass
yielding particulate
products: f p
Mass N h a s s COD in
biomass: ixB
Mass N/mass COD in
products from biomass:
ixp
Tab. 1. (Continued)
11 12 13
SND XND SALK ’
Process Rate, pj (mL - s - ’)
-isB
14
1
7Y.4
PA (KNH SNH) (A)
SNH
f K O , ~so
i-
xB,A
-1
1 -1
Kinetic parameters:
h h
Heterotrophic growth and decay:
PN, Ks, Ko,H, KNO?b H
h
Autotrophic growth and decay:
.-
I
PA, K N H Ko,A,
, ba
a
2 Correction factor for anoxic
e
0 growth of heterotrophs: np
E
v
Ammonification: k,
.-x
Y
Hydrolysis: kh, K x
a
Correction factor for anoxic
? hydrolysis: nh
414 13 Aerobic Wastewater Process Models
This analysis can be extended to actually diates and 0.34 for aromatic and aliphatic
calculating yield values by use of the YATp hy- acids (SERVIZIand BOGAN,1964). To account
pothesis, which states that producing a gram for the difference the yield equation in the ba-
of biomass requires the cell to expend a fixed sic unstructured model is written as:
amount of energy in the form of ATP. This is
superior to the attempt to predict yields based
on the free energy of formation of the sub-
strate, because there is no guarantee that mi- Y in this equation is a yield coefficient (based
croorganisms convert free energy into a form on COD, BOD, or TOC depending on the
they can use (ATP) with a constant efficiency. units of q), a parameter characteristic of the
In aerobic metabolism the production of ATP types of biomass and wastewater. The ob-
by substrate-level phosphorylation is usually served cell yield (biomass produced/substrate
small compared to that produced by oxidative consumed) in an actual reactor system is al-
phosphorylation. Thus the consumption of 1 ways less than Y, and depends not only on the
mole of O2 will produce 2 (P/O) moles of ATP values of Y and kd but also on the type of reac-
and 2 YATp(F‘/O) g of biomass. It follows tor and the biomass/substrate ratio it con-
that: tains. Finding accurate values of Y and kd
from observed yield data in batch cultures is
particularly difficult (ANDREWS,1984).
(7) The biomass decay parameter kd is a good
example of how several phenomena of second-
From Eqs. (4), (9,and (7): ary importance can be lumped together into a
single parameter with reasonable results. kd in-
YATP(P/O) corporates all of the following in an approxi-
YCOJJ = (8) mate way:
8(2+Pb(1 -x) YATP(F‘/O)/n/l)
The important feature of this equation is Maintenance. Microorganisms, like people,
that all the quantities on the right side are ap- need substrate not only for growth but also for
proximately constant. Taking the commonly staying alive; repairing spontaneous damage to
reported values YATp = 10 g/mol, (P/O)= 2.8, their macromolecules, producing energy for
Pb= 4.2, x= 0.08, M = 23 (ANDREWS,1989) activated transport of nutrients across cell
gives YcoD=0.52 g cells/g COD. In practice membranes, etc. In a true “maintenance mod-
YcoD is indeed found to be approximately con- el” the yield equation is identical to Eq. (9)
stant, but the observed yield is closer to 0.38 with kd replaced by k, Y where k, is the main-
for carbohydrates and Krebs cycle interme- tenance coefficient. However, the rate equa-
Introduction 415
tion is the Monod equation for the specific unit of “active heterotrophic biomass” con-
growth rate p: sumed (column 5 ; a negative sign indicates
consumption) produces f, units of “particulate
p=- iis products” (column 6), (1 -f,) units of “slowly
K+S degradable substrate” due to cell lysis (column
4), and (ix,B-f,i,,,) units of “particulate bio-
This model is identical to our “basic unstruc- degradable organic nitrogen” (column 12). The
tured model”, Eqs. (2) and (9), as long as units commonly used are COD equivalents.
S%Kk,/,ii which is usually true. The basic un- The total volumetric production rate of each
structured model is preferred because it works constituent is found by summing up a matrix
better at very low substrate concentrations. column
The maintenance model will not predict a
death phase in batch culture, and it tends to
predict negative substrate concentrations be-
cause the substrate uptake rate is still positive
(q = k,) when S = 0.
1.1.4 Mass Balance Equations
Endogenous metabolism. When no substrate is
available, a microorganism will start to cata- The rate and yield equations are characteris-
bolize its own protoplasm in order to obtain tic of the type of biomass and wastewater.
the energy required for maintenance. This en- They describe how the biomass reacts to
dogenous metabolism is well represented by changes in its environment (substrate composi-
the biomass decay term. tion and concentration, temperature, etc.) and
should be the same whatever type of reactor
Predation. The net effect of protozoa grazing the biomass is in. This is why parameter values
on the bacteria is that less total biomass is pro- determined in a batch culture, for example,
duced than would be expected from the yield should be applicable to an activated sludge
coefficient value and the amount of substrate plant if all the assumptions underlying the
consumed. This process is represented in a model are satisfied in both reactor systems.
semi-quantitative way by the decay term. The type of reactor is described in a model by
mass balance equations. A complete model
Cell lysis. The cell decay term in the basic un- consists of the rate and yield equations cou-
structured model can also describe the effects pled with mass conservation equations for
of cell lysis on the biomass. However, it can each constituent they involve.
not describe the other effects of lysis: release Writing mass balance equations requires as-
of protoplasm into the liquid where it becomes sumptions about the mixing conditions for the
substrate for other organisms and the residue liquid and solid phases in the reactor. The
of inert organic matter left in the biomass. equations are quite simple for plug flow (Sect.
These can only be properly described by a 2.3), pure diffusion (Sect. 2.1) and completely-
structured model that has a term “inert bio- mixed tanks (except when combined with
mass’’. This is one of the reasons why we structured kinetics models when the residence
should not expect the basic unstructured mod- time distribution of the biomass becomes im-
el to adequately describe systems like aerobic portant; Sect. 3.2). However, real reactors
sludge digestion where the cell decay processes rarely conform to these idealizations. More
are the primary, not the secondary, activity. complex reactor models such as a series of stir-
In a complex structured model a given con- red tanks with backflow (TERASHIMAand
stituent may be produced or consumed by sev- ISHIKAWA, 1985) give much better descriptions
eral different processes. The matrix presenta- of real, large-scale reactors.
tion again allows all the relevant yield values The failure to separate mass balance, yield,
to be presented in a compact form. Consider and rate equations and to clearly state the as-
for example the process “decay of hetero- sumptions on which each is based is an error
trophs” in the IAWPRC model (Tab. 1). Each that can create confusion. Use of the nomen-
416 13 Aerobic Wastewater Process Models
clature ( - dS/dt) for uptake rate of substrate plexity. The need for compromise was well de-
by biomass is one example of this. Another scribed in GLEICK’S(1987) book on non-linear
was demonstrated by the experimental obser- dynamics:
vation that cell yields in chemostats were al- “You can make your model more complex
ways lower than in batch culture. This caused and more faithful to reality or you can make it
debate (JAMES,1982) despite the fact that it is simpler and easier to handle. Only the most
an inevitable conclusion of the basic unstruc- naive scientist believes that the perfect model is
tured model (ANDREWS,1984). Attempts to one that perfectly represents reality. Such a
predict flocculation in activated sludge plants model would have the same drawbacks as a
from results on extracellular biopolymer pro- map as large and detailed as the city it repre-
duction in batch culture are another example. sents, a map depicting every park, every street,
A microorganism can produce large amounts every building, every tree, every pothole, every
of polysaccharides late in a batch culture be- inhabitant and every map. Were such a map
cause, in its recent past, it was exposed to high possible its specificity would defeat its pur-
concentrations of substrate. This is not true of pose: to generalize and abstract.”
an organism in a CMAS system. Mathemati- In recent years the wide availability of pow-
cally, the mass balance equations are com- erful personal computers has vastly increased
pletely different. The correct procedure would our ability to deal with difficult mathematics,
be to formulate a hypothesis about when bac- so the tendency has been to make fewer as-
teria produce biopolymer, construct a kinetic sumptions and produce more complex models.
model in which biomass is structured into cells This has allowed the modelling of the interre-
and biopolymer (an unstructured model can lated processes of soluble and particulate BOD
convey no information about the recent histo- removal, nitrification, and denitrification.
ry of a microorganism), solve the model for a Such models are now available for activated
batch culture, compare it with the data to de- sludge systems (HENZEet al., 1987; DOLDand
termine whether the hypothesis is correct and MARAIS, 1986), oxidation ditches (TERASHI-
what the parameter values should be, and fi- MA and ISHIKAWA, 1985), and biofilms (WAN-
nally to solve the model with the mass balance NER and GUJER, 1985; SEIGRISTand GUJER,
equations for a CMAS and see what it pre- 1987). Modern models also deal more realisti-
dicts. This has yet to be done, although some cally with liquid mixing in the reactor and are
semi-quantitative results are available (Sect. usually of the unsteady-state type to allow for
3.4). the diurnal variation in the flow and concen-
tration of wastewater (RITTMAN,1985). This
allows for theoretical investigation of different
1.2 Objectives and Advances control strategies for dealing with this varia-
tion (BARTONand MCKEOWN,1986). A start
in Modelling has been made on dropping the pseudo-single
solute assumption and modelling the removal
The basic models available for aerobic waste- of individual pollutants (WATKINand ECKEN-
water treatment systems can be found in sev- FELDER, 1984). The ultimate application of
eral textbooks (GRADYand LIM, 1980). The computers is the concept of “flexible modell-
objective of this chapter is to review more re- ing” in which the program itself “decides” on
cent developments and to show the directions the appropriate structure for the model from
in which modelling is moving and should move the information fed to it (SCHEFFERet al.,
in the future. 1985).
However, the goal of a single model capable
of optimizing the design and control strategies
1.2.1 Complexity versus Utility of all wastewater treatment systems has not yet
been achieved (JAMES, 1982), and GLEICK’S
In any field there is a direct relationship be- warning of the dangers of complexity should
tween the ability of a model to faithfully de- not be forgotten. “The goal of computing is
scribe real systems and its mathematical com- insight not numbers” (HAMMING,1973), and
Introduction 4 17
important insights that are immediately appar- Different models serve different purposes,
ent from the algebra of a simple model may be and the rule is to select the simplest model cap-
lost in the details of a computer program. A able of describing the system or phenomenon
good example is the application of the basic under study within the precision of the data
unstructured model, Eqs. (2) and (9), to the being collected. Unstructured models, strictly
mass balances for biomass over a CMAS sys- the quasi-steady state reductions of structured
tem, assuming no microbial activity in the set- models, will work for the CMAS system (and
tler and no biomass in the settler overflow. may avoid some numerical instabilities in com-
puter solutions of the complete structured
model) but not for contact stabilization. The
objective of modelling is also important. A
model designed for optimization of nitrogen
This equation shows that the outlet concen- removal in activated sludge systems (VAN
tration is independent of the inlet concentra- HAANDEL et al., 1982), for example, will nec-
tion and establishes the mean cell residence essarily over-simplify some of the factors that
time as the most important parameter in the are important in modelling reactor/settler in-
design and operation of activated sludge teractions (SHEINTUCH,1987). This is why
plants, two non-obvious and very important three models are used in this chapter. The ba-
pieces of information. sic unstructured model provides a baseline of
Another difficulty is that as the complexity expected behavior for a system. The IAWPRC
of a model increases so does the number of un- model (Tab. 1) represents the state of the art in
determined parameters, and the problem of modelling the removal of organics and nitro-
model discrimination becomes more severe. A gen by activated sludge systems. The simple
set of experimental data can be described structured model (Tab. 2) will be used for il-
equally well by more than one combination of lustrative purposes to provide insight into the
model structure, rate equations, and parame- form of rate equations (Sect. 1.1.2), floccula-
ter values. For example, in the early history of tion (Sect. 3.4), and problems arising from resi-
the contact stabilization system it was ob- dence time distribution of flocs (Sect. 3.2).
served that the MLVSS increased in the con-
tact tank and decreased in the stabilization
tank. This could be explained either by the
growth/decay hypothesis in which bacteria 1.2.2 Empirical and Mechanistic
grew in the contact tank and decayed in the
stabilization tank, or by the storage/metabol- Models
ism hypothesis in which substrate stored in the
biomass in the contact tank was metabolized in The “mechanistic” models described above
the stabilization tank. The mathematical de- start from descriptions of the basic physical
scription of either hypothesis required a struc- and biological phenomena in a process. Many
tured kinetic model; activehert biomass in models in the literature are not like this. They
the first case and active biomass/stored sub- are purely empirical, mathematical functions
strate in the second (note that JENKINSand that have been found to correlate experimental
ORHON, 1972, ignored the accumulation of data (JONES,1970). Empirical models give lit-
inert matter that accompanies biomass decay). tle insight into the effects of the system varia-
Both types of model could explain the MLVSS bles or how the system should be optimized or
data. The only way to choose between them scaled-up. This is best done by a combination
was to collect more detailed data after deciding of experiment and mechanistic modelling, us-
what specific experimental measurements cor- ing experimental data to validate the model,
respond to the abstract categories “stored sub- and then using the model to see how the sys-
strate” and “inert biomass”. Both hypotheses tem can be improved and thus what experi-
were found to be partly correct and modern ments should be done next.
models, following BUSBY and ANDREWS Consider, for example, the common empiri-
(1979, incorporate both types of structure. cal model for the variation of wastewater con-
418 13 Aerobic Wastewater Process Models
limiting
iclThick
film
I-$ nutrient
Other
nutrients
depth
Fig. 1. Concentration profiles in biofilms.
Biofilm Processes 4 19
(SEIGRIST and GUJER, 1987). For simpler KBoD + CBOD) = 0.6. The BOD has a larger ef-
processes, including the degradation of soluble fect on the rate even though oxygen is the lim-
organics, it is usually assumed that a single nu- iting nutrient. T o cover this possibility the
trient controls the metabolic rate in the bio- rate-controlling nutrient is defined as the one
film. Only one equation, Eq. (14), need then whose concentration first reaches the value
be solved but an exact algebraic solution is still that restricts the metabolic rate to 90% of its
impossible if q is given by the non-linear value under liquid-phase conditions. It can be
Monod equation (2). The usual approach has shown that this nutrient has the lowest value
been to find approximate solutions by taking of
linear approximations to the Monod equation
Ci+K )]
(Ci+10K
(ATKINSONand WILLIAMS, 1971; SHIEH et
al., 1981; ANDREWSand TRAPASSO,1985). [OYC,
j
The analysis given here, based on the work
of ANDREWS(1988a), represents a compro- Note that this is the same as the limiting nu-
mise between the accuracy of the first ap- trient if Ci%Kfor all nutrients, but otherwise
proach and the simplicity of the second. The they may be different.
first step is to identify the two nutrients that Eq. (14) for the rate-controlling nutrient
have the most influence on the metabolic rate. (which is given no subscript) can now be writ-
To d o this the qj values are eliminated using ten
the yield Eq. (9) and ( p + kd)is eliminated be-
tween the resulting equations, which can then
be integrated to give (17)
C=Ci a t z = O
dC
-=0 at z = L or when C=ECi
It follows that the limiting nutrient, the one dz
whose concentration reaches zero first in the E = 1 -(D YCi)I/(DYCi)
biofilm (Fig. l), is the one with the lowest val-
ue of (D YCi)j.The two most probable candi- The two forms of the second boundary con-
dates are organic matter (as BOD) and oxygen. dition correspond respectively to a thin film
Since YBoD/Yo= 1 by definition, diffusivities where the limiting nutrient is present through-
in biofilms are proportional to their values in out the film, and a thick film in which micro-
water. Since the diffusivity of O2 in water is bial activity is stopped deep inside the film by
2.5 x cm2/s and the diffusivities of sol- exhaustion of the limiting nutrient (Fig. 1).
uble organic compounds in water are close to The second form arises by setting C,= 0 in Eq.
1 x l o p 5 cm2/s, it follows that oxygen will be (15). The parameter E gives the relative impor-
limiting if Ci,BoD/Ci,o > 2.5 (approximately). tance of the limiting and rate-controlling nu-
For well-agitated, air-saturated water (C,,, = 8 trients in fixing the total metabolic rate in the
mg/L) this corresponds to Ci,,,>2O mg/L. biofilm. It varies between E=O when one nu-
However, the limiting nutrient is not neces- trient is both limiting and rate-controlling and
sarily the only one that influences the meta- E = 0.9 (a consequence of the definition of the
bolic rate in the biofilm. Consider the situation rate-controlling nutrient) when one nutrient is
when Ci,BoD=25mg/L, Ci,,=8 mg/L. Clear- not limiting but is very rate controlling (this
ly oxygen is limiting, yet at the point inside the happens if its Monod constant K is very
film where CBoD=7.5 mg/L, Eq. (15) gives large).
Co= 1 mg/L. Now if the Monod half-velocity H in Eq. (17) is the Heaviside function
constants are KBoD = 5 mg/L and KO= 0.1 mg/ which equals 1 if C,>O and equals 0 when
L (GRADYand LIM, 1980), the reduction in Cl=O. Mathematically it is the limit of the
metabolic rate due to oxygen restriction is q/ Monod function as K-tO, and it is what stops
4 = (Co/Ko + C,) = 0.9, while the correspond- metabolism in the film at the point at which
ing value for BOD restriction is (CBoD/ the limiting nutrient is exhausted.
420 13 Aerobic Wastewater Process Models
tanh Of
thin films: tanhe,< 1 -E2 Of = ~
Of
(20)
1 The intermediate lines in Fig. 2 were plotted
thick films: tanhBf> 1 -E2 qf = - from this formula. Interpolations of this type
Of
are clearly needed in the transition region be-
(b) Zero-order kinetics: tween thin and thick biofilms (i.e., e close to
Cj S K :0 = L (4x12D Ci) 1).
Biofilm Processes 421
I
I I
0 0.4 0,8 1.2 1,6 2 2.4
Dimensionless f i l m thickness 8 Fig. 2. Effectiveness factor.
The analysis given here is summarized in 1974). Adding more film beyond this thickness
terms of concentration profiles through the will give no increase in the substrate uptake
biofilm in Fig. 1. The interfacial concentra- rate.
tions (Ci and Ci,J may be less than the bulk-
liquid concentrations (Si and Si,Jdue to mass-
transfer limitations in the liquid phase. In a 2.2 Advances in Biofilm Modelling
very thin film (Fig. l a ) , essentially all of the
microorganisms are exposed to the interfacial Progress is needed and is being made in
concentrations, so O = 1 by definition (Eq. three areas: finding mathematical simplifica-
(19)). As the film gets thicker (Fig. l b ) , the tions to the analysis given in the previous sec-
rate of consumption of nutrients increases tion in order to facilitate its use in reactor de-
and, since this consumption is fed by diffu- sign; producing more complex models to pro-
sion, large concentration gradients appear in vide insight into important processes such as
the biofilm. For zero-order kinetics (Ci%.K), nitrification and sloughing; and finally proce-
the drop in concentration does not cause any dures for model verification and parameter
decrease in metabolic rate so that q still equals measurement for particular wastewaters.
one. For other inherent kinetics this biofilm
mass transfer resistance does cause a drop in
metabolic rate and q < 1. A truly mass-transfer 2.2.1 Mathematical Simplification
limited film is shown in Fig. 1 c. The film is so
thick that the limiting nutrient concentration Potentially the most useful simplication is
reaches zero, stopping metabolism. This can that proposed by KORNEGAY and ANDREWS
happen despite a liquid-phase concentration (1968) who showed experimentally that the
higher than that of the rate-controlling nu- substrate uptake rate per unit area of film
trient, either because it has a lower diffusivity could be correlated by
or because more of it is consumed by the film
(lower Y ) . The active depth of the film corre- S
N=gXL-
sponds to 0 = 1. It has been shown both by cal- KA+S
culation and by oxygen microprobe measure-
ments that for a n aerobic biofilm in contact KA is an apparent Monod constant whose val-
with air-saturated water this is a film thickness ue varies with film thickness. The equation is
of roughly 150 Fm (ATKINSONand FOWLER, clearly valid with KA = K (the inherent Monod
422 13 Aerobic Wastewater Process Models
constant) for very thin films (Fig. l a ) without would certainly fit within the normal scatter of
liquid-phase mass transfer resistance, because data from a pilot-scale trickling filter, for
all of the biomass is then exposed to a concen- example.
tration S of the rate-controlling nutrient. It The discontinuity in slope seen for Of=4
must also be true as S-’ co, because if S is high and Bf = 6 in Fig. 3 corresponds to O = 1, the
enough, all of the film will have adequate nu- point at which the film switches from being
trients to keep its metabolic rate at its maxi- “thick” (i.e., mass transfer limited; Fig. 1c) to
mum value 4. The question is whether this “thin”. The results for the non-mass transfer
equation interpolates well between these limit- limited, thin film sections (high S / K ) , can be
ing cases, and if so, how is KA related to the correlated by
film thickness, liquid-phase mass transfer
coefficients, etc.?
If the rate-controlling and limiting nutrients
are the same (E = 0: this assumption is inherent
in Eq. (23)) and liquid-phase mass transfer re- When working solely in the thick-film re-
sistance can be ignored, these questions can be gime (low S / K ) , the result is
answered as follows. For given values of film
thickness, Of, and nutrient concentration, S/K, KA
the effectiveness is calculated from Eq. (22), -=
K
ef
Ci,min
active depth. Second, kd is a critical parameter tivity in this region, but this is obviously only
in that small changes in its value can produce an assumption.
quite large changes in the steady-state thick- The type of activity depends critically on the
ness (see line for Ci,,). The shortage of reliable identity of the limiting nutrient. If oxygen is
values for k d , as a function of shear stress, limiting, carbon and nitrogen sources are still
wastewater type, etc., therefore, restricts the available, and anaerobic activity can continue.
use of this model. Third, there exists a sub- This can produce a variety of gases (N2, H2S,
strate concentration Ci,min = kdK/(ijx-kd) be- CH4) capable of dislodging the film. Also the
low which the “growth” and “decay” lines anaerobic bacteria will not necessarily produce
cross only at the origin, and the model predicts the extracellular biopolymer that holds a bio-
that no film will grow. This implies that for film together. If the organic matter is limiting,
many biofilm reactors (those without mixing then the metabolism in the region of interest
of the biofilm phase) there is a definite lower must be aerobic and endogenous. The mi-
limit to the effluent concentration of organics. crobes are in the same situation as in aerobic
This has been extensively studied in a series of digestion. They will first oxidize any intracel-
papers by RITTMAN(1982). The results are lular stored substrate, and then in order to
useful in a semi-quantitative way, although the generate energy to stay alive they will start to
Ci,min idea is clearly not strictly correct since hydrolyze the extracellular polysaccharide.
thin (monolayer) biofilms are found on rocks These changes are reflected in measurements
in mountain streams, a very high-shear, low- of the variation of biofilm density with thick-
substrate environment. This is probably due to ness, which show a maximum when the thick-
adsorption of substrates at the surface, and to ness equals the active depth (HOEHNand RAY,
strong microbe-surface interactions (as op- 1973; note that this casts doubt on the assump-
posed to the microbe-microbe interactions in a tion in Sect. 2.1 that diffusivities in a biofilm
true biofilm). are constant). However, attempts to build sim-
In low shear reactors including trickling fil- ple “sloughing criteria” related to the active
ters and rotating biological contactors (RBCs), depth into unstructured biofilm models have
the slow, continuous wash-off of microorgan- failed to explain the thick (order of mm) bio-
isms is less important. The biofilm grows until films found in practice (ATKINSONand WIL-
it can no longer support its own weight and LIAMS,1971). A consistent sloughing criterion
whole sections of it then “slough” off com- may only be found in a structured model that
pletely. The weakness that produces sloughing divides the film into “active biomass” and “ex-
occurs in a region not yet considered, the re- tracellular polysaccharide”. Such a model may
gion below the active depth. The theoretical also be able to predict the observed variation
model (Fig. 1c) was based on no microbial ac- in biofilm density.
Biofilm Processes 425
2.2.5 Parameter Estimation 80% of the water value is acceptable for most
purposes since this is not a sensitive parameter;
only its square root appears in the equations.
A device for the experimental measurement
of biofilm kinetics, validation of the analysis
of Sect. 2.1, and evaluation of parameter val- 2.3 Biofilm Reactors
ues for a particular wastewater, needs to be
carefully designed. It must provide a large sur-
face area of biofilm of measurable thickness The uptake rate expressions derived in the
all exposed to the same conditions, including previous sections are applicable to biofilms in
substrate concentration and shear stress. Pro- any type of reactor. In order to produce a
viding adequate aeration is another difficulty, complete mechanistic process model, they
and several researchers have chosen to work must be combined with mass balances for the
with denitrifying biofilms. various nutrients and expressions for the mass
The concentric-cylinder device (KORNEGAY transfer rates between the liquid and the bio-
and ANDREWS,1968) satisfies all of these re- film interface. These reactor-specific equations
quirements except measurement of the biofilm are studied in this section. Assuming plug flow
thickness. This device is essentially a Couette of the liquid phase, the steady-state mass con-
viscometer, and rotation of the inner cylinder servation for nutrient j in the reactor is
causes a constant shear over the entire biofilm
surface. JANSENand HARREMOES (1985) used
the device to study both aerobic and denitrify-
ing biofilms with methanol, acetic acid, and (28)
glucose as carbon sources. They confirmed the Sj=Sj,i at z=O
1/2-order kinetic model for thick biofilms.
Measuring the biofilm thickness is much easier The (kLa),term accounts for possible transfer
in the rotating-disc device used by SHIEHand of the nutrient from the gas phase into the liq-
MULCAHY(1985), which is fitted with small el- uid. The final term accounts for nutrient con-
ements that can be removed for microscopic sumption in the biofilm, which must equal the
examination. As with the concentric-cylinder rate at which the nutrient is transferred from
device, the liquid-phase mass transfer resist- the liquid to the biofilm interface:
ance can be eliminated by increasing the rota-
tional speed. However, the rotating disc has
the disadvantage that the shear stress on the
biofilm surface increases with radius. AN- Values of the mass transfer coefficients at
DREWS and TIEN (1981) used a fluidized bed the gas-liquid interface, k,,,,and the biofilm-
of coal coated with biofilm with liquid recycle liquid interface, kl,i, depend mainly on the lev-
to ensure constant substrate conditions in the el of turbulence in the liquid and vary consid-
bed. The average biofilm thickness could be erably between reactors. They are evaluated
inferred from the way the bed expanded as the from empirical correlations reviewed recently
biofilm developed, but there is n o guarantee by KISSEL(1986). In doing these calculations it
that the film thickness was the same through- is important not to confuse the nutrient that is
out the bed. limiting in the biofilm with the limiting nu-
Compared with models for dispersed-growth trient for the reactor, defined as the nutrient
systems, biofilm models require one extra pa- which will be exhausted first in the liquid
rameter, the diffusivity of the nutrient into the phase. The best example of this problem con-
biomass. Most of the experiments quoted in cerns oxygen. As long as the reactor is aerated,
this section give values in the range of 50 to the dissolved oxygen concentration can never
100% of the values in water, although some be zero (although it can be very small), thus,
lower (JANSENand HARREMOES,1985) and oxygen is not the limiting nutrient for the reac-
higher (OMUNA and OMURA, 1982) values tor. It can be, and often is, the limiting nu-
have been reported. Assuming a diffusivity of trient in the biofilm. This problem also extends
426 13 Aerobic Wastewater Process Models
to nutrients that are not supplied continuous- algebraically except when Ci is small enough so
ly. The yield equation (9) gives ( p + k d ) = that the logarithm term can be approximated
Yl q1= Y2q2= . . , . Replacing qj in Eq. (28) by two terms of its series expansion
with ( p + k d ) / Y j ,eliminating ( p + k d ) between
the equations for different nutrients and inte-
grating the result gives
Thick film, low BOD. The discussion in Sect. Thick film, high BOD. For S > 3 0 mg BOD/L
2.1 suggests that, given efficient oxygen trans- (approximately) Sect. 2.1 suggests that oxygen
fer, the organic matter is rate-controlling in becomes both limiting and rate-controlling in
the biofilm when S < 3 0 mg BOD/L (approxi- the biofilm. 6' is now defined in terms of dis-
mately). Even at these low concentrations, the solved oxygen concentrations and N can be
biofilm will grow past its active depth of ap- found only from N=No Yo/Ys (note that with
proximately 150 pm. We can, therefore, use organic matter measured as BOD, Yo/Ys= 1
the thick film asymptote for N (Eq. (26)) in by definition). In the oxygen mass balance
Eq. (29) which can be written equation the consumption by the biofilrn is
balanced by transfer from the gas phase, and
N = kL,i ( S - Ci) (32) the dissolved oxygen concentration changes
only slowly through the filter. The left side of
In principle, this equation can be solved for Eq. (28) can, therefore, be set to zero, and to-
the interfacial concentration, Ci, and the up- gether with Eq. (29) and the thick film asymp-
take rate, N. In practice this cannot be done tote the result is:
Biofilm Processes 427
ports of radius R coated with a biofilm of case due to curvature effects; ANDREWS
thickness L this is related to the biomass hold- (1988b) gives an exact analysis based on a dif-
up Eb, by: ferent definition of the effectiveness factor.)
The values y = 2 , 0 = 1 are a reasonable com-
promise between these considerations and hav-
ing unmanageably small particles. Growing
this much biofilm would cause the bed to ex-
y is a dimensionless particle radius, defined pand approximately 250% above the height of
like 0 but with L replaced by R . This is a gen- a bed of clean support particles. This is far
eralization that allows for curvature of the bio- more expansion than is allowed in practice.
film. For L & R it reduces to the “flat plate” A fluidized bed is more difficult to design
equation & b = a L used for the trickling filter than a packed bed, mainly because the support
(for spheres a = es3/R). particle and the liquid velocity cannot be set
Now consider what happens to the product independently. After fixing the particle radius
q & b in the substrate consumption term of the ( y = 2 ) and the particle density, the velocity is
mass balance Eq. (28) as the biofilm grows fixed by the need to fluidize the particles to the
thicker. For thin films q = 1 (Fig. 2) while &b required solids holdup. The height of the bed
increases so the consumption rate increases. is then fixed by the contact time needed to sa-
For very thick films (0sy ) cb= cs is a constant, tisfy the substrate removal requirement.
while q decreases (Fig. 2) so the consumption Choosing a dense particle produces a high liq-
rate decreases. Somewhere between these ex- uid velocity and a tall thin bed (and, inciden-
tremes there must be a n optimum biofilm tally, high mass transfer coefficients at the bio-
thickness. This is in contrast to the flat plate film interface), while a light particle produces
case (Fig. 4 ) where the consumption rate rises a short wide bed. This is illustrated in Fig. 5 ,
to some maximum value. Furthermore, if the the result of a design study for 90% removal
particle radius, y , is too large, q starts to de- of BOD from a waste stream of 10 m3/h con-
crease while &b is still small and the maximum taining 100 g BOD/m3 (ANDREWS,1988b).
consumption rate is low. This is a mathemati- Oxygen is provided by saturating a liquid recy-
cal statement of the obvious fact: to be effec- cle stream with pure oxygen. Repeating the
tive a bioreactor must contain a large amount study with air aeration produced very short,
of biomass, but if the biofilm is too thick some wide beds except with very small, dense sup-
of it becomes inactive due to mass transfer port particles. This type of calculation can
limitations. (Note that Fig. 2 is inexact in this eliminate many obviously unpromising designs
1000 -
-y=2l50 80meshl
\ ---Y= 5125 4OmeshI
100-
10 ,
Biofilm Processes 429
dS - d2Cj E dCj
t- = Si-S-a’N (39) Dj -- xq(cj)
=-
dt dz2 dt
The difficulty is that the uptake rate per unit where E = porosity of the biofilm. This equa-
area, N , is not constant over the disc. Some of tion must be solved twice for each nutrient,
the film is submerged in the trough, some is once for the time the biofilm is submerged,
just emerging from it with a thin film of liquid and once for the time it spends in the air.
attached to it, and some is just returning to it However, choosing the nutrients for which it
with the liquid film now thoroughly oxygen- must be solved would be difficult. It may be
ated after long exposure to the air. The value that the organic matter is limiting and/or rate
N is an average over these continuously vary- controlling over some parts of the disc while
ing conditions: oxygen becomes limiting in others.
430 13 Aerobic Wastewater Process Models
The other major problem concerns the hy- from this model identify the hydraulic loading
drodynamics of the liquid film. A common as- per unit area of biofilm ( l / ~ in’ Eq. (39)) as
sumption has been to ignore film drainage and the critical operating parameter and predict
centrifugal flow effects and to picture the film that, if this is held constant, the fractional re-
as a “slab” of liquid of constant thickness at- moval is independent of inlet substrate concen-
tached to the biofilm and not moving relative tration loading, flow rate and disc diameter.
to it. There then exists an equation (41) for These are identical to the low BOD results for
each nutrient in the liquid film, but with E = 1, trickling filters. Increases in removal caused by
Dj= diffusivity in water and, if microbial ac- increasing rotational speed, w, were smaller
tivity in the liquid phase can be ignored (an- than those observed in practice because the
other common assumption), X = 0. With model treats the liquid film thickness, 6, as an
z = distance from the biofilm interface and independent parameter. In practice 6 is an in-
6 = liquid film thickness the associated bound- creasing function of w , and this gives an extra
ary conditions are: improvement in removal. Biofilm thickness
was also shown to be a significant parameter,
t = 0 Cj = concentration in trough although no attempt was made to predict the
steady-state film thickness with a decay or
sloughing type of model. GRADY and LIM
(1980) offered a similar but more comprehen-
sive model that overcomes some of these ob-
jections and includes the effect of rotational
dC. speed on mass transfer coefficients in the
-2 = 0 (for other nutrients) trough.
dz
At the other extreme of high BOD concen-
After some time t=2(n-cos-’R1/r)/w the trations, oxygen will become the rate-controll-
liquid film returns to the trough and mixes ing nutrient in the biofilm. As in the trickling
with it. filter, the BOD removed will then be a con-
Even with this greatly simplified picture of stant ( Y o N o / Y )independent of inlet concen-
the hydrodynamics, the model requires the si- tration and dependant solely on the oxygen
multaneous solution of three equations of the transfer capacity of the system. This has been
form of Eq. (41) for each nutrient that may studied with a complex convection/diffusion
become rate-controlling. This has not been at- model by VAIDYAand PANGARKER,(1987).
tempted. Reactor design is based on purely This includes the liquid film thickness as a
empirical models, and this has led to disputes function of r, (Y and w. Several predictions are
about the relative significance of disc rpm, pe- made, including the interesting result that the
ripheral velocity, trough dissolved oxygen, total oxygen transfer reaches a maximum
etc., in scaling-up the reactor (CHESNERand when R 1 / R=0.25.
MOLOF, 1977).
The reasoning applied to limiting nutrients
in trickling filters in Sect. 2.3.1 applies equally
well to RBCs. At low BOD concentrations the
organic matter is the rate-controlling nutrient 3 Suspended Growth
in the biofilm and first-order kinetics give rea- Systems
sonable results. HANSFORD et al. (1978) con-
sidered this situation. They used all the as-
sumptions listed above, ignored concentration In modelling oxidation ponds, activated
gradients across the liquid film, and worked in sludge systems, etc., the mass transfer resist-
terms of a single substrate concentration (a ances discussed in the previous section are
function of tangential but not radial position) usually ignored. This great simplification al-
in the biofilm. This is valid for first-order ki- ;lows the characteristics of the wastewater and
netics only, because the active depth of the biomass to be described by much more de-
film is then independent of Ci. The results tailed yield and rate equations. Many sets of
Suspended Growth Systems 43 1
equations have been proposed, and the 3.2 Mass Balance Equations
IAWPRC model (Tab. 1) represents a consen-
sus of the important phenomena and how they
for a Stirred Tank
should be described mathematically (see also
SHEINTUCH,1987; SCHEFFERet al., 1985; BODEand SEYFRIED (1985) have studied the
DOLDand MARAIS,1986; DOLDet al., 1980; residence time distributions in activated sludge
BUSBYand ANDREWS,1975). systems. As in all real reactors they fall be-
tween the limits of plug-flow and complete
mixing. They are usually modelled as a series
3.1 Mass Transfer Resistance of stirred tanks (FIJIEet al., 1988), the number
in Flocs of tanks being somewhere between 1 for com-
plete mixing and infinity for plug flow. This
model is chosen for its apparent simplicity.
Most of the flocs in an activated sludge sys- The unsteady-state mass balance for constit-
tem have radii of 50-100 pm. How justified is uent j over a completely-mixed tank is
the assumption that the effectiveness factor
equals one? Repeating the calculations from
Sect. 2 in spherical coordinates shows that it is (43)
approximately valid if the liquid is well-
agitated and B (based on floc radius) There is one such equation for each constit-
<p(ANDREWS,1988a). The corresponding uent in the model (active biomass, colloidal
floc size depends on the identity and concen- substrate, etc.). r is the nominal hydraulic resi-
tration of the limiting nutrient. When oxygen dence time and rj is the volumetric rate of con-
is rate-controlling, D = 2 x lo-’ cm2/s (80% sumption of component j. For an unstructured
of the free water value), the liquid is saturated model, rj is proportional to the biomass con-
so that So= 8 mg/L, which is much larger than centration and dependent in some way on the
KO,the floc density X = 7 0 mg dry wt per cm3 component concentrations Sj. The set of Eqs.
(HOEHNand RAY,1973), and ijo=0.2 g O2per (43) can always be solved, although not neces-
g dry wt per hour, hence the limiting floc sarily algebraically. For a structured model the
radius is R = (6DS0/ij0X)”2 = 157 pm. Mass situation is far more complex. rj now also de-
transfer limitations are, therefore, insignifi- pends on the composition of a floc particle,
cant. However, taking lower but still accepta- and the flocs in a reactor do not all have the
ble values of D and So or higher but still ac- same composition. How can this be accounted
ceptable values of ijo and X produces the op- for?
posite conclusion. MIKESELL(1984), for exam- Consider what happens to a floc that falls at
ple, calculated average floc sizes from the set- time t = 0 into a stirred tank operating at steady
tling velocity data of BISOGNIand LAWRENCE state. For the purpose of illustration a simple
(197 1) and then calculated effectiveness factors structured model will be used which divides the
(with D = 10% of free-water value) as low as floc into mass mA of active biomass and ms of
0.05. stored substrate. If the floc breaks up or bits
The safest conclusion is that intra-floc dif- are washed off, these masses include all of the
fusion limitations are sometimes significant in rehlting fragments. The mass conservation
activated sludge systems. The results of Sect. equations for the floc are:
2.2.1 are very relevant here. They show that
the effect of these diffusional limitations is not active biomass: - dmA
- pm,
to invalidate the form of the Monod function dt
but to increase the apparent value of the half-
velocity constant KA. This is certainly one rea-
son for the wide range of published values for
this parameter (SHIEH, 1980).
The rates of substrate uptake, q, and growth,
p , depend on the substrate concentration in the
432 13 Aerobic Wastewater Process Models
tank, S, which is constant, and the composi- In the simplest possible case the influent of
tion of the floc particle, which varies with the a stirred tank contains a number concentration
time t it spends in the reactor. If the composi- of n flocdliter, all of which are identical in
tion is described by the ratio of stored sub- size and composition. It will be necessary to
strate to active biomass y = mS/rnA,its varia- assume that if a floc breaks up due to shear or
tion is given by: the wash-off cells, then all the fragments leave
the reactor at the same time. Unless yi = y , as
in a CMAS system (in which case growth is
"balanced" and the structured model is unne-
(45) cessary), the flocs in the reactor are not all the
same. Some will just have entered the reactor
This may be difficult to solve mathematically and therefore have a size and composition sim-
(depending on how q and p vary with y), but ilar to the entering flocs (mA=mA,i, y=yi),
what happens physically is familiar. The floc while others will have been in the reactor for
enters with a certain composition, yi, and there several days and will be much larger and have
is a lag during which the floc becomes accli- a composition y =y,. The fraction of the flocs
mated to its new environment. Mathematical- that have been in the reactor for a time be-
ly, acclimation means that the floc composi- tween t and (t+dt) is given by the residence
tion, y , reaches a value that depends on its en- time distribution function as e-'/'dt/r. Sev-
vironment, S, and thereafter, since S is con- eral important quantities must be defined by
stant, dy/dt=O. This is the quasi-steady state integrating over all the flocs in the reactor.
discussed in Sect. 1.1.2.
For the simple, linear rate equations used in active biomass n w
the model of Tab. 2, the solution to Eq. (45) concentration X A = - mAe-'/'dt
5 0
is:
(47)
stored substrate n "
concentration Xs=- S mse- '''dt
5 0
[(1 +-(k,S-km))
1/ 2
volumetric substrate n m
ym=B 4 -I] (46) consumption rate rs = ;S mAqe-'/'dt
2 k2B2
K+S (48)
B=- K = - k2
- Ym - equivalent volumetric microbial n m
KY kl Monod constant growth rate r A = - j mApe-'/'dt
5 0
This shows that the volumetric reaction rates one tank and the entire system. The former ap-
rs and rA can be found by substituting the av- proach is simpler because the equations are
erage floc composition yAV= X s / X A into the identical for each tank; only the input parame-
rate equations from Tab. 2 . It is important to ters are different. For the model of Tab. 2,
realize that this is valid only because the rate three Eqs. (43) are needed for a tank, one for
equations in Tab. 2 are linear functions of the substrate, one for active biomass, and one for
biomass composition y . Repeating the steps of stored substrate. Combined with Eqs. (49) the
Eq. (49) for the IAWPRC model (Tab. 1 ) steady-state solution is
would not give this result, because the rate of
process 7 “hydrolysis of entrapped organics” is
a non-linear (Monod-type) function of bio-
mass composition (Xs/XBA). The correct pro-
cedure in this case would be to solve Eq. (45)
to find how composition varies with the resi-
dence time of the floc (the equivalent of Eq.
(46)). Substituting this function into the rate
equations allows Eq. (48) to be integrated to
give the exact result for the volumetric rates.
In summary, there may be no alternative to Note that since B is a function of S, considera-
assuming that the reaction rates in a CSTR ble computation is needed to obtain results in
(continuous stirred tank reactor) can be based a useful form. A complete model would con-
on the average biomass composition in the sist of three such equations for the contact
reactor. The practical situation in which the tank (subscript c), three for the stabilization
reactor influent contains a range of floc sizes tank (subscript s), the obvious constraints on
and compositions, and inert matter is added to biomass composition yi,, =yAV,,, =yAV,,
the biomass structure (thus requiring two pa- and substrate composition Si,,= S,, and finally
rameters to define “biomass composition”) the mass balance equations for mixing the re-
could probably not be analyzed without this cycle stream with the influent wastewater:
assumption. However, it is exact only when
the rate equations are linear functions of floc +
Si,,( 1 + RR) = Si?,, RR S,
composition. For other models it introduces (51)
an unknown amount of error. i , c ( 1 + RR) = RR XA,
XA, s
a) C)
10
Y'
08
06
04
02
0 ,
01 02 03 04 a5
b) d)
tact tank in a contact stabilization system may as the main design and control parameter for a
only be one tenth this size (t’=0.05), and at contact stabilization system is based solely on
these low values aerating the return sludge extrapolation from conventional systems.
does make a difference. If there is no stabiliza- While this type of extrapolation can provide
tion tank and negligible metabolic activity in useful input into semi-empirical models (AL-
the settler then we require that ~ ~ ’ , ~ =Fig.
yc‘. EXANDER et al., 1980), there are obvious dan-
6 b shows that for t‘=0.5 this happens with gers in extrapolating results from the bal-
&=y: =4.0. Fig. 6 a gives the corresponding anced-growth CMAS system to the unbal-
removal as 82%, while sludge stabilization anced-growth contact stabilization system. The
that reduced yi,cto zero would give 88% re- fundamental mechanisms are different, and lit-
moval. In the storage/metabolism hypothesis tle basic insight can be gained. Truly mech-
considered here, this improved removal is due anistic models of the contact stabilization
to the ability of stabilized biomass to take up process, like the simple example given above,
substrate rapidly and store it as adsorbed col- may produce more optimal control strategies
loidal material or intracellular storage prod- based on other process parameters. It may, for
ucts. example, be preferable to keep the recycle flow
The substrate concentration and biomass (and thus 5,) constant and adjust sludge wast-
composition at the inlet to the stabilization age so as to maximize the sludge concentration
tank are the same as those at the outlet of the from the settler (XA,i,s),
whatever the resulting
contact tank. Thus if &=O, Si’,,=3.6 (Fig. 6, value. The impact of this type of strategy on
6a) and yi’,s=0.12(Fig. 6b). The inlet sludge the sludge settling behavior would obviously
concentration XL,i,sis fixed by how well the have to be considered in a comprehensive
sludge settles, and the value of 375 used in model.
Figs. 6c and 6d is fairly typical. The little sub-
strate present is rapidly consumed in the high
biomass/substrate conditions found in the sta-
bilization tank (Fig. 6c). More importantly, it 3.4 Modelling Sludge Settling
can be seen by interpolating in Fig. 6d that,
for yi:,= 0.12, the stabilization tank requires a Activated sludge systems are usually conser-
residence time of approximately 5 ’ = 0.3 (3 h) vatively designed from the point of view of
in order to reduce y’ back to zero before the substrate removal. Operational difficulties
biomass is returned to the contact tank. tend to arise not from low dissolved oxygen
levels or high BOD transients in the effluent,
Repeating this type of calculation for differ-
ent values of parameters XL,i,s,yi:c, etc., can but from difficulties with sludge separation in
suggest experiments to validate the model and the secondary settler. The need to incorporate
procedures for optimizing the design and oper- settler behavior into a comprehensive process
ation of real systems. A notable feature of the model has been understood for many years.
model is the absence of the system mean cell BUSBYand ANDREWS (1975) presented an ex-
residence time. It can be related to the model cellent model and used it to evaluate various
parameters by: control strategies based on the height of the
sludge blanket. However, settler performance
ex = total biomass in system - was described by a purely empirical correlation
wastage rate between the TSS in the effluent and the
MLVSS in the influent, which could not pre-
dict changes in settling behaviour due to bio-
mass composition, sludge age, etc. There are
several difficulties in producing a purely mech-
anistic model of settling.
The settling velocity of a floc particle de-
6, does not appear as a central feature of this pends on three things: its size, its density, and
model in the way it does for the conventional its sphericity. DA HONGand GANCZARCZYK
CMAS process (Eq. (12)). In fact, the use of 6, (1988) measured the settling velocity of flocs
436 13 Aerobic Wastewater Process Models
of 100-700 pm and found that Given sufficient time and the right conditions,
I mainly excess carbon source, bacteria will then
U (mm/s)=0.35+ 1.77 d (mm) (53) go on to accumulate extracellular polymers as
a long-term storage product. These different
Since flocs fall in the Stokes regimes (MIKE- forms of stored substrate are consumed in the
SELL,1984) we would normally expect U to be same order. Adsorbed colloids are hydrolyzed
proportional to d2. This suggests that the and used continuously. When organic matter
larger flocs were less dense, a result partly sup- is no longer available from the liquid, the or-
ported by the measurements of porosity. (The ganisms start to metabolize PHB and similar
effect of sphericity is unpredictable; a disc, for intracellular storage products. Eventually,
example, has a lower sphericity than a sphere over the time scales involved in aerobic diges-
but falls considerably faster edgewise.) How tion, this runs out, and the only energy source
can any of these factors be predicted by a available to the microorganisms is the extracel-
model? lular polymer. The results of PAVONIet al.
Predicting the size of a floc is a similar (1972) show that these polymers are normally
problem to predicting sloughing and the steady- resistant to hydrolysis, but that “old” cultures
state thickness of a biofilm (Sect. 2.4). Floc become acclimated to them which results in
size results from an equilibrium between the floc dispersion.
growth of biomass in the floc, biomass decay Incorporating this information into a model
processes, and the wash-off of cells from the would strictly require the biomass to be struc-
floc surface and possible fracture of flocs. The tured into protoplasm and three categories of
latter processes are caused by shear in the mix- stored substrate, adsorbed colloids, intracellu-
ing and aeration devices, but a floc’s suscepti- lar storage products, and extracellular biopoly-
bility to shear depends on what might be called mer. This has not yet been attempted (but see
its “cohesiveness” which is a function of its SHEINTUCH,1987), although such a model
composition. would be very useful both for activated sludge
Flocs are held together by extracellular bio- systems and for aerobic digestion, which is
polymers produced by the microorganisms, so usually analyzed only in terms of simple, first-
the amount of biopolymer produced must be order kinetics (BHARGAVAand DATAR,1988).
very important in determining cohesiveness of Some qualitative insight can be obtained from
flocs. SHEINTUCH(1987) has reviewed the evi- models in which all stored substrate is lumped
dence for activated sludge failure resulting into a single category. For example, the model
from deflocculation in the settler caused by of Tab. 2 applied to the biomass balance for
lack of natural flocculent. The production of the conventional CMAS system gives
biopolymers has been the subject of considera-
ble research and some controversy. Results
from batch experiments with pure cultures (54)
show that bacteria first accumulate intracellu-
lar storage products such as PHB, and only The total amount of stored substrate de-
later start to produce extracellular biopolymers creases with increasing mean cell residence
(PARSONS and DUGAN,1971). This is in agree- time, but from the previous discussion we ex-
ment with the correlation between biofloccula- pect that the fraction of the stored substrate
tion and PHB content reported by CRABTREE that is extracellular polymer would increase as
et al. (1966). A high C/N ratio in the media 0, increases. This would produce an optimum
produces less biomass, due to nitrogen limita- in sludge settling characteristics as a function
tion, but copious amounts of biopolymer. of Ox, which is precisely what was found experi-
This suggests that extracellular biopolymer mentally by BISOGNIand LAWRENCE(1971).
can best be incorporated into models as a form For 0.25 < 0,<2 days they found predomi-
of stored substrate. In this view, adsorbed col- nantly dispersed growth with practically no
loidal material and PHB are short-term stor- zone settling. This corresponds to our expecta-
age products that the biomass can accumulate tion that at low 0, values the flocs will contain
quickly (PADUKONEand ANDREWS, 1989). considerable adsorbed material and PHB, but
References 431
little biopolymer. For 2<Ox<6 days the flocs ANDREWS,G. F. (1989), Estimating cell and prod-
settled well, but for 6 < Ox<30 days (extended uct yields, Biotechnol. Bioeng. 33, 256-266.
aeration) the particles tended to deflocculate ANDREWS, G. F., TIEN,C. (1977), A new approach
and form pin-point floc that settled poorly. to bacterial kinetics in wastewater, J. Environ.
Eq. (54) would predict this. Even though all Eng. Div. ASCE, 1057-1074.
the stored substrate may now be in the form of ANDREWS,G. F., TIEN C. (1981), Bacterial film
growth on adsorbent surfaces, AIChE J. 23, 182-
biopolymer, there is still not enough of it to 189.
hold the flocs together. ANDREWS,G. F., TRAPASSO, R. (1985), The opti-
SHEINTUCH (1987) also showed that a mod- mal design of fluidized bed bioreactors, J. Water
el containing a purely “chemical” biomass Pollut. Control Fed. 57, 143-152.
structure such as active biomass/stored sub- ATKINSON,B., WILLIAMS,D. A. (1971), The per-
strate would never be sufficient to describe all formance characteristics of a trickling filter with
sludge settling problems. Electrostatic effects hold-up of microbial mass controlled by periodic
may also have a role in the process of biofloc- washing, Trans. Inst. Chem. Eng. 49, 215-224.
culation. Sludge bulking due to the appearance ATKINSON, B., FOWLER,H. W. (1974), The signifi-
of filamentous bacteria is a purely microbio- cance of microbial films in fermentors, in: Ad-
vances in Biochemical Engineering (GHOSE,T.
logical phenomenon, and would require the K., Ed.) Vol. 3. Berlin-New York: Springer.
biomass to be structured into “good” zoogleal BARTON,D. A., MCKEOWN,J. J. (1986), Evalua-
bacterial species and “bad” filamentous ones. tion of an aerator control strategy utilizing time
The usual explanation for this problem is that varying mathematical model simulations, Water
the filamentous organisms have both a lower Sci. Technol. 18, 189-201.
maximum growth rate and lower Monod half BENEFIELD, L. D., RANDALL, C. W. (1986), Design
velocity constants for oxygen and organics procedure for the contact stabilization activated
than the floc formers. This gives them a sludge process, J. Water Pollut. Control Fed. 48,
growth rate advantage with either low dis- 147- 152.
solved oxygen or low organic substrate con- BHARGAVA, D. S., DATAR,M. T . (1988), Progress
and kinetics of aerobic digestion of secondary
centration, and under these conditions they sludges, Water Res. 22, 37-47.
tend to dominate the biomass. EKAMAand BISCHOFF,K. B. (1965), Effectiveness factors for
MARAIS(1986) have reviewed recent results general reaction rate forms, AIChE J. 11, 351-
and models of this problem and shown how 375.
they lead to different strategies for preventing BISOGNI,J. J., LAWRENCE, A. W. (1971), Relation-
it. ships between biological solids retention time and
settling characteristics of activated sludge, Water
Res. 5, 753-763.
BODE,H., SEYFRIED, C. F. (1985), Mixing and de-
tention time distribution in activated sludge
4 References tanks, Water Sci. Technol. 17, 197-208.
BOUWER,E. J. (1987), Theoretical investigation of
particle deposition systems, Water Res. 21, 1488-
ALEXANDER, W. V., EKAMA,G. A., MARCACS, G. 1498.
V. R. (1980), The activated sludge process, Part BUSBY,J. B., ANDREWS,J. F. (1975), Dynamic
2, Application of the general kinetic model to the modelling and control strategies for the activated
contact stabilization process, Water Res. 14, sludge process, J. Water Pollut. Control Fed. 47,
1737-1747. 1055-1080.
ANDREWS,G. F. (1982), Fluidized-bed fermentors CHAIN,E. S., DEWALLE,F. B. (1975), Sequential
and analysis, Biotechnol. Bioeng. 24, 2013- substrate removal in activated sludge systems,
2022. Prog. Water Technol. 7, 235-241.
ANDREWS, G. F. (1984), Parameter estimation from CHESNER, W. H., MOLOF,A . H. (1977), Biological
batch culture data, Biotechnol. Bioeng. 26, 824. rotating disc scale-up design, Prog. Water Tech-
ANDREWS, G. F. (1988a), Effectiveness factors for nol. 9, 811-819.
bioparticles with Monod kinetics, Biochem. Eng. CRABTREE, K., BOYLE,W., McCoy, E., ROHLICH,
J. 37, BI-B37. G. A. (1966), A mechanism of floc formation by
ANDREWS,G. F. (1988b), Fluidized-bed bioreac- Zooglea ramigera, J. Water Pollut. Control Fed.
tors, Biotechnol. Genet. Eng. Rev. 6 , 151-178. 38, 1968-1980.
438 13 Aerobic Wastewater Process Models
DIRKSCHURBUSCHER
CHRISTIAN
WANDREY
Jiilich, Federal Republic of Germany
1 Introduction 445
2 Anaerobic Waste Water Purification 445
2.1 Advantages of Anaerobic Waste Water Purification 445
2.2 Microorganisms and Metabolic Pathways 447
3 Measuring Techniques 450
3.1 Biological Phase 450
3.1.1 Biomass Concentration 450
3.1.2 Biomass Activity 450
3.1.3 Identification of Methanogenic Bacteria 450
3.2 Liquid Phase 450
3.2.1 pH Value 451
3.2.2 Redox Potential 451
3.2.3 Concentrations of Volatile Substances 45 1
3.2.4 Concentrations of Dissolved Gases 451
3.2.5 Carbon Determination 452
3.2.6 Biochemical Oxygen Demand (BOD) 452
3.2.7 Chemical Oxygen Demand (COD) 452
3.2.8 Degradability Tests 453
3.3 Gas Phase 453
3.3.1 Quantity and Volumetric Flow Rate 453
3.3.2 Mass Flow 454
3.3.3 Composition 454
3.4 Level Measurements 454
3.5 Substrate Metering 454
3.6 Example of an Experimental Assembly 456
4 Mathematical Models 456
4.1 Theoretical Mass and Energy Balances 456
4.2 Substrate Consumption and Growth Kinetics 457
4.3 Steady State 459
442 14 Anaerobic Waste Water Process Models
List of Symbols
Latin letters
a model parameter
A- anion concentration
Ac- acetate ion concentration
b model parameter
C model parameter
C concentration
C' cation concentration
CO, carbon dioxide concentration
co: - carbonate concentration
COD chemical oxygen demand
D dilution rate
f state function
F L h-' fluid flow rate
G L h-' gas flow rate
G kJ mol-' free reaction enthalpy
h measuring function
H Henry constant
H' hydrogen ion concentration
HAc concentration of undissociated acetic acid
HCO; bicarbonate concentration
H2co3 concentration of carbonic acid
Z inhibitor concentration
J criterion, square error sum
k discrete time
K amplification matrix
KAC dissociation constant of acetic acid
KC dissociation constant of carbonic acid
KI substrate inhibition constant
KL a mass transfer coefficient
KS Monod constant
Kw ionic product of water
K, steady state parameter
m maintenance coefficient
M molecular weight
n order of reaction
Na +
sodium ion concentration
OH- hydroxide concentration
P pressure
P covariance matrix
Q weighting factor
Q mol L - ' h - ' mole flow
Q covariance matrix
Y mol g-' h-' specific substrate consumption rate
R J mol-' K - ' general gas constant
S variance
S covariance matrix
S mol L - ' substrate concentration
t S time
444 I4 Anaerobic Waste Water Process Models
T K absolute temperature
TO S cycle time
U input variable
U adjustment of controller input
U controller input
V L volume
vm L mol-' mole volume of ideal gases
X state vector
X stoichiometric coefficient
X system matrix
X g L-' biomass concentration
Y measured variable
Y measured value deviation
Y stoichiometric coefficient
Y measured value
yx,s g mol-' yield
Z stoichiometric coefficient
Z' mol L - ' net concentration of strong ions
Greek letters
P h-' specific maintenance metabolism rate
Y correction vector
r thermodynamic biological efficiency
e parameter vector
A forgetting factor
ru h-' specific growth rate
ru" h-' true specific growth rate
fs mass fraction of carbon in a molecule
5 h residence time
Subscripts
C carbon
cal calculated
crit critical
D differential part
des desired value
dis dissolved
exP experimental
F flow
G gas
in input
k discrete time
max maximum
meas measured
min minimum
M mass flow meter
opt optimal
R reactor
stat steady state
S substrate
theor theoretical
Anaerobic Waste Water Purification 445
tot total
Y
A
. linearized variable
X produced biomass
03 steady state
Superscripts
* saturation value
estimated value
Abbreviations
ATP adenosine triphosphate
BOD biochemical oxygen demand
COD chemical oxygen demand
DOC dissolved organic carbon
TIC total inorganic carbon
TOC total organic carbon
tion is similarly carried out by a mixed culture. processes. The different products of anaerobic
However, in this case the waste water consti- and aerobic degradation also present alterna-
tuents are largely fermented one after the other tive possibilities of obtaining energy for the
by the various bacteria into methane and car- microorganisms. For example, in the case of
bon dioxide by a type of "biological pyroly- acetic acid, the most important intermediate in
sis". the anaerobic food chain (WANDREYand AI-
The necessity of an interlinkage of the se- VASIDIS, 1983) the energy balances are:
quential degradation steps in the anaerobic aerobic (corresponding to the reactions of the
process by various microorganisms means that citrate cycle and the respiration chain) (Eq. 1):
the various steps must proceed at the same
speed in order to avoid disturbances. This CH,COO-+H' +202-'2C02+2HzO (1)
means that anaerobic processes are much more A G = -870 kJ mol-'
sensitive to disturbing influences than aerobic
with AG standard reaction enthalpy,
and anaerobic (Eq. 2):
anaerobic aerobic CH3COO- + H + C02 + CH4
-+
(2)
CHjCOOH CH3COOH A G = -31 kJ mol-'
-
in effluent
1'In
org. carbon
in treated effluent
1-5'10
Fig. 2. Comparison of the aerobic (left) and anaerobic (right) degradation of carbon (WANDREYand
AIVASIDIS,1983).
Anaerobic Waste Water Purification 447
only approximately 3070 during aerobic purifi- 2. Conversion of the organic monomers
cation (Fig. 2). by, e.g., Acetobacterium and Pseudo-
Since the elementary compositions of aero- monas to hydrogen, bicarbonate, short-
bic and anaerobic microorganisms do not dif- chain organic acids, alcohols, and meth-
fer significantly, the high growth rate of the ylamine.
aerobic microorganisms is accompanied by a 3. Oxidation of the reduced organic sub-
correspondingly high additional demand for strates to bicarbonate and acetate by
nitrogen and phosphorus. obligate hydrogen-producing anaerobes,
The following advantages for anaerobic for example, Desulfovibrio and Desulfo-
waste water purification result from the char- bacterium.
acteristics of the two processes described 4. Formation of acetate from carbon
above: little excess sludge, no aeration neces- dioxide and hydrogen according to Eq.
sary (energy savings and simpler reactors), en- (3) by homoacetate fermenters, e. g., by
ergy obtained by utilizing the biogas (direct Acetogenium kivui and Acetobacterium
combustion, electricity generation, or input wieringae:
into the natural gas grid after methane enrich-
ment), and heavy metal precipitation in the 2HCO; +4H2+
reactor by conversion into insoluble sulfides --t CH3COO - + 4 H20 (3)
(landfill sites).
The disadvantages are the slow growth of 5 . Oxidation of the reduced organic sub-
the microorganisms (low performance in the stances to bicarbonate and acetate by
stirred tank reactor, time-consuming start-up sulfate- (Desulfovibrio) and nitrate-re-
procedure) and the sequential product degra- ducers (Veillonella alcalescens).
dation (unstable system). 6 . Oxidation of acetate to bicarbonate by
Anaerobic waste water purification plants sulfate- or nitrate-reducers.
are becoming increasingly interesting due to 7. Oxidation of hydrogen by sulfate- or ni-
their simplicity and low energy requirements. trate-reducers with the reduction of sul-
TEMPERet al. (1986) have provided a survey fate to hydrogen sulfide or of nitrate to
of the industrial applications of this process. nitrite, nitrogen, or ammonium.
8. Acetoclastic splitting of acetate into car-
bon dioxide and methane according to
2.2 Microorganisms and Metabolic Eq. (4), for example, by Methanosarcina
and Methanothrix.
Pathways
CH3COOH + CH, + CO2 (4)
The anaerobic conversion of polymeric or-
ganic substrates into methane and carbon 9. Methanogenesis of carbon dioxide and
dioxide is a complex process in which various hydrogen by methane bacteria according
microbial populations play a part because of to Eq. (9,for example, by Methanobac-
their different substrate and product specifici- terium and Methanosarcina:
ties (food chains) (see NYNS, 1986). Nine dif-
ferent steps can be differentiated (HARPER COr + 4H2 CH4 + 2H20
+ (5)
and POHLAND,1987).
Fig. 3 shows a greatly simplified model of the
1. Organic polymers are hydrolyzed by ex- food chain of complex organic macromole-
tracellular enzymes of facultative or cules up to biogas, in which only three degrad-
obligate anaerobic bacteria, e. g., Clos- ation steps are differentiated: acid formation
tridium (degradation of compounds con- from complex organic molecules (see above
taining cellulose and starch) and Bacillus l.), acetic acid formation with the simulta-
(degradation of proteins and fats), to neous formation of hydrogen and carbon
monomeric constituents (amino acids, dioxide (see above 2., 4.), and methane forma-
fatty acids, and sugar). tion (see above 8., 9.).
Anaerobic Waste Water Purification 449
Since methane bacteria are responsible for The Methanosarcinae, approx. 1-2 pm in
the final stage in the anaerobic degradation of diameter, have the greatest metabolic versatili-
hydrogen, the microorganisms belonging to ty of the methane bacteria. Most varieties can
the kingdom Archaebacteria are of decisive utilize methane/carbon dioxide, methanol,
significance for the entire process. The Ar- methylamine, or acetate as methanogenic sub-
chaebacteria form an independent group, strates. They have the property of forming ag-
which is phylogenetically differentiated from glomerates. Tab. 1 shows the energy gains ob-
the so-called classical bacteria (Eubacteriales) tainable with anaerobic substrate degradation
as well as from the higher microorganisms (eu- by methane bacteria. AGO' represents the
karyotes). Archaebacteria live under extreme standard reaction enthalpy under physiological
conditions not unlike those that prevailed dur- conditions (THAUERet al., 1977). This ex-
ing the earlier developmental stages of the plains the differences from the values given in
earth's surface. They are found wherever or- Fig. 1 for standard reaction conditions (25 "C,
ganic substances are decomposed in the ab- lo5 Pa).
sence of oxygen: in swamps, sediments of bod- If one assumes that about 37 kJ mol-' of
ies of water, the digestive tracts of animals free energy is required for the formation of
and, of course, in anaerobic waste water pu- ATP, then - depending on the thermodynamic
rification plants. Isolates have also been ob- efficiency - 2 to 3 mol of ATP can be gener-
tained from geothermal springs and hydrother- ated per mole of methane formed with the
mal craters. The chemoautotrophic, strictly above substrates, with the exception of ace-
anaerobic methane bacteria represent the tate. At most, half a mole of ATP is synthe-
group with the greatest variety of forms, which sized per mole of acetate. Since ATP is neces-
obtain the energy for COz evolution from the sary for anabolism, this explains the slow
reduction of carbon dioxide to methane (Eq. 6): growth and low cell yields of microorganisms
450 14 Anaerobic Waste Water Process Models
on acetate. Various metabolic pathways exist following sections, and particular attention
for methanol, depending on the hydrogen par- will be paid to their suitability for on-line ap-
tial pressure. If sufficient hydrogen is present, plication in process control.
then methanol is reduced directly to methane.
In the absence of hydrogen, the necessary re-
duction equivalents are obtained by oxidizing 3.1 Biological Phase
part of the methanol.
In order to achieve high bioreactor efficien-
cy, and in particular to accelerate the growth 3.1.1 Biomass Concentration
behavior, the biomass concentration in the
reactor must be increased in comparison to Apart from its activity, the concentration of
that of the stirred tank. Various possibilities the biomass is the most important parameter
present themselves for this purpose: pelletiza- for reactor performance, and at the same time
tion, filtration, adsorption, covalent bonding, the most difficult to determine. In order to de-
and colonization (AIVASIDIS and WAN- termine the biomass, the solids contained in a
DREY,1985; HEIJNEN,1984). Immobilization liquid sample are centrifuged or filtered off.
achieves a high biomass concentration in the Those solids remaining after drying contain,
reactor and a decoupling of the residence times among other constituents, the biomass. After
of the substrate and bacteria (WANDREY and weighing, the biomass can be selectively dis-
AIVASIDIS,1983). In this respect fixed-bed solved by the addition of acid or alkali, or else
reactors have the disadvantage that in the by incineration at 550°C. After weighing the
course of time the biomass grows out of the residue, the difference in the two measure-
carriers and into the free space, thus resulting ments can be assumed to be the biomass.
in channel formation and a poorer substrate The biomass can be measured indirectly,
supply to the microorganisms. This type of e.g., by measuring the total nitrogen or the to-
reactor must therefore be regenerated at inter- tal organic carbon of the filtered solids.
vals of a few months. The excess biomass is
sheared off and flushed out either by a short-
term increase in the power of the circulating 3.1.2 Biomass Activity
pump or by blowing in inert gas. As an alter-
native reactor design, fluidized-bed reactors do The fluorescence of the coenzyme Fd2,,,pres-
not have this disadvantage. ent in all methanogenic bacteria, can be used
to measure the activity (SCHNECKENBURGER
et al., 1984). This measurement can rarely be
used for technical media because of the inher-
ent turbidity.
dissolved gases in the liquid. This can be done ing BOD5 with different dilution stages of the
by taking gas specimens via suitable membrane waste water.
systems in conjunction with mass spectrome- Due to the long measuring time, this analy-
ters (JOUANNEAU et al., 1980; SCOTTet al., tical method is not suitable for on-line process
1983; WHITMORE et al., 1987). control but only for determining degradability.
On the other hand, further measuring proce-
dures can be introduced and calibrated on the
3.2.5 Carbon Determination basis of this definition. It is relatively difficult
to correlate the BODS value with other meas-
Three types of situation must be distin- ured values such as the TOC, and it depends in
guished in the quantitative determination of particular on the substrate used.
carbon compounds in aqueous solution, for A short-time toximeter with an analysis du-
each of which several analytical methods are ration of 15 minutes is appropriate for on-line
available (EHRENBERGER, 1979): use. The oxygen consumption of a microbial
Total Inorganic Carbon (TIC): inorganic population in a test reactor is kept constant by
carbon comprises the dissolved carbon in the controlling the feed ratio of one waste water
form of carbonate, bicarbonate, and carbon and one dilution water pump. This inlet ratio
dioxide. is then the measure of the waste water pollu-
Dissolved Organic Carbon (DOC): after re- tion ( K ~ H Net
E al., 1986).
moving the TIC, the dissolved organic carbon BOD probes make use of immobilized mi-
can be detected by oxidizing the dissolved car- croorganisms as receptors and an oxygen elec-
bon compounds and determining the resulting trode as the transducer (SCHUGERLet al.,
carbon dioxide. The carbon contained in par- 1987).
ticles (e. g., microorganisms) is not included. The use of data obtained with aerobic mi-
Total Organic Carbon (TOC): after oxida- croorganisms for anaerobic populations is
tion at a very high temperature in a catalyst problematic in view of the completely different
furnace, the carbon contained in the particles environment and the different substrate spec-
can also be determined. trum. Due to the superior reproducibility of
Titrimetric measurements or analysis by in- the measured data, the BOD determination is
frared absorption are suitable methods for de- frequently replaced by a determination of the
termining the carbon dioxide. The overall Chemical Oxygen Demand (COD).
analysis requires a certain amount of sophisti-
cated equipment but it can be implemented
quasi-on-line. 3.2.7 Chemical Oxygen Demand
(COD)
3.2.6 Biochemical Oxygen Demand The Chemical Oxygen Demand specifies the
(BOD) volume-related oxygen quantity required to
completely oxidize the constituents of the
The biochemical oxygen demand is the pa- waste water. The COD value is suitable as a
rameter most commonly used for quantifying summary measure for determining the pollu-
biologically degradable water pollution, and as tion of the waste water or the quality of the
such it had to be standardized. For this pur- purification procedure. It is the most impor-
pose the concept of BODS was defined as the tant parameter in the waste water legislation.
oxygen demand of a Pseudomonas culture For Germany the method of determination is
over five days. The limitation of the time inter- mandated in the DIN 38409 (OLIVIERand
val necessarily results in an information mix SCHENDEL,1983). It is possible to reduce the
from waste water pollution and inhibition of analysis time to two hours, but nevertheless
the biological system, so this has to be decou- this analytical method can only be taken into
pled by means of a toxicity test. It is possible consideration for slow stirred-tank processes
to calculate the inhibiting effects by determin- for quasi-continuous process control (RENARD
Measuring Techniques 453
et al., 1988). Oxidation is carried out in a de- measuring ranges (> 20 L h - I ) . For this rea-
fined sulfuric acid-potassium dichromate solu- son attention will only be drawn here to a few
tion (KZCr207).Potassium dichromate is re- principles covering the lower measuring range.
duced by the organic constituents of the waste Wet gas meters are frequently used and are
water to C r 3 + . The excess potassium dichro- characterized by a high precision of about
mate quantity is inversely proportional to the f 5 V o to k l q o over a wide range of 1:20
COD determination based on photometry. (STROHRMANN, 1980) as well as by insensitivi-
On-line measuring instruments have recent- ty to fluctuations of the gas composition and
ly become commercially available, permitting corrosiveness of the gas constituents. Howev-
COD determination within 3-5 minutes. In er, these meters measure the gas quantity, not
these instruments hydrogen peroxide or ozone the gas flow.
are used for oxidation. Such instruments are also available with a
slotted disk so that pulses can be counted pho-
toelectrically. In order to calculate a volume
3.2.8 Degradability Tests flow, the pulsed signals must then be filtered
and mathematically differentiated. ERDMAN
A more reliable basis for the process devel- and DELWICHE(1985) provide instructions on
opment of anaerobic waste water purification how to construct such a slotted disk, including
plants can be obtained from anaerobic degrad- evaluation electronics.
ability tests with the microbial population of In a method developed by BEAUBIEN et al.
the industrial process. In this procedure the (1988) a glass flask is filled to a certain pres-
gas production is measured directly or indi- sure with biogas which is then discharged. The
rectly by the internal pressure in a closed con- gas volume flow can be determined automati-
tainer (SHELTONand TIEDJE,1984; DOLFING cally over a range of 1-100 mL min-' by the
and BLOEMEN, 1985; CONCANNONet al., number of pulses, the dead volume of the
1988; SCHOTT,1988). The methane fraction in flask, and the gas space in the fermenter.
the gas can be measured after sampling. The The gas is also frequently collected in a
final pressure level and the type of curve pro- flask from which a liquid has been displaced.
vide information about the total pollution and Acidification of the liquid prevents the carbon
toxicity of the waste water. Work on an on- dioxide from dissolving in it. Conversely, in
line implementation of such degradability tests the case of batch experiments, the gas can be
for the monitoring of waste water purification fed into a flask filled with alkali solution
plants has been carried out by SCHORBOSCHER(Mariotte flask). The carbon dioxide is dis-
(1989). solved in the liquid phase, whereas the me-
thane gas is collected in the head space of the
flask and pushes the liquid into a measuring
3.3 Gas Phase cylinder through an outlet located at the bot-
tom.
Apart from the pH value, the gas measure- Quasi-continuous methods based on the dis-
ment can be regarded as the simplest and most placement principle have also evolved in which
important measurement. This includes mea- the filling level of the flask can be interrogated
surement of the gas quantity and flow, and by a limit detector. After automatic deaeration
also of the gas composition. of the flask, the displaced liquid can flow back
again according to the principle of communi-
cating vessels. The cycles can be counted and
can serve as a measure of the gas quantity
3.3.1 Quantity and Volumetric (MOLETTAand ALBAGNAC,1982; GLAUSER
Flow Rate et al., 1984).
3.3.2 Mass Flow cal models, since carbon dioxide is very soluble
in water. Especially with non-steady-state
In the low gas flow range (<20 L h-I), the processes, this leads to a delayed transition
measuring principle of the thermal mass flow into biogas.
meter is one of the (few) possibilities for ob- Apart from methane, hydrogen is the most
taining a direct on-line signal for the gas flow. important gas component. Measuring instru-
The gas flows through a capillary tube equip- ments developed for medical applications are
ped with two resistance thermometers and a used for the measurements.
heating coil. The heating spiral is located in
the center between the two measuring sensors.
The gas is warmed at the heating spiral as it 3.4 Level Measurements
flows through the capillary. The resulting tem-
perature difference between the two sensors is Due to the low demands made on the bio-
proportional to the gas flow and causes a reactor (no sterility but, on the other hand,
change in sensor resistance. The current neces- high corrosiveness), reactors of glass or plastic
sary for balance via a bridge circuit is used as are frequently used for studies on a laboratory
the measuring signal. Since the temperature scale. In this case, the filling level can be easily
difference is dependent on the heat capacity of controlled by measuring the liquid level (vol-
the respective gas, the sensor must be cali- ume regulation).
brated for the gas or gas mixture to be mea- In this connection, conductometric measur-
sured. The instruments are therefore adjusted ing principles can be applied. One sensor con-
to a particular gas mixture by the manufactur- sists of two wires, one of which must be in
er (e. g., 50% COz, 50% CHJ. The gas is dried continuous contact with the liquid. The other
before entering the sensor in order to avoid is guided vertically through the reactor lid in
measuring distortions and corrosion. Unfortu- such a way that its tip marks the desired filling
nately, in spite of these measures such sensors level. When the contact is closed, the discharge
are affected over the course of time by the or- pump is switched on by an electronic device.
ganic sulfur compounds contained in the bio- These instruments have the disadvantage that
gas, so that in the final analysis this solution is they react to foam. Furthermore, the switching
not satisfactory. pumping process causes a noise signal in the
gas flow so that it first has to be filtered before
further mathematical processing (left part of
3.3.3 Composition Fig. 5 ) .
Another possibility would be to use an ana-
The most important compounds, methane log filling level sensor providing a signal pro-
and carbon dioxide, are measured by the infra- portional to the filling level. Such sensors
red absorption of the gas constituents. The make use of, for example, the principle of the
wavelength of the absorption bands character- Wheatstone bridge. The sensor consists of a
ize the gas type, whereas the extent of absorp- steel probe dipped into the liquid and thus of-
tion is a measure of the concentrations of the fers an infinite number of resistances. A signal
components. The gas flows through a measur- proportional to the depth of submergence is
ing cuvette for each gas component to be mea- generated. The performance of the discharge
sured. The measuring cuvettes are irradiated pump can then be continuously controlled by a
one after the other with light of wavelengths at suitable PID controller (right part of Fig. 5 ) .
which the component to be measured has its
absorption maximum and no absorption, re-
spectively. In this way one obtains in succes- 3.5 Substrate Metering
sion a concentration-independent and a con-
centration-dependent signal. The difference is An important parameter for continuous bio-
a stable measure of the concentration. technical processes, in particular for studies on
The methane concentration in the gas is par- a laboratory scale, is the residence time or the
ticularly important for compiling mathemati- flow rate. Since the waste water is sometimes
Measuring Techniques 455
L
2
~
loo
20
0
0.1 0.8 1.6 2.0 2.1
Time l h l
Fig. 5. Elimination of noise in the gas measurement.
Waste water Alkali Fixed bed Cleaned Fig. 6 . Scheme of the experimental
loop reactor effluent setup.
456 14 Anaerobic Waste Water Process Models
corrosive and polluted with solids, not all feed ed. A model will then be given that assumes a
systems can be taken into consideration. Flexi- constant p H value, followed by a model also
ble tube pumps are particularly suitable for suitable for simulating changes in the pH val-
this purpose. ue. All models are derived for continuous reac-
A metering system permits a very accurate tors with ideal stirred tank behavior for bio-
determination of the substrate flow in the mass and substrate.
course of which a flexible tube pump, gear
pump, or membrane pump transports the sub-
strate from an intermediate store on a balance 4.1 Theoretical Mass
into the reactor. The balance and the micro-
processor-controlled pump are connected to and Energy Balances
each other so that it is possible to exchange
data (VETTERand CHRISTEL,1988). The me- The composition of the biogas can be calcu-
tering system uses the weight decrease of the lated from the mass balance by the stoichiome-
substrate on the balance as the controller in- try for the anaerobic degradation of a complex
put. substrate C,H,O, (Eq. 7).
tion are measured on a standard basis. The By analogy, the chemical oxygen demand for
substrate is continuously metered with a flexi- the same substrate results from Eq. (8):
ble tube pump, and the filling level is con-
trolled with a level sensor by a discharge
pump. The contents of the column-shaped
reactor is stirred by a circulating pump that
continuously pumps liquid from the top of the
reactor to the bottom. The substrate pump and
+ -X2 HzO
the reactor temperature can be controlled by a X
connected personal computer. A chemical oxygen demand of 2 n + - - y mol
The p H value is checked off-line daily. The 02/mol C,H,O, thus results. 4
COD value is determined by a cuvette test, and In the case of a generalized application of
the individual concentrations of volatile com- the balances with the introduction of the de-
ponents by a gas chromatograph. gree of reduction, including not only the chem-
ical elements but also the electrons available,
the electric charge and the energy are also tak-
en into consideration (SCHUGERL,1985). The
degree of reduction of a given organic sub-
4 Mathematical Models stance C,H,O,N, is calculated as:
+
7 = (4 X - 2 y - 3z)/n (8 a)
Since anaerobic processes display very long
time constants, particular importance is at- The mass fraction of carbon in the molecule
tached to mathematical modelling and the im- results from:
plementation of studies by means of simula-
tions. First, the possibilities provided by the
basic mass and energy balances will be present- __
Mathematical Models 457
the double reciprocal plot after Lineweaver- deed the poorest of the linear transformations.
Burke (Eq. 15): For this reason, methods of non-linear optimi-
zation are applied almost exclusively today; in
1 - K S 1 1 the following, mainly the Rosenbrock method
-+- (15) (HOFFMANNand HOFMANN, 1971). Fig. 7
r ( S ) rmax S rmax
shows the substrate consumption kinetics cal-
However, due to distortions these transforma- culated according to both methods with the
tions do not lead to statistically perfect state- following values: r m a , = 0 . 0 1 2 5 mol g - ' h - '
ments (CORNISH-BOWDENand EISENTHAL, and Ks =0.0067 rnol L-'.
1974). The Lineweaver-Burke method is in-
'"1
u 7-
0
I I
LO 60
I
120
I
160
I
It can be seen that with linear plotting the with pZaXtrue specific growth rate and
small substrate concentrations have an inap- P specific maintenance metabolism
propriately large influence on the result. Up to rate.
this point is was possible to evaluate the data The measurable growth represents the
without any statement about the nature of the growth converted into biomass, the concept of
coupling between microbial growth and sub- true growth is a fictive growth including the
strate consumption. maintenance metabolism (BERGTER, 1983).
2.0 0.030
I x o Real growth
................................................... I
I/ I+
0 , I x 0.3 mol L-'
I
0.6 rnol L-' 10 0.9 rnol C' Ix 0.3 mol L'' I+ 0.6 rnol L'' 100.9 mol L-l
0 0.001 0.008 0.012 0.016 0.020 0.021 0 0.01 0.08 0.12 0.16 0. 0
Growth rate (h-') Substrate concentration ( m o l L-ll
Fig. 8. Growth kinetics.
If the maintenance metabolism is taken into The following values result for the param-
consideration, the substrate consumption rate eters pZax= r,,, Yx/smax= 0.0263 h - I , Ks =
results from the growth of the biomass and its 0.0067 mol L-I, ~ = m Y x ~ s , , x = 0 . 0 0 4 6h-I.
maintenance metabolism (Eq. 16): A survey of the derived equations and the
parameters is given in Tab. 3.
Gas Phase
G c H ~ =V, V R W X v,=5 L
Vm=25.5Lrnol-’
Liquid Phase
dS
- = (Sin-S)D-r(S)X
r,,, = 0.0125 rnol g -’ h - ’
dt
S
r(S) = rmax-
Ks +S
Biological Phase
p&,,=0.0263 h-’
Ks = 0.0067 rnol L
P = 0.0046 h -’
The minimum substrate concentration that results, but not an increase in substrate con-
just covers the requirements for maintenance centration.
metabolism can be calculated from ,u = 0 (Eq. However, Fig. 9 also shows a definite dis-
22). crepancy between the model and measured val-
ues, particularly at high flow rates. This can be
m attributed to growth on the wall, which occurs
smin. =- Ks=0.0014mol L - ’ (22) especially if sessile microorganisms are used,
rmax -m
This has a particularly distorting effect if mi-
Curves typical of a continuous process can be croorganisms with very long generation times
seen in the plots versus the flow rate. If the are involved. As a rule, experiments with small
substrate concentration in the inlet is in- flow rates are not implemented due to the long
creased, an increase in biomass concentration time constants. For example, a flow rate of
Mathematical Models 461
Gcal=2VmVRr(S)X (25)
with Vm=25.5 mol L - ' (at 38°C).
Only approx. 88% of the theoretical gas
production can be measured. This can be ex-
plained by the high solubility of carbon
dioxide in the liquid. That is to say, part of the
carbon dioxide does not leave the reactor with
the biogas but is flushed out with the waste
water (Eq. 26):
-
- Qexp -
-
QG,CH, + QG,CO, + QF,CO,
-
- Qexp
4.5 Application
-- of the Steady-State Fig. 11 shows the experimental runs and the
simulations made with the model determined
Model to Non-Steady-State during stationary experiments. It can be clear-
-
Experiments ly seen that these results are not satisfactory.
In experiment I, substrate concentrations are
Three experiments were carried out by reached which were not set during the steady-
TRETTER (1987) according to the program state experiments to determine the kinetics.
shown in Tab. 4. This first inhibits the microorganisms, but
I I1 111
-- 0,
I ..I , 1 I I I
m rn
-.
I
. . . I
, I *
0.8
0.1
0.2
0
-
0.6 -
---
<'
-
measured ]-calculated
. ..be**. . ... .
. . ..
0.7 ! I
0 2 1 6 8 1 0 1 2 1 1
Time I d 1 lime ( d l
Experiment III
0.15-
0.1 0 .-
0.05-
01
0.2
0-
- ...--
0 2 1 6 8 1 0 1 2 1 1 0 2 1 6 8 1 0 1 2 1 1
lime ( d 1 lime ( d 1
Fig. 11. Non-steady-state courses of experiment and simulation.
Mathematical Models 463
(37)
dt
dS
- = (Si,-S)D-r(HAc)X
dt In calculating the pH value, the carbon dioxide
produced by the microorganisms during degra-
dation of the acetic acid and thus dissolved in
(33) the waste water must be taken into considera-
tion. Carbon dioxide is present in various dis-
A further differential equation permits the solved forms in the liquid:
simulation of titration with sodium hydroxide
solution. It is assumed that the sodium hy- =
c02t0, C02dis+ H2C03 + H c o ; + c0:- (38)
droxide solution is always completely disso-
ciated: with CO,,,, total concentration of all forms
of C02 in the liquid phase
d Na+ - (Nu; -Nu +)D
--
C02dis concentration of carbon dioxide
(34) dissolved in the liquid
dt
H2C03 concentration of carbonic acid.
In order to calculate the pH value, the net con- Due to the low stability of carbonic acid, its
centration of strong ions concentration is negligible in comparison with
the dissolved C02. This is taken into consider-
ation in defining the dissociation equilibrium
between bicarbonate and the total concentra-
is introduced into the charge balance: tion:
Z++Na'+H+= KCICO2ciis
HCO; = (39)
= OH- + A c - + HCO; + 2COf (35) H+
S =Ac-+HAc with Kcl dissociation constant of bicarbonate
(5.01 x lo7 mol L-' at 38°C).
KAc The following is true for the dissociation
Ac- = S (36) constant of bicarbonate to carbonate:
H f +KAc
3- 0.010
2 0.000
1 I l l , , , , , I I I , , , , , ,
The carbonate concentration can thus also be The following calculations assume an acetic
neglected; Eq. (42): acid concentration of Si,=0.36 mol L-'. It
thus follows that: Z ' = -0.012 mol L - I .
These negatively charged ions are mainly sul-
fate and sulfite. Fig. 12 shows the pH values
After combining Eqs. (39) and (42) the follow- arising in the reactor with certain specific con-
ing is obtained for the dissolved carbon version rates. Since a pH value of 6.4 is re-
dioxide: garded as optimum according to experience,
the waste water is pre-titrated to pH = 4 (left
(43) figure) in order to achieve this value at degrad-
ation rates of more than 80% (right figure).
The quantity of sodium hydroxide solution
required is calculated as:
The saturation concentration for the dissolved
carbon dioxide resulting from contact with the No+ = A c - -H' -z+ (46)
gas phase can be calculated by the Henry-Dal-
ton law (Eq. 44): with A c - =0.067 mol L-' and H + = mol
L - ' from which it follows that Na+ =0.079
mol L - ' .
Fig. 13 shows a titration curve calculated
with H Henry constant = with the derived concentration in comparison
3.01 x lo-' mol L - ' Pa-' with the actual measurement. At first glance
(at 25 "C) the agreement seems very good, but particular-
2.11 x mol L - ' Pa-' ly in the pH range between 6 and 7, which is of
(at 38 "C) interest here, there are great deviations. Thus,
pco, C 0 2 partial pressure in the gas for example, a slight deviation in the sodium
phase. hydroxide solution can result in a calculated
The concentration of Z + in the waste water pH value of between 7 and 13. In comparison,
can be calculated from the ion balance with the real system is much more strongly buf-
the pH value of the substrate present (here fered. Simulations undertaken with this model
pH = 1.9). At this pH value the OH - concen- can therefore only provide qualitative state-
tration can be neglected due to the ionic prod- ments.
uct of water at 25°C; K w = H + * O H - = 1 0 - ' 4 With pH values adjusted in the reactor be-
(see Eq. 35): tween 6 and 7, the hydrogen ion and hydroxide
ion concentrations can be neglected.
Z + = A c - -H+ (45)
466 14 Anaerobic Waste Water Process Models
12 - -12
10-
-10
I
CL
-8
-6
NaOH ( m t ) NaOH ( m L J
Fig. 13. Titration of waste water with sodium hydroxide solution.
After the dissociation equilibria for acetic sorption effects) that it is very difficult to cal-
and carbonic acid have been substituted in the culate even with model solutions (SCHUMPE,
ion balance (Eq. 35), the following results as a 1985). Further information on calculations of
simplified balance: pH values can be found in BLIEFERT(1978),
PONSet al. (1990), and WELLINGER (1985).
KAC KCl c d i s The total balance for the carbon dioxide in
Z + +Nu+=S
H++KAc
+- H+
(47) the liquid is now as follows:
Gas Phase
R T VR v R = 5L
dpcO, = - -(PQGCO, -PCO, Qc)
dt P VG Vm=25.5Lmol-'
p = lo5 Pa
R=8.3143 JmolV' K-'
GcH,= V m VRQX,CH~
QG = Qcco, + Qx,CH,
Mass Transfer
KLa
H=2.11~10-~molL-'Pa-'
Liquid Phase
KAc= 1 . 7 0 lo-'
~ mol L-'
Kcl = 5.01 x
dZ dNa Sin= 0.36 mol L-'
-- +
- (ZL - Z + ) D
+
-= (Na; - N a + ) D
dt dt Z + = -O.O12molL-'
dCOztot N a + =0.065 mol L-'
-= Qx,co,- COztotD - QCCO,
dt
Biological Phase
-d_X - (p(HAc)-D)X
dt
Observation of Non-Measurable Variables 469
fore requires a selection of the really important The Luenberger observer is in the first in-
parameters to be adapted and exclusion of oth- stance only applicable to linear systems. For
ers that can be assumed to be constant (BAS- non-linear systems it would have to be linear-
TIN et al., 1983; BOLLEet al., 1986; HAVLIK et ized once by a defined operating point or else
al., 1984). continuously along the trajectory. A non-lin-
ear observer proposed by ZEITZ(1977) can be
generally solved for second-order systems. It
cannot be applied to higher order systems due
to non-linear growth and substrate kinetics
5 Observation of and the associated considerable analytical
mathematical efforts. It is possible to use non-
Non-Measurable Variables linear systems with the Kalman filter, although
a numerical differential equation solver (e. g.,
Runge-Kutta) is then required.
5.1 Kalman Filter
are not deterministic but rather stochastic. The (large s), this leads to a lower weighting of the
Kalman filter then calculates its optimum ei- error of observation.
genvalues itself on the basis of the given values A variance is similarly allocated to the state
by means of the noise behavior of the system. variables through the matrix Q . Its values have
The following results for the amplification ma- a directly proportional effect on the increase
trix: of the matrix P over time and thus once again
on the amplification vector K .
The covariance matrix P enters noise into the 5.2 On-Line Calculation
deterministic differential equation system and
thus represents a measure of the inaccuracy of of the Biomass Concentration
the mathematical model. The standard devia-
tions of the individual state variables can be In the case of anaerobic waste water purifi-
calculated by extracting the root from the ele- cation, for example, the Kalman filter can be
ments of the principal diagonal so that they used to attempt to calculate the biomass con-
can, for example, be included in the graphic centration and its growth rate from the change
representation (GILLESand SCHULER,1981). in the biogas flow:
A further differential equation system results
in order to calculate the covariance matrix: dk
- = ( p -D)R
dt
ference between the calculated and measured It can be seen that the influence of the error
methane flow is not included, a rapid conver- of estimation for the initial concentration de-
gence of the calculated substrate concentra- creases exponentially with time, i. e., conver-
tions cannot be achieved. After substitution, gence is ensured in any case. For short cycle
Eq. (64) follows from the differential equation times, the differential equation can also be di-
for the substrate concentration (Eq. 62) and rectly converted into a difference equation by
the measuring equation (Eq. 63): rectangular integration:
S ( k )= S ( k - 1 ) +
-dS_ - (Sin- S ) D - -
P W X or
dt
dS
- - - (Sin- S )D - r ( S ) X
yx/s
(62)
+
[ Vm1
( S i n - S ( k - l ) ) D - L To (68)
GCH VR
5
S('=O) GCH,
= jdt
o
(65)
(Sin- S )D-- with GM, gas flow measured by the mass flow
vm VR meter during time interval i
1' V, gas volume measured through the
with GCH, = - GCH4(t)
dt,
to gas meter.
the relations follow, assuming that GCH, re- The correction factor k2 takes the composi-
mains constant for sufficiently short times: tion of the waste water into consideration.
Since the balances for acetic acid were set up
D # 0: S ( t )= S ( t = 0) e - D f + as reference components, the biogas fraction
+ sin- GCH, ( 1 -e-"')
( ~
DvmvR
) attributable to conversion of the acetic acid is
estimated by means of this factor.
bCH,
D=O: S(t)=S(t=O) -- t
Vm VR
or, in the time-discrete notation: The substrate balance is fitted to the laborato-
ry analyses bylthe factor k3. The acetic acid
D*o: t$k=L!?k-le-DTo+ concentration S ( t J estimated at time tl is veri-
GCH4 ) ( 1 -ePDTo)
fied by means of the concentration Smeas(tl)
+ sin- ___
( DvmvR
determined by a laboratory analysis. If devia-
tions occur here it leads to adaDtation of the
factor k3 by means of the amplification factor
(67) fi
472 14 Anaerobic Waste Water Process Models
0.028
* Flow rate
0.026---Growth rate
3
0
0.021- u-
0.022-
W
0.0204 0.5
0.018 ' I
0.1 1 1 -
I
!-I
Ul 1.2-
-e
I
c
.= 1.0-
z 0.8-
U
W
.9
m
m
0.2-
..
I Y .
0 5 10 15 0 5 10 15
Time I d I Time ( d l
Fig. 15. Observation of state variables (experiment I).
5.4 Results
The fact that it is worthwhile to apply the
Kalman filter even with very few measured I I I I I I I
points has been shown in the evaluation of the 0 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.10
non-steady-state experiments already present- Subslrate concentration (rnol t-']
ed in Fig. 11. The value for the gas flow aver-
aged over one day served as the data base for Fig. 16. Comparison of non-stationary with steady-
state growth rates (experiment I).
the evaluation in Figs. 15, 17, and 18. The gas
composition was not measured in this experi-
ment. The measured methane flow was there-
fore assumed to be 57% of the measured total
gas flow (Eq. 27). ing the biomass concentration was calculated
Since the experiments were started from the from the actual biogas flow:
stationary state, the actual flow rate ii = D can
be selected as the initial value for estimating
the growth rate. The initial value for estimat- (73)
Observation of Non-Measurable Variables 473
0.0227 i
Flaw rate
-Growth rate
,
Y. I"
.. I
0.05-
. . 8 .
. * . . I
-5.0
0 5 10 1
Time ( d 1 Time I d )
Fig. 17. Observation of state variables (experiment 11).
0.022 7- 2.0
--- 0.020 .
--calculatedl measurec
-
= 5 0.018
w -w
I
.k 1 . 5 1 \
3 2 0.016
3 =
2 3 0.011
U VE
0.012
-Growth rate
0.01 0
0 4 e 12
Time ( d 1 Time ( d )
Fig. 18. Observation of state variables (experiment 111).
474 14 Anaerobic Waste Water Process Models
As described above, in pH-auxostatic opera- HOLT and KOMMEL, 1979; ISERMANN, 1987b;
tion, the pH value in the reactor is controlled JACOBS et al., 1980) (Fig. 19).
by the addition of acid waste water through For this purpose a black-box model de-
the substrate pump. Switching controllers are scribing the process must be found whose pa-
used to turn the pump on and off. If the flow rameters are then continuously adapted to the
rate is to be used as a measured value for more input and output behavior of the process in a
complex mathematical evaluations, then it is parameter identification carried out in parallel
best to use continuous controllers permitting with the process. This type of experimental
continuous operation of the pump. In this modelling (identification) replaces the theoreti-
way, the control signals can also be used as in- cal formulation of mechanistic models wher-
directly measured variables without further fil- ever they would involve unacceptably great ef-
tering. The system dynamics are altered in the forts in order to achieve a certain accuracy.
course of time due to the growth of the bio- This is true more of chemical engineering proc-
mass and the necessary periodic regeneration esses than of mechanical and electrical proc-
of the fixed-bed reactor, and also to activity esses, and accordingly should be of great val-
fluctuations and variations in temperature or ue, particularly in biotechnology (HENSELet
waste water. If a conventional PID controller al., 1986).
were used, its parameters would have to be In the present case it is not possible to de-
continuously adapted to the operating state. scribe the control section by a physical model
The use of a parameter-adapted controller due to the complex pH behavior. A SISO (sin-
would therefore be appropriate. These con- gle-input single-output) model can be used as a
trollers make use of a mathematical model of black-box model with the substrate flow as the
the process, continuously adapting its parame- input and the pH value as the output variable.
ters to the actual course of the process, and on Since digital computers read the data in fixed
the basis of a predefined control criterion can cycle times, i. e., discretely, time-discrete dif-
correspond to the instantaneous system dy- ference equations must be introduced for the
computer-internal representation of contin-
uous differential equations. A linear process
without dead time as the simplest form of sys-
tem behavior can be described by an ordinary
output y difference equation of the nrh order. If the
parameters for the difference equation system
Process
of the nth order are combined, then the devia-
tion values generally found in automatic con-
trol engineering are:
+
y ( k ) a, y (k - 1)+ . . + a, y (k - n) =
*
=b,u(k-I)+ ... + b , u ( k - n )
u (k)= U ( k )- um (74)
The cycle time is no longer explicitly given, with the measuring vector x and the parameter
but is included in the parameters. This must be vector B (Eq. 78):
remembered if a different cycle time is to be
used. wv = [a1(W, * * a, (N,b1 ( N ) ,
9
(94)
Experiments with this control have been imple-
mented by RENARD et al. (1988). They The following thus results for the controller:
checked the COD value every two hours. The DOCHAIN and BASTIN(1985) had already de-
value C, was substituted as a constant, where- rived such an integral approach for a two-stage
as the variable C1 was substituted in various reactor (methanation stage with an acidifica-
experiments sometimes as a constant and tion stage connected in series).
sometimes as proportional to the gas flow
C, (t)= C1G ( t ) . Rwas little changed during on-
line fitting. 6.4 Hydrogen Regulation
COSTELLOet al. (1989) therefore kept this
parameter constant during their simulations. Since hydrogen plays an important part in
They demonstrated in simulations that in the the interactions between the individual micro-
controls pH
Temperature
controls f l o w r a t e
pH Value
and conversion r a t e constant
I
bial populations, and the rise in its concentra- ance is particularly steep. Fig. 22 shows two
tion is an important indication of an overload- experiments carried out with a fixed-bed reac-
ing of the fermenter, regulation of the hydro- tor with different temperature increase gra-
gen concentration may be a possible operating dients.
strategy. WHITMOREet al. (1987) measured
the dissolved hydrogen and methane concen-
trations with a membrane and a mass spec-
trometer. It was demonstrated in experiments
that acidification of the reactor can be pre-
vented by regulating the dissolved hydrogen
increase of
concentration to 1 pmol by means of the flow
rate. temperature
Observation of
reactor
6.5 Optimization
of Operating Parameters
and storage
Greatly differing data are given ..i the litera- of characteristic
ture for the optimum temperature and pH temperature
ranges in anaerobic waste water purification.
Furthermore, data are rarely given for the
method used to determine the optimum values.
For this reason, experiments were undertaken
to determine such parameters by computer.
These methodological developments are natu-
rally of general significance for biotechnology.
However, they can be tested particularly effec- characteristic
tively with the anaerobic waste water purifica- temoerature
tion process, since this process is not operated
continuously or under sterile conditions, and
the pH value can be easily measured on-line as reactor
an indirect measured variable for the residual performance
substrate concentration and the biogas flow.
MARKLet al. (1983) proposed a computer-au-
tomated pH optimization which was tested by performance performance
MATHER(1986) in simulations. EBERHARDT declining?
(1989), FREYER(1988), and SCHURBUSCHER
(1989) developed a computer-controlled tem-
perature optimization that functions according Fig. 21. Flow diagram for temperature optimiza-
to the diagram shown in Fig. 20. tion.
The temperature is varied by the computer
according to the diagram shown in Fig. 21. A
change in temperature causes an altered de-
gradation rate and would cause a change in pH In spite of different experimental sequences,
if the parameter-adaptive pH controller dis- the experiments converge towards an optimum
cussed in Sect. 6.1 did not keep this value con- temperature of 42°C with a performance in-
stant. An increase in the flow rate is thus ob- crease between 25 and 30%. The time con-
tained as a valuable signal for reactor perform- stants of the system must be taken into consid-
ance. The algorithm also recognizes so-called eration with the configuration of the optimiza-
characteristic temperatures in the case of a per- tion sequence. The influence of temperature
formance change in which the rise in perform- on cell reproduction is not accounted for with-
480 14 Anaerobic Waste Water Process Models
13
-
t 12-
g
c
11-
rn
l 10-
t
a
:3 9 -
38 -
6.5 6.5
r
n
6.1 - 6.1 -
6.3 I I I I , 6.3i I I I I I
controllers for waste treatment by anaerobic di- HSIA,T. C. (1977), System Identification, pp. 69ff,
gestion, Environ. Technol. Lett. 6 , 584-593. 83, Lexington: Lexington Books.
DOLFING,J., BLOEMEN,W. (1985), Activity meas- ISERMANN, R. (1974), ProzeBidentifikation, pp. 45-
urements as a tool to characterize the microbial 49, Berlin: Springer-Verlag.
composition of methanogenic environments, J. ISERMANN, R. (1986), Experimentelle Identifikation
Microbiol. Methods 4, 1-12. und rechnergestiitzter Regler-Entwurf bei techni-
DROSTE,R. L., KENNEDY, K. J. (1988), Dynamic schen Prozessen, Chem. Ing. Tech. 58, 875-887.
anaerobic fixed-film reactor model, J. Environ. ISERMANN,R. (1987 a), Digitale Regelsysteme, Ber-
Eng. 114, 606-620. lin: Springer-Verlag.
EBERHARDT, R. (1989), Rechnergestiitzte Regelung ISERMANN, R. (1987 b), Stand und Entwicklungs-
biotechnischer Prozesse, Dissertation in prepara- tendenzen bei adaptiven Regelungen, Automati-
tion, Institut fur Biotechnologie, Forschungszen- sierungstechnik 35, 133-143.
trum Julich KFA. JACOBS,0. L. R., HEWKIN,M. A., WHILE, C.
EHRENBERGER, F. (1979), Zur Bestimmung von (1980), Online computer control of pH in an in-
Sauerstoffbedarfs- und Kohlenstoff-Kennzahlen dustrial process, IEE Proc. 127, 161-168.
(TOD, TOC, DOC usw.) in der Wasserqualitats- JOUANNEAU, Y., KELLEY,B. C., BERLIER,Y.,
bestimmung, GIT Fachz. Lab. 23, 738-747. LESPINAT,P. A., VIGNAIS,P. M. (1980), Con-
EISENTHAL, R., CORNISH-BOWDEN, A. (1974), The tinuous monitoring, by mass spectrometry, of H2
direct linear plot, Biochem. J. 139, 715-720. production and recycling in Rhodopseudomonas
ERDMAN,M. D., DELWICHE,S. R. (1985), Low- capsulata, J. Bacteriol. 143(2), 628-636.
cost digital counting interface for fermentation KLEINSTREUER, C., POWEIGHA,T. (1982), Dy-
gas measurement, Biotechnol, Bioeng. 27, 569- namic Simulator for anaerobic digestion proc-
571. esses, Biotechnol. Bioeng. 24, 1941-1951.
FREYER,S. (1988), Rechnergestiitzte Optimierung KOHNE, M., SIEPMANN,F. W., TE HEESEN,D.
eines anaeroben Abwasserreinigungsprozesses, (1986), Der BSBS und der kontinuierliche Kurz-
Diplomarbeit, Institut fur Biotechnologie, For- zeit-BSB (BSB-M3) im Vergleich, Korrespondenz
schungszentrum Jiilich KFA. Abwasser 33, 787-793.
GILLES,E. D., SCHULER,H. (1981), Zur frtlhzeiti- MACARIO,A. J. L., CONWAYDE MACARIO,E.
gen Erkennung gefahrlicher Reaktionszustande (1985), Antibodies for methanogenic biotechnol-
in chemischen Reaktoren, Chem. Ing. Tech. 53, ogy, Trends Biotechnol. 3, 204-208.
673-682. MARKL, H., MATHER,M., WITTY, W. (1983),
GLAUSER,M., JENNI,B., ARAGNO,M. (1984), An Men- und Regeltechnik bei der anaeroben Ab-
inexpensive, automatic gas meter for laboratory- wasserreinigung sowie bei Biogasprozessen, Miin-
scale methane digesters and other gas-evolving chener Beitriige zur Abwasser-, Fischerei- und
systems, J. Microbiol. Methods 2 , 159-164. FluJbiologie, Vol. 36, Miinchen: R. Oldenbourg
GRAEF,S. P., ANDREWS, J. F. (1973), Mathemati- Verlag.
cal modeling and control of anaerobic digestion, MATHER,M. (1986), Mathematische Modellierung
Water, 101-131. der Methangarung, Fortschr. Ber. VDI, Reihe
HARPER,S . R., POHLAND, F. G. (1987), Enhance- 14, No. 28. Diisseldorf: VDI-Verlag.
ment of anaerobic treatment efficiency through MOLETTA,R., ALBAGNAC, G. (1982), A gas meter
process modification, J. WPCF 59, 152-161. for low rates of gas flow: Application to the
HAVLIK,I., VOTRUBA,J., SOBOTKA,M., VOLES- methane fermentation, Biotechnol. Lett. 4, 3 19-
KY, B. (1984), Parametric sensitivity in modeling 322.
of anaerobic digestion, Biotechnol. Lett. 6, 607- MONZAMBE, K. M., NAVEAU,H. P., NYNS,E.-J.,
610. BOGAERT,N., BUHLER,H. (1988), Problematics
HEIJNEN,J. J. (1984), Technik der anaeroben Ab- and stability of on-line pH measurements in
wasserreinigung, Chem. Ing. Tech. 56, 526-532. anaerobic environments: The jellied combined
HENSEL,H., ISERMANN, R., SCHMIDT-MENDE, P. electrode, Biotechnol. Bioeng. 31, 659-665.
(1986), Experimentelle Identifikation und rech- NYNS, E. J. (1986), Biomethanation processes, in
nergestiitzter Regler-Entwurf bei technischen Biotechnology (REHM, H.-J, REED, G., Eds.),
Prozessen, Chem. Ing. Tech. 58, 875-887. Vol. 8, pp. 207-268, Weinheim: VCH.
HILL,D. T., BARTH,C. L. (1977), A dynamic mod- OLIVIER,M., SCHENDEL,F.-A. (1983), Gesetzge-
el for simulation of animal waste digestion, J. bung (Stand und Trend); analytische Erfassung
WPCF, October, 2129-2143. der Abwasserinhaltsstoffe, Papers Verfahrens-
HOFFMANN, H., HOFMANN, U. (1971), Einfiihrung technik der Biologischen Abwasserreinigung,
in die Optimierung, p. 106, Weinheim: Verlag Krefeld: VDI-Gesellschaft Verfahrenstechnik und
Chemie. Chemieingenieurwesen.
References 483
ONER, M. D., ERICKSON,L. E., YANG, S . S. der anaeroben Abwasserreinigung, Berichte der
(1984), Estimation of yield maintenance, and Kernforschungsanlage Jiilich Nr. 2325.
product formation kinetic parameters in anaero- SCOTT,R. I., WILLIAMS, T. N., LLOYD,D. (1983),
bic fermentations, Biotechnol. Bioeng. 26, 1436- Oxygen sensitivity of methanogenesis in rumen
1444. and anaerobic digester populations using mass
OZTURK,S. S., PALSSON,B. O., THIELE,J. H. spectrometry, Biotechnol. Lett. 5(6), 375-380.
(1989), Control of interspecies electron transfer SHELTON,D. R., TIEDJE, J. M. (1984), General
flow during anaerobic digestion: Dynamic diffu- method for determining anaerobic biodegrada-
sion reaction models for hydrogen gas transfer in tion potential, Appl. Environ. Microbiol., 850-
microbial flocs, Biotechnol. Bioeng. 33, 745- 857.
757.
PHELPS,T. J., CONRAD,R., ZEIKUS,J. G. (1985), SHINSKEY, F. G. (1974), Adaptive pH controller
Sulfate-dependent interspecies H2 transfer be- monitors nonlinear processes, Confrol Eng.,
tween Methanosarcina barkeri and Desulfovibrio Feb., 57-59.
vulgaris during coculture metabolism of acetate SOBOTKA,M., VOTRUBA,J., HAVLIK,I., MINKE-
or methanol. Appl. Environ. Microbiol. 50(3), VICH, J. G. (1983), The mass-ernergy balance of
589-594. anaerobic methane production, Folia Microbiol.
PONS, M. N., GARRIDO-SANCHEZ, L., DANTIGNY, 28, 195-204
P., ENGASSER,J. M. (1990), pH Modelling in STEPHANOPOULOS, G., SAN, K.-Y. (1984), Studies
fermentation broths, Bioprocess Eng. 5 , 1-6. on on-line bioreactor identification, BiotechnoL
POWELL,G. E., ARCHER,D. B. (1989), On-line ti- Bioeng. 26, 1176-1218.
tration method for monitoring buffer capacity STROHRMANN, G. (1980), Einfuhrung in die Me&
and total volatile fatty acid levels in anaerobic di- technik im Chemiebetrieb, p. 257, Miinchen -
gesters, Biotechnol. Bioeng. 33, 570-577. Wien: R. Oldenbourg Verlag.
RENARD,P., DOCHAIN,D., BASTIN,G., NAVEAU, TEMPER,U., PFEIFFER,W., BISCHOFSBERGER, W.
H., NYNS, E.-J. (1988), Adaptive control of (1986), Stand und Entwicklungspotentiale der
anaerobic digestion processes - a pilot-scale ap- anaeroben Abwasserreinigung unter besonderer
plication, Biotechnol. Bioeng. 31, 287-294. Beriicksichtigung der Verhlltnisse in der Bundes-
ROBINSON,J. A., TIEDJE,J. M. (1984), Competi- republik Deutschland, Bericht aus Wassergute-
tion between sulfate-reducing and methanogenic wirtschaft und Gesundheitsingenieurwesen, Tech-
bacteria for Hz under resting and growing condi- nische Universitat Munchen.
tions, Arch. Microbiol. 137, 26-32. THAUER,R. K., JUNGERMANN, K., DECKER,K.
ROELS, J. A. (1980), Application of macroscopic (1977), Energy conservation in chemotrophic
principles to microbial metabolism, Biotechnol. anaerobic bacteria, Bacteriol. Rev. 41, 100.
Bioeng. 22, 2457-2514. TORRE,A., STEPHANOPOULOS, G. (1986), Mixed
ROELS,J. A. (1983), Energetics and Kinetics in Bio- culture model of anaerobic digestion: Applica-
technology, Amsterdam: Elsevier Biomedical tion to the evaluation of startup procedures, Bio-
Press. technol. Bioeng. 28, 1106-1 118.
SCHNECKENBURGER, H., REUTER,B. W., SCHO- TRETTER,M. (1987), Untersuchungen zur Reakti-
BERTH, S. M. (1984), Time-resolved fluorescence onstechnik des anaeroben mikrobiellen Essigsau-
microscopy for measuring specific coenzymes in reabbaus und zur Wachstumsentkopplung me-
methanogenic bacteria, Anal. Chim. Acta 163, thanogener Mikroorganismen am Beispiel von
249-255. Methanosarcina barkeri, Diplomarbeit, Institut
SCHOTT(1988), Schott-Anaerob-Testeinheit. Pro- fur Biotechnologie, Forschungszentrum Jiilich
dukt-Information Nr. 6191d, Mainz: Schott KFA .
Glaswerke. VETTER,G., CHRISTEL,W. (1988), Durchfluflkon-
SCHUGERL,K. (1985), Bioreaktionstechnik I, trolle kleiner Dosierpumpen bei stetiger und pul-
Frankfurt am Main: Sauerllnder. sierender Stromung, Chem. Ing. Tech. 60, 672-
SCHUGERL,K., LUBBERT,A., SCHEPER,T. (1987), 685.
On-line-Prozenanalyse in Bioreaktoren, Chem. WANDREY, C., AIVASIDIS, A. (1983), Zur Reakti-
Ing. Tech. 59, 701-714. onstechnik der anaeroben Fermentation, Chem.
SCHUMPE,A. (1985), Gas solubilities in biomedia, Ing. Tech. 55, 516-524.
in Biotechnology (REHM, H.-J., REED, G., WELLINGER, A. (1985), Process parameters affect-
Eds.), Vol. 2, p. 159, Weinheim: VCH. ing methane production in mesophilic farm diges-
SCHURBUSCHER, D. (1989), Rechnergestiitzte Opti- ters, Process Biochem., October, 131-137.
mierung und ProzeBftihrung kontinuierlicher WHITMORE, T. N., LLOYD,D., JONES,G., WIL-
biotechnischer Prozesse untersucht am Beispiel LIAMS,T. N. (1987), Hydrogen-dependent con-
484 14 Anaerobic Waste Water Process Models
trol of the continuous anaerobic digestion proc- YANG, S . T., OKOS, M. R. (1987), Kinetic study
ess, Appl. Microbiol. Biotechnol. 26, 383-388. and mathematical modeling of methanogenesis
WIESMANN, U. (1988), Kinetik und Reaktionstech- of acetate using pure culture of methanogens.
nik der anaeroben Abwasserreinigung, Chem. Biotechnol. Bioeng. 30, 661-667.
Ing. Tech. 60, 464-474. ZEITZ,M. (1977), Nichtlineare Beobachter fur che-
WITTY,W., MARKL, H. (1985), Reaktionstechni- mische Reaktoren, VDI-Fortschrittsberichte, Rei-
sche Aspekte der Methangiirung am Beispiel der he 8, No. 27.
Vergarung von Penicillinmycel, Chern. Ing.
Tech. MS 1400185.
15 Bioprocess Kinetics and
Modelling
of Recombinant Fermentation
DEWEYD.Y. RYU
JEONG-YOON KIM
SUN BOK LEE
Davis, California 95616, U.S.A.
1 Introduction 486
2 Bioprocess Kinetics 488
2.1 Bioprocess Modelling and Simulation 488
2.1.1 Molecular Model 488
2.1.2 Single-Cell Model 488
2.1.3 Population Model and Bioreactor Model 489
2.2 Kinetics of Gene Product Formation 490
3 Determination of Genetic Parameters Related to Productivity 491
3.1 Development of the Method - Kinetic Model 492
3.2 Application of the Method 495
4 Optimization of Recombinant Fermentation (Selected Examples) 496
4.1 Optimization of Dilution Rate in a Two-Stage Continuous Culture System 497
4.1.1 Effect of Specific Growth Rate on Plasmid Content 498
4.1.2 Effect of Specific Growth Rate on the Gene Expression Rate 499
4.1.3 Effect of Specific Growth Rate on Plasmid Stability 500
4.2 Optimization of Temperature Control in a Two-Stage Continuous Culture System 501
4.2.1 Effect of Temperature on the Gene Expression Rate 501
4.2.2 Effect of Temperature on the Plasmid-Harboring Cell Fraction 502
5 Conclusion 503
6 References 503
Biotechnology Second, Completely Revised Edition
Edited by H.-J. Rehm and G.Reed in cooperation with
A. Puhler and P. Stadler
copyright@WILEY-VCH Verlag GmbH, D-69469 Weinheim (Federal Republic of Germany). 2001
,
486 15 Bioprocess Kinetics and Modelling of Recombinant Fermentation
et al. (1974), there have been exciting develop- these areas, and they should not be underesti-
ments in the areas of pharmaceutical and med- mated; this chapter, however, deals with only
ical applications. Some of these gene products the scientific and technological aspects of bio-
are already available on the market, and others process kinetics and modelling of recombinant
are well on the way. The important front-run- fermentation.
ners among these new gene products include: Recombinant DNA technology enables us to
insulin, growth hormones, interferons, inter- identify, clone, and transform the desired
leukin-2, tissue plasminogen activator, vac- genes, and to have the host cell express them
cines, diagnostics, amino acids, and enzymes very efficiently. On the other hand, bioprocess
(WEBBER,1985; OF F I C E O F TECHNOLOGY AS- technology, if and when properly developed
SESSMENT, 1982, 1984). Significant progress and optimized, enables us to produce gene
has also been made in the areas of more tradi- products economically by making good use of
tional “engineering biotechnology” or biopro- recombinant fermentation, purification, and
cess engineering, which include fermentation bioseparation technologies. Although many
technology, enzyme technology, and biosepa- underestimated the importance of bioprocess
ration technology. technology during the early period of biotech-
Successful commercial development of gene nological development, it is now widely recog-
products requires not only recombinant DNA nized, and many of us, especially in the bio-
technology and bioprocess engineering but chemical engineering community, are now fo-
also many other strategic planning tasks re- cusing on the problems, namely scale-up prob-
lated to clinical testing, handling of regulatory lems related to recombinant fermentation and
affairs, marketing, and legal and business as- bioseparation of gene products. Fig. 1 high-
pects involved in the commercial development lights the important strategic planning task
of new products and processes. In recent years components required for the commercial de-
we have gained considerable experience in velopment of gene products.
-Gene 7
cloning Gene
expression andT-)--Fermentation
optimization
recombinant
strain
process
technology
1- large-scale
production
$-
Bioseparation
technology
and application
Fig. 1. Commercialization of gene products: devel- Approval for
opment strategy and pathway. marketing
488 15 Bioprocess Kinetics and Modelling of Recombinant Fermentation
It is also appropriate to consider the “sys- 1983; MOSER,1985) (see Fig. 2). The first two
tems engineering” approach to the biotechno- levels are microscopic process kinetics, and the
logy process optimization task, since the next two levels are macroscopic process kinet-
rDNA technique used in the “front end” of the ics; a combination of all the levels gives the
research and the bioprocess technology used in overall bioprocess kinetics. We have very little
the “rear end” of the process development are information about the first two levels of com-
highly interactive, and a well-coordinated and plexity in bioprocess systems.
concerted research and development endeavor
could bring about a very efficient commercial
development of the gene products. The selec- 2.1.1 Molecular Model
tion of host organism, gene expression system,
and excretion system, and their product sepa- Recent advances in molecular biology have
ration processes, are good examples. enhanced our understanding of the regulatory
Recently, as more and more of the second- mechanisms of microbial metabolism and have
and third-generation gene products were well enabled us to analyze them more accurately
on their way to large-scale production and than ever before. Analysis of the molecular
commercialization, the importance of biotech- model provides insight into metabolic regula-
nology process engineering and the scarcity of tions taking place at the molecular level. The
a related knowledge base came to be recog- molecular level models have been built up
nized. Once a fundamental biochemical engi- based on genetic control and regulatory mech-
neering study is established for many host-vec- anisms. For instance, the control mechanisms
tor systems of practical importance, our in- of the lac promoter-operator and the hdv plas-
depth knowledge base can be significantly ad- mid replication, which are now well under-
vanced, and a rational strategy for improving stood, have been studied in detail using molec-
and/or optimizing recombinant fermentation ular mechanism models (LEE and BAILEY,
processes can be developed for practical appli- 1984a, b, c, d).
cation. Our ultimate aim is to develop the strategy
that gives the best gene expression efficiency.
The trp or APL promoters with the autorepres-
sor control system have been found to yield
high gene expression efficiency. Some model
2 Bioprocess Kinetics simulation studies of the effect of rDNA de-
sign on the transcription efficiency indicate
that the best strategy is cloning the repressor
Once the recombinant fermentation process gene with the target gene on the plasmid and
is optimized, we must then deal with many putting the repressor formation under the au-
more genetic parameters of recombinants. torepression control system.
Thus, a deeper understanding of the recombi-
nant fermentation process and quantitative
analyses of bioprocess kinetics are necessary 2.1.2 Single-Cell Model
for optimization.
It is convenient to use the specific growth
rate as a lumped parameter representing the
2.1 Bioprocess Modelling and overall metabolic activities of the whole cell
system, although other representations can
Simulation also be used. Some attempts have been made
to develop a cellular model that relates the spe-
Our bioprocess model is organized accord- cific growth rate to the plasmid DNA and oth-
ing to five levels of complexity: (1) the molecu- er biosynthetic activities of recombinant or-
lar level model, (2) the single-cell model, (3) ganisms (LEE and BAILEY,1984e). Model sim-
the population model, (4) the bioreactor mod- ulation of hdv replication shows that the plas-
el, and (5) the bioplant model (BAILEYet al., mid DNA content decreases with an increase in
Bioprocess Kinetics 489
4. Bioreactor models
(bioprocess kinetics)
1 “Macroscopic process kinetics”
I Yield concept I
t
3. Population models Pure, mixed populations,
gene concentration distribution, and morpho-
t
2. Single-cell models “Microscopic process kinetics”
T
1, Molecular level models
Control of enzyme reactions
Genetics, gene regulation
Basic stoichiometry
the specific growth rate in accordance with the dynamics must be considered in the design and
experimental results obtained with plasmids analysis of recombinant fermentation proc-
R1, pBR322, and ColE1 (ENGBERGand esses. From the model simulation there ap-
NORDSTROM,1975; STUEBERand BUJARD, pears to be an optimal gene concentration that
1982; SIEGELand RYU, 1985) (see Fig. 3). corresponds to maximum productivity (AIBA
et al., 1982; LEE et al., 1985). The optimal
gene concentration may change according to
transcription and translation efficiency, which
2.1.3 Population Model and depend on the promoter strength, ribosome
Bioreactor Model binding site, and the specific growth rate. Al-
though the rate of gene product formation is
The recombinant fermentation bioreactor related to gene concentration, too high a gene
system contains both plasmid-harboring and concentration can put a severe metabolic strain
plasmid-free cells, and mixed cell population on the primary metabolism of the host cells.
I5 Bioprocess Kinetics and Modelling of Recombinanf Fermentaiion
Ibl
t -
-80
3
x
e
-6 2
c
0
.c
_
E
-4
c
8
c
u
-2 ;
P
L
-03
-
rn
L
d
e
c
-
c
O C
0 1 2 0 1.o 2.0
[generations h-ll u, [generations h-ll
Fig. 3. Effect of the growth rate ,u on the plasmid copy number. (a) Experimental data for R1 plasmid;
(0)plasmid, (0,A) chromosome (ENGBERG and NORDSTROM, 1975). (b) Simulation results of hdv mod-
el (BAILEYet al., 1983; LEE and BAILEY,1984~).
where
Now the product formation kinetics may be ficiency (E), a lumped parameter representing
expressed as: the combined efficiency of promoter strength,
transcription, and translation; and the growth
dP ratio ( a = p + / p - ) , the ratio of specific growth
dX'
- = ko&G,-
dt dt
+
kocG,bX+ rate of plasmid-harboring cells, p + , to that of
plasmid-free cells, p - .
The productivity of recombinant organisms
depends not only on these genetic parameters
and other genetic characteristics of recombi-
A=koEG, and B = A b (13) nants but also on the environmental and oper-
ating conditions in a bioreactor system, such
Eq. (12) is in the form of a Leudeking-Piret as medium, temperature, and dilution rate. An
equation, and the coefficients A and B show overview of the relationships between the im-
biological significance far more meaningfully portant kinetic parameters affecting the pro-
than do the empirical constants proposed by ductivity of recombinants is shown in Tab. 1.
LEUDEKING and PIRET(1959). Thus, it is important to assess the effect of op-
erating conditions on these genetic parameters
and, in turn, the effect of the genetic parame-
ters on the productivity of recombinant organ-
3 Determination
Tab. 1. An Overview of the Relationships between
of Genetic Parameters the Important Kinetic and Genetic Parameters that
Affect the Productivity of Recombinant Organisms
tion depend on the genetic information con- Gene level: Plasmid concentration (G,,), Specific
tained in the recombinant plasmid and the plasmid loss rate (O), Gene expression efficiency ( E ) ,
host-vector interaction. Thus, basic under- Gene expression kinetics, etc.
standing of those genetic parameters and their t
effects on the dynamics of a heterogeneous re- Environmental and operational parameters: temper-
combinant population is essential to the eco- ature ( T ) , dilution rate (Dl and D2), substrate con-
nomic processing of recombinant fermentation centration (&), etc.
systems.
492 15 Bioprocess Kinetics and Modelling of Recombinant Fermentation
isms when recombinant fermentation processes For a chemostat culture system with a heter-
are to be designed and optimized. ogeneous recombinant cell population, one
From the point of view of bioprocess engi- can obtain the following mass balance equa-
neering, it will be extremely useful to: (1) de- tions:
velop a methodology by which those genetic
parameters of recombinants can be accurately dX+
-= -DX++~+(I-~)X+ (14)
determined, (2) assess how the parameters are dt
affected by the operating conditions such as
temperature and dilution rate, (3) assess the ef- dX-
fect of the genetic parameters on the produc- -= -~~-+p+ex++p-x- (15)
dt
tivity of recombinants in terms of gene expres-
sion efficiency, and (4) provide a rational By introducing SZ as the ratio of plasmid-free
strategy for design, scale-up, control, and op- to plasmid-harboring cell concentrations at a
timization of large-scale recombinant fermen- given time,
tation processes.
(29)
Rearrangement of Eq. (22) gives At apparent steady state with respect to the to-
tal cell population, Eq. (28) gives:
In (SZ + A ) = B t + In ( S Z , + A ) (23)
D=paPP (30)
where A and B represent
The expression for the apparent specific
0 growth rate, papp,can be further modified to
A=- and B = A + O (24)
A+@ contain S2 and a:
The method described so far needs one numer- Eq. (32) can be solved by using the method of
ical value of specific growth rate, either p or
+
separation of variables with an appropriate in-
p - , to estimate 0 (see Eq. (19)). Since the cell itial condition, provided that a and 8 are prac-
population is continuously changing in this tically constant during a certain period of cul-
heterogeneous population, and the limiting tivation under well-defined experimental con-
substrate concentration varies with the dilution ditions. The solution of Eq. (32) with the ini-
rate, it is not easy to measure p or p - accu-
+ tial condition of SZ = Qo at t = 0 (after the cul-
rately. Thus, it is possible to estimate the value ture system reaches an apparent steady state)
of 0 only when p + or p - are almost the can be expressed as:
same.
An alternative approach is to find one more +
ln(SZ + A ’ )= B’t’ 1n(Qo+ A ’ ) (33)
equation containing two variables among the
three unknowns, provided that one of these where
parameters is easily measurable. This can be
done by incorporating a new parameter, a,the
growth ratio ( a = p + / p - ) , and measuring it (34)
from a new correlation.
When a is introduced, Eq. (18) becomes:
(35)
Tab. 2. Comparison of Parameters Found in Eq. (38) with the Corresponding Parameters Found in Eqs.
(23) and (33)
Corresponding Parameters
Eq. (38) Eq. (23) Eq. (33)
B (=A+@)
A‘(=
t t’ [ = D t - l n I)%(
B ‘ ( = i-a+ae)
Growth parameter: A =B - A B
cr-ae
A’B‘
Instability parameter: 0= AB e=-
1 +A’B‘
+
Eq. (23) In (52 + A ) = B t ln(52, + A )
Eq. (33) ln(Q+A ’) = B’t’+ ln(R,+ A ’)
Eq. (38) In (52 + () = ar + 1n(Qo+ <) (as a generalized equation)
A comparison of Eqs. (23) and (33) reveals Tab. 3. The Kinetic and Genetic Parameters that are
that the form of these equations is identical Important for Recombinant Fermentation as Deter-
and can be given in the following generalized mined According to the Method Described in Sect.
form: 3.1
4.1 Optimization of Dilution Rate crease productivity. Fig. 7 clearly shows that a
two-stage continuous culture system has ad-
in a Two-Stage Continuous Culture vantages over a single-stage continuous culture
System system. A brief discussion of how the recombi-
nant fermentation processes can be optimized
A two-stage continuous culture system has by controlling the dilution rate in a two-stage
been used to study recombinant fermentation continuous culture system is presented below.
processes (SIEGELand RYU, 1985; LEE et al., The control of the dilution rate in a contin-
1988; PARK and RYU, 1990). The system uous culture system means the control of the
shown in Fig. 6 is composed of the growth apparent specific growth rate of a heterogene-
stage and the production stage. The separation ous population of plasmid-free and plasmid-
of the growth and production stages is a design harboring cells.
strategy to decrease plasmid instability and in-
J.4
1.2
08
E -
82
04
Fig. 7. Comparison
of plasmid stability @
and product forma-
10 20 30 40 50 60 70 l l tion P (SIEGELand
tlhl RYU, 1985).
498 15 Bioprocess Kinetics and Modelling of Recombinant Fermentation
For the growth stage: ent growth rate, while the plasmid concentra-
tion under the expressed condition exhibits a
maximum.
(45) The steady-state plasmid concentrations for
both stages may be described from a plasmid
balance as described below.
For the growth stage:
IT
dXlT
and at the apparent steady state, -= 0:
dt where rpl is the rate of plasmid synthesis with-
in the cell when there is no gene expression or
under the repressed condition.
(47)
dX ~ T
and at the apparent steady state, -= 0:
dt
0 . o I
0.0 0.2 0.4 0.6 0.8 1.0 12
Specific growth rate,u ( h - 1 I
All the important parameters such as plasmid
content, plasmid stability, and gene expression Fig. 8. Effect of specific growth rate on plasmid
efficiencies are subject to the specific growth content. Symbols and W are experimental data
from the repressed stage and the expressed stage, re-
rate of host cells. It is therefore important to
spectively. Lines - and --- are from Eqs. (52) and
investigate the optimum dilution rate by (53), respectively (SIEGELand RYU, 1985; PARKet
studying the effect of the specific growth rate al., 1990).
on productivity.
-
plasmid concentration under the repressed
condition decreases with the increasing appar-
Optimization of Rec:ombinant Fermentation (Selected Examples) 499
This expression was also found compatible ciated parameters toward the product biosyn-
with a rate expression derived from a molec- thesis (b):
ular mechanism for plasmid replication. Using
this rate expression for the repressed state and qp/Gp= k o o(
~ + + b) (54)
D 1= p l ,the plasmid concentration becomes:
A two-stage continuous culture system was
employed to evaluate these kinetic parameters
accurately (SIEGELand RYU, 1985; LEE et al.,
1988). Under the apparent steady state condi-
The equation describes a hyperbolic decrease tions of a two-stage continuous culture system,
of plasmid content with growth rate. The pa- qp and p + in Eq. (54) are given:
rameters for the experimental system (PARK,
1988) were evaluated from a least squares fit as
a = 0.98 mg/g cells/h, b = 0.52 mg/g cells, and
c = 0.40 h. The plasmid concentration data for
the second stage for the condition D1=0.9,
or in terms of the plasmid concentration of
the stream coming from the first stage, where 4 and p;pp represent the dilution rate
G,, = 0.2 mg plasmid/g cell, can be estimated and apparent specific growth rate in the sec-
from Eq. (50) with a rate expression of the ond stage, and XITand X2, are the cell con-
form: centrations in the first and second stage (Fig.
6). The productivity equation is rearranged in
Eq. (57):
(53)
$ 06-
4.1.3 Effect of Specific Growth &04.
Rate on Plasmid Stability .
sr
2 0.2-
The gene-product yield is also a direct func-
tion of the plasmid-harboring cell fraction in 0.04 '
both stages. Even though the total cell concen- 0.0 0.1 0.2 0.3 0.4 0.5 0.6 On
Apparent specific growth rate &PP( h-11
tration may remain constant, because of plas-
mid instability the fraction of plasmid harbor- Fig. 11. Effect of ,u;pp on plasmid stability
ing cells will continue to decrease with time, (Y1=0.75, T 2 = 4 0 " C ,D2=0.7 h-') (PARKet al.,
and the rate of decrease depends on many fac- 1990).
tors. The stability of the plasmid-harboring
cells is usually characterized by two kinetic pa-
rameters related to growth rate and genetic
characteristics, namely, the growth ratio a and
the plasmid loss rate 8.
A mathematical analysis for the two-stage
continuous culture system has been carried out
in detail by LEE et al. (1988). According to where
their study, the fraction of a plasmid-harbor-
ing cell population at infinite time, Q2( a),be-
comes:
Optimization of Recombinant Fermentation (Selected Examples) 501
The ratio Q,2(co)/Q,i,was plotted against pipp 4.2.1 Effect of Temperature on the
in Fig. 10. Here, was varied in the range of
0.25 to 0.95 and a, and 13, were fixed at 0.5
Gene Expression Rate
and 0.05, respectively. The figure shows that
the early decrease in @z(co)/Q,l values at low Transcription starting from the APL promot-
,u;"~ is steeper as values are lower. The ex- er is induced by inactivation of the tempera-
perimental results showing the effect of pipp ture-sensitive repressor molecules at tempera-
on plasmid stability are given in Fig. 11. Here, tures above 38 "C. ACKERet al. (1982) showed
Y, the fraction of plasmid-harboring cells that the probability of repression of the AP,
among the colony-forming cells, is used in- promoter can be quantitatively related to the
stead of Q, because some X + cells lose their change of repressor concentration. Likewise,
colony-forming ability before Xfcells lose the we assume that the transcription efficiency of
ability to produce the cloned-gene product or the AP, promoter will vary from 0 to 1.0, de-
completely lose their viability (DIPASQUAN- pending on the degree of repressor inactivation
TANIO et al., 1987; SIEGEL, 1985), and it is at a constant specific growth rate. It is also as-
very difficult to measure the Xf concentration sumed that the translation efficiency is con-
accurately (PARKand RYU, 1990). Y 2 / Y ,val- stant when the specific growth rate is main-
ues at a constant value of 0.75 are comparable tained constant; hence, the gene expression
to those shown in Fig. 10. rate is limited at the level of transcription. Us-
ing the experimental data of PARK(1988), an
adequate kinetic model is established as fol-
4.2 Optimization of Temperature lows:
Control in a Two-Stage Continuous
Culture System
Induction time and inducer concentration
for gene expression were shown to be impor-
tant in optimizing the recombinant fermenta-
tion processes (BOTTERMANN et al., 1985).
When the APL promoter is controlled by the / -E\
temperature-sensitive repressor (cIw), the
K ( T )=/3 T exp
gene expression is regulated by changing the
temperature. Usually a temperature lower than (1 + exp exp -
RT
37 "Cis employed for cell growth, and a slight-
ly higher temperature for the expression of the
cloned gene product. A two-stage continuous
culture system in combination with the tem-
perature-controlled gene switching system has
been used to optimize the productivity of the
foreign protein (LEE et al., 1988; PARKand
RYU, 1990; PARKet al., 1990). This section
will deal with the optimization of temperature
control for the production of the cloned gene
product using kinetic models with parameters
evaluated from the experimental data.
38 39 40 41 42 43
T I"C 1
Fig. 12. Specific gene expression rate as a function
of temperature. Symbols are experimental data
(PARK,1988).
502 15 Bioprocess Kinetics and Modelling of Recombinant Fermentation
E
-"
L
a'
The decrease of the plasmid-harboring cell
fraction with the cell generation time due to se- 0.01J I
0 100 200
gregational instability has been characterized Timef I h l
in terms of a relative segregation rate parame-
ter 8 and the growth rate ratio a. In terms of Fig. 13. Plasmid-harboring cell fraction as a func-
these parameters the transient behavior of the tion of temperature and time. Symbols are experi-
plasmid-harboring cell fraction for the growth mental data (n 35 "C, 0 39 "C, 41 "C) (PARK,
1988).
and production stages is given by:
model for jldv plasmid replication: Analysis of within the 3’ untranslated region of the pyruvate
wild-type plasmid, Plasmid 11, 151-165. kinase messenger RNA upon its stability and
LEE, S. B., BAILEY,J. E. (1984d), A mathematical translation in Saccharomyces cerevisiae, Nucleic
model for jldv plasmid replication: Analysis of Acids Res. 15, 7951-7962.
copy number mutants, Plasmid 11, 166-177. REMAUT,E., STANSSENS, P., FIERS,W. (1983), In-
LEE, S. B., BAILEY,J. E. (1984e) Analysis of ducible high level synthesis of mature human fi-
growth rate effects on productivity of recombi- broblast interferon in Escherichia coli, Nucleic
nant Escherichia coli populations using molecu- Acids Res. 11, 4677-4688.
lar mechanism models, Biotechnol Bioeng. 26, RYU,D. D. Y., SIEGEL,R., LEE, S.B. (1985), Ki-
66-73. netics of gene expression in recombinant E. coli
LEE, S . B., SERESSIOTIS, A., BAILEY,J. E. (1985), K-12 AH1 Atrp/pPLc23trpAI, Paper presented
A kinetic model for product formation in unstab- at the Annual Meeting of Am. Inst. Chem. Eng.
le recombinant populations, Biotechnol. Bioeng. SATYAGAL, V., AGRAWAL, P. (1989), A generalized
27, 1699-1709. model of plasmid replication, Biotechnol.
LEE, S. B., RYU,D. D. Y., SIEGEL,R., PARK,S. Bioeng. 33, 1135-1144.
H. (1988), Performance of recombinant fermen- SCHOTTEL,J. L., SNINSKY, J. J., COHEN,S. N.
tation and evaluation of gene expression efficien- (1984), Effects of alterations in the translation
cy for gene product in two-stage continuous cul- control region on bacterial gene expression: Use
ture system, Biotechnol. Bioeng. 31, 805-820. of cat gene constructs transcribed from the lac
LEUDEKING, R., PIRET,E. L. (1959), A kinetic promoter as a model system, Gene 28, 177-193.
study of the lactic acid fermentation. Batch proc- SHEN,S.-H. (1984), Multiple joined genes prevent
ess at controlled pH, J. Biochem. Microbiol. product degradation in Escherichia coli, Proc.
Technol. Eng. 1, 393-412. Natl. Acad. Sci. USA 81, 4627-4631.
MOSER,A. (1985), General strategy in bioprocess- SHEPARD,H. M., YELVERTON,E., GOEDDEL,D.
ing, in: Biotechnology (REHM,H.-J., REED, G., V. (1982), Increased synthesis in E. coli of fibro-
Eds.), Vol. 2, pp. 173-197, Weinheim-Deerfield blast and leukocyte interferons through altera-
Beach/Florida-Basel: VCH. tions in ribosome binding sites, DNA 1, 125-
NILSSON,G., BELASCO, J. G., COHEN,S. N., VON 131.
GABAIN,A. (1984), Growth-rate dependent regu- SHIRAKAWA, M., TSURIMOTO, T., MATSUBARA, K.
lation of mRNA stability in Escherichia coli, Na- (1984), Plasmid vectors designed for high-effi-
ture 312, 75-77. ciency expression controlled by the portable recA
OFFICE OF TECHNOLOGY ASSESSMENT (1982), 3. promoter-operator of Escherichia coli, Gene 28,
Genetic Technology: A New Frontier, Washing- 127- 132.
ton, D. C. SEO, J.-H., BAILEY,J. E. (1986), Continuous culti-
OFFICE OF TECHNOLOGY ASSESSMENT (1984), 4 . vation of recombinant Escherichia coli: existence
Commericial Biotechnology: A n International of an optimum dilution rate for maximum plas-
Analysis, Washington, D. C. mid and gene product concentration, Biotechnol.
PALVA,I., LEHTOVAARA, P., KAARIAINEN, L., SI- Bioeng. 28, 1590-1594.
BAKOV,M., CANTELL,K., SCHEIN,C. H., KA- SINGH,A., LUGOVOY, J. M., KOHR, W. J., PERRY,
SHIWAGI, K., WEISSMANN, C. (1983), Secretion L. J. (1984), Synthesis, secretion and processing
of interferon by Bacillus subtilis, Gene 22, 229- of a-factor-interferon fusion proteins in yeast,
235. Nucleic Acids Res. 12, 8927-8938.
PARK,S. H. (1988), Ph. D. Thesis, University of SIEGEL,R. (1985), Ph. D. Thesis, University of Cal-
California at Davis. ifornia at Davis.
PARK,S. H., RYU,D. D. Y. (1990), Effect of oper- SIEGEL,R., RYU,D. D. Y. (1985), Kinetic study of
ating parameters on specific production rate of a instability of recombinant plasmid pPLc23rrpAI
cloned-gene product and performance of recom- in E. coli using two-stage continuous culture sys-
binant fermentation process, Biotechnol. Bioeng. tem, Biotechnol. Bioeng. 27, 28-33.
35, 287-295. SPANJAARD, R. A., VAN DIJK, M. C., TURION,A.
PARK,S. H., RYU, D. D. Y., KIM, J. Y. (1990), J., VAN DUIN,J. (1989), Expression of the rat
Biotechnol. Bioeng., accepted. interferon-alpha 1 gene in Escherichia coli con-
PARSELL,D. A,, SAUER,R. T. (1989), The struc- trolled by the secondary structure of the transla-
tural stability of a protein is an important deter- tion-initiation region, Gene 80, 345-351.
minant of its proteolytic susceptibility in Esche- STUEBER,D., BUJARD,H. (1982), Transcription
richia coli, J. Biol. Chem. 264, 7590-7595. from efficient promoters can interfere with plas-
PURVIS,I. J., BETTANY,A. J., LOUGHLIN,L., mid replication and diminish expression of plas-
BROWN,A. J. (1987), The effects of alterations mid specified genes, EMBO J . 1, 1399-1404.
References 505
TALMADGE, K., GILBERT,W. (1982), Cellular loca- technology firms, Chem. Eng. News, 25-60, No-
tion affects protein stability in Escherichia coli, vember 18.
Proc. Natl. Acad. Sci. USA 19, 1830-1833. YAMAKAWA,M., SUGISAKI,K., MORIMOTO,M.,
THOMPSON,R. (1982), Plasmid and phage M13 TANAKA,M., YAMAMOTO,M. (1989), Effects of
cloning vectors, in: Genetic Engineering (WIL- gene dosage on the expression of human growth
LIAMSON,R., Ed.) pp. 1-52, New York: Aca- hormone cDNA in Escherichia coli, Biochim.
demic Press. Biophys. Acta 1009, 156-160.
UHLIN,B. E., MOLIN,S., GUSTAFSSON, P., NORD- YARRANTON, G. T., WRIGHT,E., ROBINSON,M.
STROM,K. (1979), Plasmids with temperature-de- K., HUMPREYS,G. 0. (1984), Dual-origin plas-
pendent copy number for amplification of cloned mid vectors whose origin of replication is con-
genes and their products, Gene 6 , 91-106. trolled by the coliphage lambda promoter PL,
WEBBER,D. (1985), Consolidation begins for bio- Gene 28, 293-300.
IV. Control and Automation
16 Control of Bioreactor Systems
HENRYC. LIM
KYU-SUNGLEE
Irvine, California 92717, U.S.A.
1 Introduction 511
2 Low-Level Control 513
2.1 Sequence Control 514
2.2 Single-Loop Control of Environmental Variables 5 14
2.2.1 Flow Control 514
2.2.2 Gas Pressure Control 515
2.2.3 Temperature Control 515
2.2.4 Dissolved Oxygen Control 515
2.3 Controller Selection and Tuning Methods 515
2.3.1 Controller Selection 515
2.3.2 Tuning Methods 516
2.3.2.1 Ultimate Gain Method 516
2.3.2.2 Process Reaction Curve Method 517
2.3.2.3 Automatic Tuning 518
2.4 Advanced Process Control 518
2.4.1 Feedforward Control 519
2.4.2 Multiloop Control 519
2.4.2.1 Cascade Control 520
2.4.2.2 Ratio Control 520
2.4.3 Adaptive Control 520
2.4.3.1 Self-Tuning Control 521
2.4.4 Computer Control 521
2.4.4.1 Supervisory Control 522
2.4.4.2 Setpoint Control (SPC) 522
2.4.4.3 Direct Digital Control (DDC) 522
3 Optimization and High-Level Control 523
3.1 Constant Setpoint Optimization for Low-Level Control and Continuous Bioreactors 524
3.2 Profile Optimization 525
3.2.1 Batch Bioreactors 528
3.2.2 Profile Optimization for Fedbatch Bioreactors 530
510 16 Control of Bioreactor Systems
Introduction 5 11
cated tasks such as on-line identification and (an optimization problem). These objectives
adaptive control and optimization of fermen- are to be achieved under various constraints
tation processes. such as safety, environmental regulations, lim-
Until recently the fermentation industry has ited resources, and operational constraints.
lagged behind other process industries in im- Problems associated with the design of a
plementing control and optimization technolo- control system include:
gy. There are many reasons for the delay. Fer-
mentation processes are much more complex (a) What are the objectives?
than other industrial processes, involving a (b) What should be measured?
large number of complex and dynamic bio- (c) What should be manipulated?
chemical reactions and transport phenomena, (d) Which control variables should be
many of which are not well understood. There- paired with which measured output?
fore, it is very difficult to develop a realistic Should some disturbances be measured
model involving a modest number of key var- and used to manipulate control varia-
iables. In addition, on-line measurement of bles?
these key physiological and biochemical pa- (e) How should controllers (typically PID
rameters is very difficult if not impossible. controller parameters) be tuned?
Consequently, instruments able to measure (f) How does one decide the best operating
more than a few simple parameters are not conditions (constant values or variable
available. Finally, the organisms have intracel- profiles for the setpoint)?
lular regulatory mechanisms through which
they perform internal regulation. Without We best illustrate these points by consider-
proper understanding of these mechanisms we ing a specific problem, that of controlling the
can only manipulate the extracellular environ- temperature ( T ) and the dissolved oxygen con-
ment, hoping to affect intracellular mecha- centration (DO)of a fermentor at some de-
nisms in order to optimize bioreactor perform- sired values ( T d ,DOd) by manipulating the
ance. coolant flow rate (F,) and the aeration rate
The primary objectives of control systems (R,) in the presence o f external disturbances
are to provide quality assurance and economic such as changes in the inlet coolant tempera-
incentives. Slow dynamics associated with mi- ture (Ti),the oxygen concentration of the inlet
crobial bioreactors can result in off-specifica- gas (Oi), or the microbial growth rate, which
tion products over a significant period for con- can alter the oxygen demand and the tempera-
tinuous operation or after a long period for ture. Here the problem statement itself pro-
batch operations. Unforeseen disturbances in a vides the answers to the questions concerning
continuous bioreactor can result in a failure the objective and manipulation, items (a) and
(washout) of the bioreactor, requiring new (c) above. Obviously T and DO must be mea-
start-ups. Failures or off-specification prod- sured, which answers the question concerning
ucts can have catastrophic economic conse- measurement (b). The question of proper pair-
quences. Therefore, there is a strong economic ing (d) is a simple matter. Intuition tells us that
incentive for proper control and optimization the temperature ( T ) should be paired with the
of bioreactors. coolant feed rate (FJ and the DO measure-
The purpose of control is to manipulate the ment with the aeration rate (R,) so that the
control variables to: (1) maintain the desired temperature is controlled by the coolant flow
outputs at a constant desired value (a regula- rate and the dissolved oxygen concentration by
tion or a setpoint problem) by suppressing the the aeration rate. In certain cases the answer is
influence of external disturbances and/or forc- not obvious and one must resort to a systemat-
ing the outputs to follow a desired profile (a ic method such as a relative gain array method
servo or profile problem), (2) stabilize unstable to obtain a proper pairing (BRISTOL,1966).
or potentially unstable processes such as con- The next question is how to tune the control-
tinuous cultures (a stabilization problem), and lers, item (e), which is dealt with in many ref-
(3) optimize the performance as defined by erences (COUGHANOWRand KOPPEL, 1965;
measures such as yield, productivity, or profit DOUGLAS,1972; SHINSKEY,1979; COHENand
Low-Level Control 5 13
COON, 1953). The remaining question, item sired value (setpoint), and calculates the con-
(f), is how to determine the optimal operating trol signal by applying a certain algorithm to
conditions; here the desired setpoint values the error signal. This control signal then ma-
( T d ,DOd), are covered in Sect. 3. nipulates a process input (control element) to
In this chapter we cover very briefly the ba- reduce the error. The controllers are named
sic ideas of control. We then pay more atten- after the particular algorithm they use on the
tion to advanced control techniques and the error signal to generate the manipulated varia-
optimization of batch, fedbatch, and contin- ble, m. The most common controller is called
uous bioreactors. a PID controller and uses proportional (P), in-
tegral (I), and derivative (D) actions. It is satis-
factory for about 80% of control applications.
2 Low-Level Control
Measurement
--+4
tors in the fermentation industry: manual, au-
tomatic, and computer controls. While more
recent fermentation plants rely more on auto-
matic and computer controls, there are still
Setpoint
*-
- Controller Motor
Ra
Eioreactor
c*
00
implement the control action. Various types of where K,, 7[,and TD are the proportional, inte-
computer controls are used: a direct digital gral, and derivative constants, respectively.
control (DDC), a supervisory control, a set- Simpler controllers are P, PD, and PI control-
point control (SPC), and a modern control. lers that lack the derivative and integral ac-
A classical automatic control system in the tions, the integral action, and the derivative
form of feedback control is shown in Fig. 1. action, respectively. The resulting manipulated
Since classical control techniques are covered variables, m , are obtained from the above by
extensively in standard references (COUGHA- setting to zero I/T, and T,, T,, and l/rI, re-
NOWR and KOPPEL, 1965; DOUGLAS,1972; spectively. A special case of a proportional-
SHINSKEY, 1979), the treatment given here for only control (P) is on-off control. Here the
the sake of completeness is at most cursory. controller gain, K,, is infinite, but the control
Almost all feedback control systems use nega- action is restricted physically within the limits
tive feedback. The controller generates an er- of the control element. On-off controllers are
ror signal, e , by subtracting the measured usually used with on-off control elements
process output (controlled variable) from a de- (e.g., solenoid valves, fixed-speed pumps, con-
514 16 Control of Bioreactor Systems
stant-load heaters, etc.). This type of control is ture for the desired period. When the desired
satisfactory when the changes in the controlled time has lapsed, the timer triggers the next
variables are small and vary slowly with time. process. Cooling the fermentor to the desired
In many fermentation processes the tempera- temperature is an event that can trigger the
ture and pH loops have these characteristics, opening of a valve for inoculation.
and thus on-off control is adequate. Although With the advent of inexpensive and compact
most recent applications of PID controllers microprocessors, mechanical devices such as
utilize analog electronics or digital computers, cam controllers and timers formerly used to
there are many pneumatic controllers in opera- control sequential operations have been re-
tion. This is due to two factors. In the .past, placed with a modern version of sequence con-
pneumatic instrumentation was the mainstay trol that involves programmable logic systems
of the process industry. Due to their rugged- and computers. Using hardware or software,
ness and reliability, many existing installations various basic discrete control functions (direct
still have pneumatic controllers. Second, pneu- on/off control, interlock control, sequential
matic valves have many desirable characteris- control) are developed to control discrete de-
tics, and they are still specified for some new vices such as on-off valves, pumps, or agita-
installations. tors, based on the status (ON-OFF) of equip-
ment or values of process variables.
A discrete device can have only two states,
2.1 Sequence Control such as ON/OFF, OPEN/CLOSE, and
ENABLED/DISABLED. The triggers can be
Since batch and fedbatch processes vary time-based or event-based (temperature, pH,
with time, the controller system must deal with pressure, level, etc.). Sequencing requires
time- and event-based process conditions and reaching an end condition (trigger) before the
transition phenomena. In batch and fedbatch system can proceed to the next step. In sequen-
fermentations there are many sequential oper- tial control one trigger or a combination of
ations that need to be followed in each run. In triggers may decide a step transition.
a batch fermentation the bioreactor must first Improved sequencing can result in signifi-
be cleaned, filled with the fermentation me- cant economic gains. Since the sequencing op-
dium, sterilized, brought to proper fermenta- erations may be done discretely, computeriza-
tion conditions, inoculated with the inoculum, tion is an excellent method for their attain-
allowed to ferment for a specified time or to ment. The time base for the sequence is more
an end event, and emptied. The starting of accurate and does not need to be calibrated
pumps, the opening of valves, the completion with a computer. The time base and the logic
of charging the bioreactor with medium and can also be readily changed by computer.
inoculum, and completion of fermentation are
all events in time, requiring different control
actions. Sometimes the control system includes 2.2 Single-Loop Control
automatic start-up and shut-down proce-
dures. of Environmental Variables
Sequential control includes all control func-
tions (discrete and continuous/regulatory) trig- Examples of some common single-loop con-
gered by time or events. In sterilization, for trols, such as flow control, pressure control,
example, the fermentor may be brought to a temperature control, and dissolved oxygen
desired temperature, maintained there for a control, are given below.
specified period, and then cooled to a desired
fermentation temperature. Clearly the steam
valve must initially be opened and a tempera- 2.2.1 Flow Control
ture controller must operate to reach the de-
sired temperature. When the latter is reached, Control loops for flow of gas or liquid are
or when the desired event takes place, the tim- characterized by fast response times (seconds)
er is then triggered to maintain the tempera- with essentially no delays. Most disturbances
Low-Level Control 5 15
result from high-frequency noise, though gen- 2.3 Controller Selection and Tuning
erally not of large magnitude, due to stream
turbulence, valve changes, and pump vibra- Methods
tions. Proportional-Integral (PI) flow control-
lers are generally used. The presence of recur- After determination of types of control and
ring high-frequency noise rules out the use of location of controllers, it is necessary to spe-
derivative action. cify the modes of control and the parameter
values of controller settings (Kc, T], and rD)
necessary to obtain a satisfactory response.
2.2.2 Gas Pressure Control Given below are the general characteristics of
various controller modes and some practicable
It is relatively easy to control gas pressure, tuning methods used in industry.
because the gas-pressure process is self-regulat-
ing unless the gas is in equilibrium with a
liquid. P I controllers are normally used with a 2.3.1 Controller Selection
small amount of integral control action. Due
to the short response time compared with oth- No fixed criteria exist to determine which
er process operation times, derivative action is modes to use in a specific application. Fre-
normally not used. quently a PID controller is selected simply be-
cause it is most likely to accomplish satisfacto-
ry control. But it is beneficial to note through
2.2.3 Temperature Control a typical process response to load changes the
following general characteristics (COUGHA-
Due to the variety of heat transfer equip- NOWR and KOPPEL, 1965):
ment, and because processes differ significant-
ly in time scales, general guidelines for temper- Proportional control results in a re-
ature control loops cannot be provided. The sponse with a maximum offset, a high
presence of time delays and/or multiple ther- maximum deviation, and a moderate pe-
mal capacitances (high-order process with or riod of oscillation that ceases gradually.
without dead time) leads to an upper limit on Proportional-integral control results in
controller gain. PID controllers are commonly n o offset, but at the expense of a high-
used to provide more rapid responses than er maximum deviation, a longer period
those obtainable with P I controllers. of oscillation, and a longer time re-
quired for oscillations to cease than in
the case of proportional control.
2.2.4 Dissolved Oxygen Control Proportional-derivative control general-
ly results in the shortest time to steady
Dissolved oxygen can be regulated by aera- state with the least oscillation and the
tion rate, agitator speed, or a combination of smallest maximum deviation, but at the
the two. In industrial-scale fermentors agitator expense of an offset which, though
speed may be fixed, in which case the only smaller than that for the proportional
means of controlling the dissolved oxygen con- control, is still significant.
centration is by varying the aeration rate. A P I Proportional-integral-derivativecontrol
controller may be used for this purpose. In re- may result in a combination of the ad-
search-scale bioreactors dissolved oxygen is vantages of proportional-integral and
often regulated by the cascaded control system proportional-derivative controls. The
that is discussed in Sect. 2.4.2. integral action and the derivative action
serve to eliminate the offset and lower
the maximum deviation, respectively.
The derivative action also eliminates
some oscillations that occur in PI con-
trol.
516 16 Control of Bioreactor Systems
It is important to emphasize, however, that tion curve method uses an open-loop test to
the above observations do not always apply obtain process dynamic information.
and that exceptions occur. Nevertheless, the
basic characteristics attributed to each mode
of control generally persist even in a complex 2.3.2.1 Ultimate Gain Method
system, and therefore selection of a particular
mode of controller should be guided by the First proposed by ZIEGLER and NICHOLS
above observations. (1942), this has been called the loop tuning or
the continuous cycling method. The controller
is operated in a closed-loop with the system to
2.3.2 Tuning Methods be controlled. The first step is to determine ex-
perimentally the ultimate gain by using the
Controller tuning is the adjustment of con- proportional-only controller and increasing K ,
troller settings to obtain a satisfactory per- by small increments until continuous cycling
formance (response) of the control system. Al- of the system variables first occurs. This value
though the manufacturers of fermentor instru- of K, is called the ultimate gain and is denoted
mentation packages usually supply their con- by K,. The period of the resulting sustained
troller's settings, on-site tuning is normally re- oscillation is called the ultimate period, P,.
quired to meet individual requirements. The controller settings are then calculated
Acceptable overshoot, settling time, or stea- from K, and P, using the Ziegler-Nichols
dy-state oscillation depend on the overall proc- (Z-N) tuning rules (the original settings in
ess dynamics, user requirements, and the ob- Tab. l), which were empirically developed to
jectives of the process. For example, in tem- provide a quarter-decay ratio. These tuning
perature control it may be desirable to operate methods were commonly used in industry and
the process at or near the optimum tempera- served as a basis for comparing other control
ture for cell growth and/or metabolite produc- schemes. Although the original Z-N settings
tion. This is characteristically near the maxi-
mum temperature for microorganisms. In this
situation an overshoot of even a few degrees Tab. 1. Controller Settings Based on the Ultimate
above the setpoint can be disastrous. Also, a Gain Method
significant deviation of pH from the setpoint (ZIEGLERand NICHOLS, 1942)
can lead to a disastrous result. On the other
hand, the requirement for dissolved oxygen Controller Kc 51 TD
Controller K, 71 TD
P
--(1+2)
1 7
K Td
PI
518 16 Control of Bioreactor Systems
in which simple single-feedback control is in- measurable, and the effect of the disturbances
adequate, however, and more advanced con- must be known a priori. In addition, feedfor-
trol is required. More advanced control can ward control scheme is rarely used alone, but
improve the quality of control over that is combined with feedback control, because
achievable by simple feedback control. Ad- without the feedback control there is no way
vanced controls frequently used in industry are to correct errors caused by imperfect knowl-
discussed here. edge in predicting the effect of disturbances on
the output. A combination of feedback and
feedforward control is shown in Fig. 3. This
2.4.1 Feedforward Control permits the major effect of the disturbance to
be corrected by the feedforward controller
In this scheme (Fig. 2), unlike in normal leaving the fine trimming to the feedback
feedback control, one does not wait until a dis- loop.
turbance actually affects the output. One in-
stead measures the disturbances (TI, Oi) and
applies corrective control actions (Fc, R,) in 2.4.2 Multiloop Controls
anticipation of the expected effect. This is a
better control for eliminating the effect of dis- In practice, most processes involve more
turbance. However, to implement a feedfor- than one input (manipulated variables) and
ward control scheme, the disturbances must be more than one output (measured variables).
Fig. 1 depicts a two-input-two-output system.
The outputs (T, DO)as affected by the distur-
....................................................... bances (Ti, Oi) are measured and fed back to a
comparator where they are compared to the
Measurement = desired setpoints. The differences ( E ) between
the desired and measured variables are used to
Controller Motor
ya- I DO generate controller signals that in turn ma-
nipulate the control variables (Fc, R,) to force
Bioreactor the outputs to match the desired value. In this
T
Controller Valve situation coolant flow rate, F,, affects primari-
ly temperature, T , and the aeration rate, R,,
Measurement :
affects primarily dissolved oxygen. Thus, there
is very little interaction, and single-loop tech-
....................................................... niques (COHEN and COON, 1953; ZIEGLER
Fig. 2. Feedforward control. and NICHOLS,1942) can be applied to the sep-
arate loops. In other situations any one input
can significantly affect more than one output;
therefore, the interaction is significant and the
Measurement usual single-loop techniques cannot be used.
I fiMeasurement I_( I 0;
' 1 Sometimes even the problem of proper pairing
of the output to input variables cannot be re-
solved intuitively as we have done in Fig. 1. A
relative gain array method (BRISTOL,1966)
and a singular value analysis (SMITH et al.,
1981; MOORE,1986) are used to determine the
best pairing. Proper pairing and tuning for
multiple-input-multiple-output systems are
found in some references (SEBORGet al.,
1989).
“best” for DO control at low cell concentra- troller settings in the feedback control are au-
tions, for example, may need to be altered at tomatically adjusted by the loop optimizer.
high cell concentrations. As shown in Fig. 6, a This type of controller is called self-tuning or
typical adaptive control system has two major self-adaptive. In most real-time parameter-esti-
elements, one to identify the changes (process mation schemes an external forcing function
identification) from the input and output data, that excites the process most effectively is in-
troduced to obtain accurate process model pa-
Disturbances rameters.
whether to control with cooling or heating. It where A T is the sample time (controller time
would be convenient to design the low-level increment), m [nA TI and E [nA T] are the cur-
controller so that a positive-valued setpoint rent manipulated variable and error, respec-
would effect control by heating and a negative- tively, and E [(n- i) A TI are previous errors.
valued setpoint would control by cooling. This is known as the position algorithm, since
When it would be desirable to have the bio- it calculates the position of the manipulated
Optimization and High-Level Control 523
variable (control element). A more convenient ning. The evolution of microcomputers and
and therefore more frequently used form is the hierarchical computer systems has been a tre-
so-called velocity algorithm, which calculates mendous help in this area. Low-level control
the incremental change in the manipulated can be carried out by microcomputers while
variable, higher computers do other tasks more effi-
ciently. Also, in multi-computer systems, the
m [ n A TI = K ,
AT
i
E [ n A TI - ~ [ ( -
n 1 ) A TI + burden of computer failure is reduced because
failure of one computer will not cripple the en-
+-
TI
E [nTI + -( E [nA TI -
TD
AT
tire process or plant.
\
+ E [(n- 2) A TI)
- 2 E [(n- 1) A TI
J (7)
balance equations with all necessary kinetic ing n algebraic equations for the n unknowns
rate expressions having kinetic parameters as must be solved for optimal operating condi-
functions of T and pH. If no model relating tions:
the performance index to the operating varia-
bles is available, or if the dependence of any ap(x)/ax,=O, for i = l , 2, 3, ... n (8)
one of the kinetic parameters on T and pH is
missing, the above-described analytical meth- When a quantitative model relating the cel-
od cannot be applied. An experimental search lular productivity ( Dx) of the continuous cul-
technique with on-line optimization must then ture to these operating variables, T , pH, and
be used to obtain the solution. This type of D , is known ( D x ( T ,pH, D)) it is simple to ob-
problem is considered in detail in Sect. 5.1. tain the best steady-state values:
Trajectory optimization refers to the deter-
mination of temporal or spatial functions, not a(Dx)/aT=O, a(Dx)/a(pH) = 0 ,
constant values, which lend an optimal value a(Dx)/aD=O (9)
to the given performance index. For example,
the determination of pH or temperature pro- which must be solved simultaneously to obtain
file as a function of time during fermentation the optimum operating temperature ( T ) , pH,
to maximize the product concentration at the and dilution rate (0).
end of batch or fedbatch fermentations is a For simplicity we consider as an example
trajectory optimization. There is no obvious the determination of the dilution rate that best
reason to expect that temperature or pH maximizes the cellular productivity, D x , of a
should be kept constant for batch or fedbatch continuous culture. Consider an ideal contin-
bioreactors; what is best for cell growth may uous bioreactor that produces cell mass as its
not be best for product formation. The per- product. The steady-state cell mass and sub-
formance index is a functional, a function of strate balance equations are:
variables that themselves are also functions.
Therefore, a trajectory optimization that re-
quires determination of functions that opti-
mize the performance index is a problem in the
so-called variational calculus.
where D is the dilution rate, s and x are the
substrate and cell mass concentrations, respec-
3.1 Constant Setpoint Optimization tively, sF is the feed substrate concentration,
for Low-Level Control and and Y is the yield factor. First, we must ex-
press the performance index, D x , solely in
Continuous Bioreactors terms of D before we take the total derivative
of D x with respect to D and set it to zero.
As stated above, constant setpoint optimiza- From Eqs. (10) and (11) we have:
tion is concerned with finding the best set
of constant values of the operating variables.
For the above example the optimal constant
temperature, pH, dilution rate, and so on in which s must be eliminated in terms of D by
are: x = ( T , pH, D)T= (x1,x2,x3, . . . xJT,which solving Eq. (lo), p(s) = D , for s. Denoting this
lead to the optimum value of the given per- by:
formance index, p , that in turn depends
on many operating variables, p ( x )= D x = S= b]-'D (13)
p ( T , pH, D). This is a problem in ordinary cal-
culus. To obtain the solution the performance so that for the Monod model, (u=pmaxs/
index must be expressed in terms of the operat- (K+s), s = ~ ] - ' D = D K / & , ~ ~ - D and
) , Eq.
ing variables, all the partial derivatives of the (12) is now written as:
performance index with respect to each operat-
ing variable must be set to zero, and the result- D X = Y D ( ~-fD K/(umax- 0)) (14)
Optimization and High-Level Control 525
and d (Dx)/dD = 0 yields: TRYAGIN et al. (1962) and singular control the-
ory and seek solutions by their use.
Do,, = pmax[ 1 - ( K 1 / 2 / ( +
K SF)] I 2 ] (15) Briefly, the essence of the Maximum Princi-
ple is as follows. Consider an ideal bioreactor
The optimum dilution rate given by Eq. (15) that can be described by four unsteady-state
results in maximum cell-mass productivity. mass balance equations for cell mass, sub-
It is now desired to obtain the optimum strate@), product, and overall mass:
temperature that results in maximum produc-
tivity. This is obtained by setting to zero the
partial derivative of D x with respect to T,
a(Dx)/aT = O :
and withdrawal (Fo)rates as functions of time. where the adjoint vector A must satisfy the fol-
The fermentation time, tf, may be assumed to lowing ordinary differential equations:
be given (fixed) or open to determination
(free). Thus the objective is to maximize the
performance index P by choosing optimally
the temperature, pH, and feed and withdrawal
rates as functions of time: with the final conditions depending on the
functional form of P and given by:
Max =P[x,,P f ,Sfl (21)
T ( 0 ,PH(0,F(O,Fo(O
basis of Eq. (31) or (32). This type of problem To solve this equation for @(tf; 0) we must
is called a two-point split boundary-value solve the n2 equations in Eq. (34), starting at
problem in the sense that for the state vector, t = t f ,with the final condition @(tf; tf) = I , and
y , the initial conditions are known, while for integrating backward in time to t = 0.
the adjoint vector, A, the final conditions are It is now largely accepted that the boundary
known. There are two numerical techniques condition iteration procedure outlined above is
that can be used to solve this problem: a generally inferior to the control vector itera-
boundary condition iteration and a control tion procedure described below. There are ex-
vector iteration. ceptions, however. There are many reasons for
Boundary condition iteration refers to a this statement, the main one being that the
technique in which one set of missing bound- backward integration of generally stable state
ary conditions is estimated and iterated until equations results in an unstable system, so that
the calculated boundary conditions agree with a small error in the estimation of the final con-
the specified boundary conditions. For exam- ditions on the state vector can lead to a very
ple, the final conditions on the state vector, large discrepancy between the calculated and
y'(tf), are first estimated, and then the state given initial state vector values. This makes the
and adjoint vector differential equations, Eq. conversion very slow.
(25) and (28), are integrated backward starting The.contro1 vector iteration starts with an
with the estimated final conditions on the state estimated control vector, for, e.g., pH and
vector and the specified final conditions on the temperature as functions of time,
adjoint vector, Eq. (29), with the control vec- m ( t )= (pH, 7'). Using the estimated control
tor ( T , pH) selected by the necessary condi- vector, m'(t), and the known initial condi-
tions, Eq. (31) or (32). This backward integra- tions, Eq. (26), the state equations, Eq. (25),
tion is continued until the initial time ( t = O ) , are integrated forward in time from t=O to
and the resulting calculated initial conditions t = t f ,while the adjoint equations, Eq. (28), are
on the state vector, y'(O), are compared with integrated backward from t = t f to t = O using
the given initial conditions, yo, Eq. (26). If the the specified final conditions, Eq. (29). Then
calculated initial conditions do not agree with the Hamiltonian function, H i , Eq. (27), is
those given, a new set of final conditions on evaluated, and a new improved control vector
the state vector, yi"(tf), is proposed and the is calculated using the necessary conditions,
procedure is repeated. An iterative procedure Eq. (31) or (32). The process is repeated with
is applied until the calculated initial conditions the improved control vector until there is negli-
on the state vector agree with the given initial gible change in the Hamiltonian and therefore
conditions. Then the resulting control vector also negligible change in the control vector or
( T , pH) profiles are the optimal profiles that the performance index. The question of how
maximize the Hamiltonian function, Eq. (30), to obtain an improved control vector for the
and the performance index, Eq. (21). The iteration is provided by the following equa-
problem of how to improve on the final condi- tion:
tions after an unsuccessful estimation is now
considered. This is based on the variational aH i
m'+'(t)=m'(t)+w(t)- (35)
equations for Fy. It can be shown (KOPPEL, am
1968) that:
where W ( t )is a positive definite weighting ma-
Y i + l ( t f ) = y i ( t f ) + @(tf; 0 ) b 0 - ~ ' ( 0 ~ 1 (33) trix. If the weighting matrix, W ( t ) , is chosen
to be as follows, the method is known as the
where @(tf; 0) is the so-called transition matrix steepest ascent:
starting at t = O and ending at t = t f , which G - I (t)6s
must satisfy the following differential equa- W ( t )= (36)
tions:
[i(g) (g) '"
'G-' dt]
trix defining the distance that is the finite dis- Optimization of Feed Flow Rates, F(t), for
tance by which m is to be moved: Continuous Bioreactors
ff
For steady-state continuous bioreactors the
(6s)'= (Sm)TG(z)6md z (37) feed flow rate and the withdrawal rate are
0
equal, F ( t )= Fo(t). The performance index
This steepest ascent method has been found ef- must here be defined. If a product other than
fective for a large class of problems encoun- cells is involved, the volumetric productivity
tered in profile optimization. may be chosen as the performance index, D p .
In addition to the flow rate, the steady state
temperature and pH also may be optimized.
Optimization of Feed Flow Rates, F(t), for The details of this optimization scheme are
Fedbatch Bioreactors given in Sect. 3.2.2 and are therefore not re-
peated here.
We now consider maximizing the perform-
ance index by manipulating the flow rate, F(t),
for fedbatch bioreactors (Fo(t)=O). The feed 3.2.1 Batch Bioreactors
flow rate F ( t ) appears linear so that maximiza-
tion of the Hamiltonian depends on the sign of Assuming that the medium composition has
the coefficient of F, ATa. That is, if ATa is po- already been optimized, the variables that can
sitive we choose the maximum flow rate and if be manipulated to optimize the performance
it is negative we take the smallest F (F=0, or a of batch bioreactors include temperature, pH,
batch period). But, if a'A is identically zero initial inoculum size, initial substrate concen-
over a finite period of time, the Maximum tration, and fermentation time. Of these, only
Principle fails to yield a solution. This time pe- temperature and p H are functions of time, and
riod is called the singular interval, and the feed the rest are constant. The basics involved in
rate is called the singular control (intermediate optimizing temperature and p H profiles have
values). Therefore, the optimal feed rate pro- been presented above. Here we give an exam-
file consists of periods of maximum flow rate ple of temperature profile optimization re-
(F,,,), minimum flow rate (F=O,or a batch ported by CONSTANTINIDES et al. (1970a, b)
period), and intermediate flow rate, Fs: for batch penicillin fermentation in the context
of the Maximum Principle given above. In
their work optimal temperature profiles were
determined for a batch penicillin fermentation
(38) model.
The exact sequence and the times at which the Optimum Temperature Profile for a Batch
flow rate shifts from one period to another are Penicillin Fermentation Model
yet to be determined. Since ATa is zero over
the finite time interval, its time derivatives (the The following model is based on the mass
first, second, and so on) also must vanish. In balance equations of the cell and penicillin
general, for fedbatch bioreactors the second- while ignoring the substrate concentration ef-
order derivative expression results in a term fect altogether:
that contains F(t) explicitly, so that ATa = 0
and the first two derivatives are sufficient to X= k , [l - ( x / ~ , ) ] x (39)
allow the determination of the feed flow rate
during the singular interval in terms of the
state and adjoint vectors, y ( t ) and A(t). Since
the details of problem formulation, necessary where x and p represent cell mass concentra-
conditions, and numerical examples are given tion and penicillin concentration, respectively,
in Sect. 3.2.2 we shall not go into them here. and k l , k2, k3, and k4 are empirical constants
Optimization and High-Level Control 529
whose temperature dependencies are given be- The necessary condition for maximum given
low: by Eq. (30) is:
where T is temperature in "C and the numeri- i1= -A1 k1+2Al(kl/kl)~-A3k3A l ( t f ) = O (47)
cal values for constants c1 through cg are re-
ported elsewhere (CONSTANTINIDES et al., = A 2 k4 A 2 V f ) = 1 (48)
1970b). It should be stated here that most pen-
icillin fermentations are carried out in fed- If decreasing cell mass is not allowed, a con-
batch culture, in which the substrate feed rate straint is imposed on cell mass concentration:
is varied with time. Assuming that for a parti-
cular strains the model proposed is valid, we
proceed to look at the details involved.
The value of k2 is equivalent to the cell concen-
tration at infinite time, i.e., the maximum
v)
growth for constant temperature. This con-
straint implies that the net rate of cell forma-
tion becomes zero when the temperature drops
below a certain level. When the constraint is
-- -
4
reached, Eq. (39) becomes:
.-
c
x=O whenxrk, (50)
n -
The Hamiltonian then changes to:
H = A 2 k 3 ~ - A 2k4p (51)
cell concentration. When the constraint was tion and induction and repression, which lead
reached, i.e., when the rate of cell formation to unimodal rate expressions. Some chemical
reached zero, the optimal temperature shifted reactions also lead to unimodal rate expres-
rapidly to a lower level, approximately 18 "C, sions; these are autocatalytic, adiabatic exo-
maximizing the net rate of penicillin formation thermal, and Langmuir-Hinshelwood type ca-
(the difference between the rate of formation talytic reactions. Such reactions have not been
and the rate of degradation). After that, the subjected to thorough optimization in terms of
temperature decreased gradually, remaining reactor operations - batch, continuous, or
below 18 "C. The final penicillin potency was semibatch operations - until recently (WAGH-
76.6% higher than that obtained with the best MARE and LIM, 1981).
constant temperature of 25 "C. Isothermal reactor operations for simple
reactions have been optimized (WAGHMARE
and LIM, 1981) by optimal control theory. It
3.2.2 Profile Optimization has been shown that whenever the rate expres-
sion goes through a maximum, i.e., the rate
for Fedbatch Bioreactors expression is a non-monotonic (unimodal)
function of the reactant, a semibatch reactor
Early in the 20th century it was found that (a variable-volume batch reactor with a pro-
veast Droduction was maximized by adding grammed feed, i.e., a so-called fedbatch cul-
wort a; intervals of time rather than all at once ture) is likely to outperform either a contin-
(WHITAKER,1980). Since then, many indus- uous reactor or a batch reactor. In other
trially important fermentation processes have words, if the specific rate is a non-monotonic
been carried out in a semibatch manner. Alco- function of the substrate concentration, a fed-
hols, amino acids, antibiotics, enzymes, micro- batch bioreactor may lead to a better fermen-
bial cells, organic acids, vitamins, and various tation result than that achievable with a batch
recombinant cell products are among the prod- bioreactor or a continuous bioreactor. Thus,
ucts for which semibatch operations have been one should fully explore the possibility of us-
used or tested (MODAKet al., 1986). In a typ- ing fedbatch cultures and apply optimization
ical semibatch mode of operation, the so- theory as developed in this chapter to deter-
called fedbatch operation, the nutrients neces- mine the best feed-rate profile as a function of
sary for cell growth, the precursors for prod- time. A logical deduction from this is to anti-
uct formation, or the inducers are fed intermit- cipate that whenever there are two opposing
tently, continuously, or in a lump during an effects such as activation and inhibition or in-
otherwise batch operation, and the fermenta- duction and repression, a semibatch operation
tion broth is harvested either fully or partially (or fedbatch) is a prime candidate to be con-
at the end. The whole process may be repeated sidered for optimum yield or productivity.
either with a fresh inoculum when the harvest This anticipation had been proven correct. It
is complete, or with the remaining cells acting so happens that the general category of auto-
as the inoculum for the next cycle when the catalytic reactions, of which fermentation in-
harvesting is partially done. volving free cells is a classic example, is also
Fedbatch operation has been found particu- appropriate for semibatch operations. The
larly effective for processes in which effects problem is then the determination of the opti-
such as substrate inhibition, catabolite repres- mum feed rate of substrate as a function of
sion, product inhibition, glucose effects, and time, which optimizes the given performance
auxotrophic mutation are important (MODAK index such as productivity, yield, or profit.
et al., 1986). These phenomena lead to unimo- The nutrient limiting the growth of cells in a
dal reaction rate expressions that exhibit a fedbatch bioreactor provides an excellent
maximum with respect to a single reactant con- means of controlling the growth rate and the
centration or in terms of two or more reactant metabolism of the cell. Thus, fedbatch bio-
concentrations. A living cell possesses a com- reactors may be operated in a variety of ways
plex internal control system involving oppos- by regulating the feed rate in a predetermined
ing phenomena such as activation and inhibi- manner (feedforward control) or using a feed-
Optimization and High-Level Control 53 1
where F, is an intermediate flow rate yet to be Since Eq. (70) represents three linear equations
determined in the finite interval t, 5 t 5 t, + in five adjoint variables, only two adjoint var-
over which 6' is identically zero. All that is iables need be determined to completely de-
known at this point is that the optimal profile scribe the remaining three. Although the infor-
consists of periods of maximum flow rate mation obtained is not sufficient to deduce the
(Fmax), of minimum flow rate (F=O, or a optimal time profile, use of Eq. (70) considera-
batch period), and of intermediate flow rate bly reduces the computational burden. Analy-
(F,). The exact sequence and timing of flow- sis involving asymptotic behavior and limiting
rate shifts from one period to another are yet cases directly amenable to analytical solutions
to be determined. The interval over which 6' is allow deductions to general situations. Here
identically zero is known as the singular inter- we will briefly present the general characteris-
val and the control as the singular control, F,. tics of optimum feed-rate profiles and compu-
Since e(t) is zero over the finite interval, its tational schemes to obtain them. Details are
time derivatives (the first, second, and so on) available elsewhere (MODAKet al., 1986; LIM
also must vanish: et al., 1986).
ations is used to provide inoculum for large the flow rate sequence just described does not
bioreactors. maintain the substrate concentration constant
Finally, there is the purely theoretical situa- at the value corresponding to the maximum
tion in which the amount of inoculum is very specific growth rate, but results instead in vari-
large and the initial substrate concentration is able substrate concentrations that maximize
low (case (d) in Fig. 9). For this case the opti- the total product formation rate, T C X V .As with
mal profile can miss a batch period after the type I when the initial substrate concentration
period of maximum flow, resulting in a period is high (case (b)), the initial period of maxi-
of maximum flow rate followed by a period of mum flow rate disappears, and the optimal
singular control and then a batch period, bang profile consists of a batch period followed by a
(maximum)-singular batch. period of maximum flow rate and a batch peri-
od, bang (maximum)-singular-batch. In an
ideal situation in which the amount of inocu-
Type II. Specific Growth Rate Exhibiting a lum and the initial substrate concentration are
Maximum and Monotonically Increasing Spe- chosen just right (case (c)), the optimum feed
cific Product Formation profile begins with a period of singular flow
rate followed by a batch period.
A less common situation than type I is that
in which the specific growth rate exhibits a
maximum while the specific product formation Type III. Both Specific Growth and Product
rate increases with substrate concentration. Formation Rates Exhibiting Maxima
Reports suggesting this may be the case include
glutamic acid fermentation on ethanol and Since p and TC both show maxima, the feed-
vitamin BI2 fermentation. Single-cell protein rate profile must take advantage of this fact.
production involving microbial cell mass is a This is the least common type of fermentation.
limited version of this type, since the product An example is ethanol fermentation from fruc-
itself is cell mass. The initial conditions ((a) tose. Typical profiles are shown in Fig. 11.
through (d) in Fig. 10) dictate the optimal pro- When the initial amount of inoculum and the
file sequences, as we have seen above. When substrate concentrations are low (case (a)), the
the amount of inoculum and the initial sub- feed profile resembles that of case (d), type I-a
strate concentrations are small, a situation de- period of maximum flow rate followed by a
noted by case (a) in Fig. 10, the optimal profile period of singular flow rate and a batch peri-
consists of a period of maximum flow fol- od, bang (maximum)-singular-batch. A physi-
lowed by a period of singular flow, a period of cal interpretation can be envisioned by consid-
maximum flow, and a batch period. A simple ering a limiting case in which both p and TC
physical explanation can be provided here. show their maxima at the same substrate con-
Since the specific product formation rate is a centration and the yield coefficients are also
monotonically increasing function of substrate constant. Here one should apply the maximum
concentration, whereas specific growth first in- flow rate so that the substrate concentration in
creases and then decreases with substrate con- the bioreactor reaches in a minimum time the
centration, going through a maximum, all that value that maximizes the specific rates; one
is needed to maximize the total product forma- should then apply a singular feed rate to main-
tion rate, T C X V , is to force the substrate con- tain the substrate concentration at this level
centration to reach the value that maximizes p until the reactor is full. After that a batch op-
and keeps it there as long as possible. This is eration should be continued until it is no long-
accomplished by applying the feed at the maxi- er economical, i.e., the maintenance cost out-
mum rate and then switching the singular flow weighs the profit realized by converting the re-
rate to maintain the substrate concentration at maining substrate to product. Apparently,
a value corresponding to the maximum specif- when the peaks in the specific rates do not
ic growth rate. This physical interpretation is coincide, and when the yield coefficients are
correct if the yield coefficients are constant. not constant but vary with substrate concen-
When the yield coefficients are not constant, tration, the singular feed rate maximizes total
536 16 Control of Bioreactor Systems
7 . Compare the calculated adjoint varia- a period of maximum feed rate, t l =O. There-
bles at the final time, A3(tf)and A,(?,), fore, we can set tl = 0 in step 1 and skip steps 2
with the specified values, Eq. (64), and 11. For the situation shown in Fig. 9(c), in
A 3 (tf) = l/tf, and As ( t f )= -pf/t:. which initial conditions are such that the proc-
8. If the calculated values do not agree ess is on the singular arc, we can set tl = t2 = 0
with the specified values, improve the and skip 2, 3, 10, and 11. For Fig. 9(d) we can
estimated values of A3 (t,) and As (t,) by set t, = tl and skip steps 3 and 10.
a Newton-Raphson method and go to
step 5.
9. If there is agreement between the cal- Type 11Fermentation Processes
culated and the specified values of
A3(t2)and A,(?,), store the values of the We will develop a computational algorithm
switching times, tl and t2, and the cor- for the general profile given in Fig. 10(a). Al-
responding value of the performance gorithms for other profiles are then obtained
index, g [x(tf)]. by modifying that for Fig. lO(a). As in type I
10. Change or increment t2 by A t , , and if fermentation we will use as the performance
the entire range of t, has not been cov- index the maximization of product productivi-
ered, go to step 2. If covered, go to ty with a free final time.
step 11.
11. Change or increment tl by A t , , and if 1. Choose tl and t,, tl ct,.
the entire range of tl has not been cov- 2. Integrate forward the state equation
ered, go to step 2. If covered, go to with the given initial conditions, Eq.
step 12. (60), using F= F,, until t = tl.
12. After sets of switching times have been 3. Estimate the unknown variables at tl,
tried, choose the one that yields the A 3 (?I>, and ?S ( t l )= A s (0.
maximum value, or improve switching 4. Integrate, A3 equation, Eq. (63), and
times by further narrowing the range. the state equations, Eq. (60), forward
from tl to t2 using the singular flow
The above approach to searching over entire rate, F,, given by Eq. (69), as calcu-
ranges of tl and t2 is based on the assumption lated by the integrated A3, the assumed
that the ranges of t , and t2 are known. For a A,(?), the integrated state variables,
given maximum flow rate, tl has an upper lim- and the remaining three calculated ad-
it, and it is the time required to fill the bio- joint variables calculated from Eq.
reactor with the maximum flow rate, i.e., (70).
tl Ivmax/Fmax. There is also an upper limit 5. Using F=F,,,, integrate forward the
of t,, which is the time at which the substrate entire sets of the state and adjoint
concentration becomes very small as the singu- equations, Eqs. (60) and (63), from f 2
lar feed rate must initiate before the cells to t3,the time at which the bioreactor
starve or near the substrate concentration cor- is full.
responding to the optimum product formation 6. Continue integrating forward using
rate. In actual computations the range of tl F= Fmin = 0 from t3 to t f , at which
and t2 can be narrowed further through analy- either the specified final time or condi-
tical results as previously discussed (MODAKet tion is met.
al., 1986). 7 . Compare the calculated adjoint varia-
For other special initial conditions the opti- bles, A3 (tf) and A, ( t f ) ,with the specified
mal feed-rate profiles degenerate, as shown in values, A3(tf)= Vtf, and As ( t f )= -pf/t:.
Fig. 9(b)-(d), and thus the above algorithms 8. If there is no agreement between the
need to be modified. Actually, these situations calculated values and the specified val-
require simplification of the above general al- ues, improve the estimated values of
gorithm. Fig. 9(b) represents a situation in A3 (tl) and As ( t l ) by a Newton-Raphson
which the initial substrate concentration is method and go back to step 4.
high. The optimal control sequence now lacks 9. If there is agreement, store switching
538 I6 Control of Bioreactor Systems
times t, and t2 and the corresponding until the final time or condition is met,
value of the performance index, t =tf.
g [X(tf)l. 6. Compare the calculated adjoint varia-
10. Change or increment t2 by At2, and if bles, A3(tf)and A,(&), with the speci-
the entire range of t2 has not been cov- fied values, A3(tf)= l/t, and
ered, go to step 2. If covered, go to A,@,) = -pf/t:.
step 11. 7. If there is no agreement, improve the
11 Change or increment tl by A t , , and if estimated values of A3 (t,) and A, (t,) by
the entire range of tl has not been cov- a Newton-Raphson method and a go
ered, go to step 2. If covered, go to back to step 4.
step 12. 8. If there is agreement, store the switch-
12. After sets of switching times have been ing time tl and the corresponding value
tried over the entire feasible range, of the performance index, g [ x ( t f ) ] .
choose the one that yielded the maxi- 9. Increment tl by At,, and if the entire
mum value of performance index. If range of t , has not been covered, go to
necessary, further narrow the grid step 2. If covered, go to step 10.
points. 10. After various values of tl have been
tried, choose the one that gives the
When the optimal feed-rate profile is given maximum value for the performance
by Fig. 10(b), it is only necessary to modify index. If necessary, further narrow the
step 2 using F=Fmi,=O instead of F=F,,,. interval, say by a golden search.
For the optimal feed profile given by Fig.
lO(c), we set t , = O in step 1 and skip steps 2 For the profile given in Fig. ll(b), use
and 11. The profile given in Fig. 10(d) is a con- F= Fmin= 0 and t = t2 in step 2. For the profile
stant fedbatch process, and switching time in Fig. ll(c), steps 1 and 2 are skipped, tl is set
t3= [vmaX- v(0)]/Fmax is predetermined. to zero in step 3, and steps 9 and 10 are also
skipped.
1. Choose t,.
2. Integrate forward the state equation
with the given initial conditions, Eq. -
(60), using F= F,,, until t = t,.
3. Estimate the unknown variables at t l ,
A3 (td, and ( t l )= A 5 (t).
4. Integrate forward the A3 equations, Eq. ’OD-= 50-
(63), and the state equations, Eq. (60),
from tl until the fermentor is full, Time (hl
t = t 3 2 using F=Fs given by Eq’ (69)9 as Fig. 12. Optimal glucose feed rate and correspond-
calculated from Eq. (70). ing cell, glucose, and penicillin profiles: low initial
5 . Continue to integrate the state equa- glucose concentration.
tions, Eq. (60), and the entire sets of x,= 10.5 g, so=io-6g, p o = o g , v 0 = 7 L, v f = iOL,
adjoint equations, Eq. (63), from t = t 3 F,,= 10mL/h, sF=500g/L (LIM et al., 1986).
Optimization and High-Level Control 539
tail. The computational techniques are sum- Production of Baker’s Yeast Saccharomyces
marized above (the details are given elsewhere cerevisiae
(LIM et al., 1986), and these are used to gener-
ate the numerical results below. The model developed by MODAK(1988) for
baker’s yeast is given below, consisting of
mass balance equations for glucose, ethanol,
Penicillin Fermentation by Penicillium chryso- and cell mass, and the overall mass balance
genum equations, along the line of Eqs. (53) through
(56):
The penicillin fermentation model of BAJ-
PAI and REUSS(1981) given below has three
d(xv) - p(xv)
--
component mass balance equations for sub- dt
strate (glucose), cell mass, and penicillin, and
one overall mass balance equation, Eqs. (53)
through (56), with the following specific (73)
rates:
d(Ev) - (n-q)(xv)
--
+
p = 0.1 1 ~ / ( 0 . 0 0 6 ~S ) dt
(74)
n=0.004 s/(O.OOOl + s + 10s’)
(71)
0 =p/0.47 + n / l .2 + 0.029 (75)
k=0.01 h - ’
where G and E stand for glucose and ethanol
The authors state that the model with the concentrations, respectively; various rates are
above rates adequately describes the experi- given below:
mental data available in the literature (PIRT
and RIGHELATO,1967; HOSLERand JOHN- ki G + k2G2
SON, 1953; Mou, 1983). The performance in- P = P G + P E=
k3+ k,G+ G2
+
dex to be maximized is the amount of penicil-
lin at an unspecified final time. This is a free + (k,+ k7ak+SEE) ( 1 + k7a)
time problem, so the optimization procedure
must be used to determine the optimum time
profile for feed rate that maximizes the total
mount of penicillin produced at various final 1+k,G2
times and the one must be chosen that gives R ( G )=
the largest amount of penicillin. k l o+ k9G2
This is a typical example of type I fermenta-
tion due to the non-monotonic specific penicil- X= k13OR (79)
lin production rate, n. The optimal glucose rl = rUdkl.4 (80)
feed-rate profiles for glucose, cell mass, and
penicillin are given in Fig. 12. The optimal glu- The numerical values of these parameters are
cose feed profile consists of a period (11.2 h) given in Tab. 5.
of maximum flow rate, a batch period
(17.6 h), a singular flow rate period (95 h), and
a negligibly small batch period. Note that the Tab. 5. Kinetic Model Parameters (in g/g) (MODAK,
profiles clearly show the experimentally ob- 1988)
served phenomena: a tropophase in which al- ~~~ ~
Many industrially important fermentation where x, sl, sz, and p represent the concentra-
processes involve microbial cells that require tions of cells, starch, caseinate, and a-amy-
more than one substrate for their growth. It lase, respectively. v is the bioreactor volume,
has long been realized that the production of F, and Fz the feed rates of starch and casei-
antibiotics and enzymes requires precise con- nate, slFand S2F the feed concentrations of
Optimization and High-Level Control 541
starch and caseinate, and k the a-amylase hy- (F,47,). Therefore, the feed rate of carbon
drolysis rate constant. The specific rates p, o,, source, F l , is neglected in the overall mass bal-
02,and n are functions of starch and caseinate ance equation of Eq. (88).
concentrations: We define as state variables x1= x v ,
x 2 = s 2 v ,x 3 = p v , x 4 = v , and x 5 = t and as con-
0.086 ~ 1 ~ 2 0,= - P trol variables m l = F 2 and m 2 = s 1 .Eqs. (81)
P= through (85) can be expressed in terms of new-
(2.0 +sl+s:/33.0) 0.68
ly defined state and control variables as:
P
n= 117.7 e ~ p - " ~ ~ ' ' z poz = -
1.05 dx
-= a ( x , m 2 ) + b m ,
dt
and k = 0.18. The operating conditions used in
the optimization study are and
Max = { P = P v ( t f ) - e t f } (88)
FIVFZ
g/L. By this fact the original optimization minimum feed rate (22.51-24.89 h). For a-
problem with two control variables is reduced amylase production the specific growth rate of
to an optimization problem with only one (ca- the cells increases monotonically with increases
seinate feed rate). This is a standard four-di- in caseinate concentration, whereas product
mensional singular control problem with a sin- yield is inhibited at high caseinate concentra-
gle control variable. Interested readers are re- tions. Therefore, rapid cell growth can be
ferred to our previous publications (MODAKet achieved by supplying caseinate as rapidly as
al., 1986; LIM et al., 1986; MODAKand LIM, possible (maximum feed rate) and then shut-
1989) for details of the computational algo- ting off the feeding (minimum feed rate) to al-
rithm. low the cells to grow on caseinate. This is fol-
Fig. 14 shows the optimal caseinate feeding lowed by a period of singular feed rate to
policy for the a-amylase production process. maintain the caseinate concentration at the
The optimal feed rate has a period of maxi- level that does not inhibit product yield. The
mum feed rate (0-2.1 h), a period of minimum last period of minimum feed rate is a result of
(batch) feed rate (2.1-16.5 h), a period of sin- the constraint on the volume of the bioreactor.
gular feed rate (16.5-22.51 h), and a period of The concentration profiles resulting from the
optimal feed rate are shown in Fig. 15. As ex-
pected, the high concentration of caseinate in
1.2 I I10 the initial period (0-16.5 h) allows cells to
grow rapidly, whereas during the singular in-
terval (16.5-22.5 1 h), caseinate concentration
is lower in order to achieve a higher product
yield. The rate of addition of starch required
to maintain a constant level (8.12 g/L) of
starch can be calculated by rearranging the
starch mass balance equation from the set of
mass balance equations:
4 Estimation Techniques
For better control and optimization of fer-
mentation processes it is essential to measure
on-line many key physiological parameters.
Such measurements are difficult, if not impos-
sible, for many of these parameters. There-
fore, any control or optimization based on key
physiological parameters cannot be imple-
Fig. 15. Cell mass, caseinate, and a-amylase profiles mented unless values can be measured on-line
with optimal starch and caseinate feeding strategy to provide the necessary information required
(MODAKand LIM, 1989). by the controller or optimizer. Although ef-
Estimation Techniques 543
forts to develop such sensors are underway, sumption or production of many species can
the availability and reliability of instruments be calculated by taking a simple species bal-
are very limited. There is thus an urgent need ance around the fermentor. Among the quan-
to provide the necessary information by esti- tities most easily acquired on-line and most
mating these parameters from others that are frequently calculated are the oxygen uptake
simpler and easier to measure. In this section rate (OUR), the carbon dioxide evolution rate
we present parameter and state estimation al- (CER), and the respiratory quotient (RQ).
gorithms that can be used to estimate a key pa- These provide very useful information and are
rameter or a physiological state of a fermen- very good indicators of cellular respiratory ac-
tor. tivities. Other variables obtained from indirect
It is true for any processing industry that measurements include the overall oxygen
balance equations can provide much needed mass-transfer rate, metabolic heat-evolution
information. Material balance around a fer- rates, the specific growth rate, cell yields, sub-
mentor is extremely valuable in this respect. strate utilization rates, and secondary metabol-
Energy balance equations have not played as ite production rates. A list of the calculated
important a role as material balance equations variables (ARMIGERand HUMPHREY,1979)
because fermentation processes take place in and a set of straightforward step-by-step data
aqueous solutions, and because the heat of fer- analysis schemes are available (NYIRI, 1971,
mentation is usually rather small. Balances of 1972). We discuss below a few schemes that
elements such as C, H, N, 0, and others can are more practical.
be used with off-gas data to estimate some The specific growth rate was measured indi-
quantities not directly measured, such as bio- rectly in a turbidostat system by controlling
mass or product concentrations. Since the the cell concentration between the upper and
state of instrumentation is still unsatisfactory, low limits via a continuous optical density
many measurements contain a high level of measurement (VERES et al., 1981). The specific
noise. Thus, raw measurement signals should product formation rate can be determined by
be passed through suitable filters before they applying a similar technique to a pH control
are made available to control bioreactors. scheme and by carefully monitoring the cumu-
lative amount of acidlbase added to achieve
neutralization. Acetic acid production by
4.1 Indirect Measurements Escherichia coli (SANand STEPHANOPOULOS,
1984) and gluconic acid production (VERES et
and Correlations al., 1981) have been estimated in this way.
Another important indirect measurement is
Certain data collected from bioreactors can the volumetric oxygen transfer coefficient,
be used directly, while others have very little kLa, as described by the following equation:
physical significance by themselves and must
therefore be combined with other measure-
ments to provide physically significant infor-
mation. For example, low level measurements where qo,(t) is the volumetric oxygen transfer
such as temperature, pH, and dissolved oxy- rate, Cb, ( t )is the liquid-phase oxygen concen-
gen can be directly fed back to actuate control tration in equilibrium with the gas phase, and
such devices as the heater/cooler, pumps for Co2( t ) is the liquid-phase oxygen concentra-
acid/base addition, and controllers for the gas tion. It is important to note that kLais a func-
flow rate or the impeller rotation speed. On tion of time due to changes in fermentor con-
the other hand, gas flow rate data are some- ditions such as agitation, air dispersion, and
what uninformative unless they are combined rheological properties; it therefore needs to be
with other measurements to form so-called estimated and constantly updated by combin-
“gateway” sensors (HUMPYHREY, 1971). Indi- ing the measurements of gas flow rate, gas-
rect measurements and gateway sensors pro- phase oxygen concentration, and dissolved
vide information on cellular metabolism and oxygen concentration.
fermentation conditions. The rates of con-
544 16 Control of Bioreactor Systems
Two general methods are used for estima- which no metabolite is produced, and there-
tion of kLa: static and dynamic methods. Both fore the product is cell mass. The basic feature
methods assume that the bioreactor is well of the method is to represent the biological
mixed, that kLa is a function of time but not conservation of substrate to cell mass by an
space, and that the dynamics of oxygen are overall chemical reaction as follows:
much faster than the dynamics of biomass,
substrate, and products. The last assumption aC,H,O,+ bOz+cNH3 -+
CER, or the concentrations of various species Two more equations are needed to solve for
over a short period during which the stoi- the six unknown stoichiometric coefficients a
chiometric coefficients are assumed to be con- through f . As above, the offgas exchange in-
stant: formation provides another equation:
The specific growth rate can be estimated by plied with initial conditions that are often at
rewriting Eq. (24) as: best rough guesses. Thus, a good noise filtra-
tion algorithm should be employed to improve
1 dx Yo, the reliability of the estimated values before
p = -- = -OUR-mo, Yo, (113)
x dt b they are used for control purposes.
(119), the standard Kalman filter problem is prediction of error covariance matrix
based on a set of sequential observations
Z ( k )= (z(l), z(2), . . ., z(k)}. The objective is P(kIk- 1)= @(k)P(k- 1 Ik- l)QT(k) + Q(k- 1)
to find linear, unbiased, minimum-variance es- (121)
timates of the true state xu), denoted by Corrector:
fg’l k). Depending on the relative values of j state estimate
and k , the estimation is referred to as predic-
tion u > k ) , filtering u = k ) , or smoothing f(kl k) =f(kl k - 1) + K ( k ) .
g’<k). f ( k l k ) will be denoted by f(k). * [z(k)- H ( k )f (k I k - 1)l (122)
error covariance matrix
Tab. 6 . Standard Assumptions and Definitions of
the Discrete Kalman Filter P(k I k ) = [ I - K ( k )H ( k ) ]P(k I k - 1) (123)
E[w(k)l = o Kalman gain matrix
COV[W(k), WOII=Q(k)djk a
E[v(k)]=O K ( k )=P(k I k - 1) H T ( k ) *
C O V [ V ( vOI1
~ ) , =R(k)Jjk * [H(k) P(k I k - 1) H T ( k )+ R (k)]- 1 (124)
cov[x(O), w(k)] =0, for all k
cov[x(O), vO]]=O, for a l l j initial conditions
cov [w( k ) , vO)] = 0, for all j , k
E [x(O)I = m (0) f(O)=m(O) (125)
COV[X(O), X(0)l = p m
P(0)= Px (0) (126)
a djk is the Kronecker delta function
where P ( k ) is the symmetric error covariance
matrix of the estimation, defined by:
The Kalman estimation is described by the
following set of recursive vector filtering dif- P(k)= [x(k)-a(@] [X(k)-f(k)lT
ference equations: =Z(k)fT(k) ( 127)
Predictor:
prediction of state A schematic diagram depicting implementa-
tion of the above equations to obtain the best
f(kl k - 1) = @ ( k ) f ( k - 1 I k - 1) = @(k)P(k- 1) estimate is given in Fig. 16. The error covar-
(120) iance matrix P ( k ) is obtained by integrating
E [ k l k-1 j
c
Delay
kIk-l/k-ll
@(k-1) -
E[k / k - l j
Fig. 16. System model and discrete Kalman filter.
Estimation Techniques 549
Eqs. (121) and (123), and the Kalman gain ma- Kalman filter and giving small errors,
trix K ( t ) is calculated from Eq. (124). The esti- f ( f l f k ) = x ( t ) - f ( f I tk). One would expect that
mated noise, the difference between the cor- the estimator would approach the linear case
rupted measurement z(k) and the estimated Kalman filter for vanishingly small non-linear-
noise-free measurement Hf (k) is multiplied by ities.
the Kalman gain and added to the noise-free The extended Kalman filter estimation is de-
predicted value, f(kl k- l), which is calculated scribed by the following set of vector filtering
from Eq. (120) using the best estimate of the equations:
previous step. Predictor:
A prediction of estimate is produced by inte-
gration of the differential equation
Extended Kalman Filter
x*(tIt,)=f(x(tIfk), t) (130)
We consider a more general, non-linear dy-
namic model and a discrete measurement prediction of error covariance matrix
model:
system model:
i = f ( x ( O , t ) + G ( x ( t ) ,t ) w ( t ) (128)
Corrector:
measurement model: state estimate
z (fi) = h ti), ti) + v (ti) (129)
where f is a vector of a non-linear function of
x ( t ) and t , and h is a vector of non-linear func-
tions of x ( t i ) and t i . The standard assumptions error covariance matrix
and definitions are presented in Tab. 7.
E [ ~ ( t =) 0]
cov[w(t),w(t)]= Q ( t ) s ( t - ~ ) ~
E [ v ( i ) ]= O
cov [v(i),v01=R(06, (135)
cov [x(to), w(t)]= 0,for all t z to
cov [x(to), v(ti)] = 0,for all j
cov[w(t),v(ti)]=0,for all ? = t oand t , r t o
E [x(to)l= m (0) initial conditions
cov [x(to), x(tiJ)l= p m
a 6( .) is the Dirac delta function
For the model given by Eqs. (128) and Implementation of the extended Kalman fil-
(129), the extended Kalman filter problem can ter is essentially identical to that of the Kalman
be stated as follows: given a measurement se- filter. The only difference is that in the ex-
quence Z(k) = { z ( t o ) ,z ( t l ) , . . ., z ( f k ) )find
, an tended Kalman filter algorithms F and H are
estimator to provide estimates of x ( f ) , denoted the linearized matrices given by Eqs. (135) and
by x ( f l f k ) , patterned after a linear case (136).
552 16 Control of Bioreactor Systems
understand and quantify the complex intricate uous culture of Saccharomyces cerevisiae,
networks of biochemical reactions present in a more commonly known as baker’s yeast, is
living cell with sophisticated regulatory sys- maximized by manipulating both temperature
tems governing the rate of growth and synthe- and dilution rate. This system is of industrial
sis of various products. Therefore, it is practi- importance due to the use of yeast in the bak-
cally impossible to develop a model that can ing and brewing industries and in the manufac-
adequately describe the process over a wide ture of various yeast products using non-re-
range of operating conditions and yet be sim- combinant and recombinant yeast strains. A
ple enough to be used for optimization calcula- continuous bioreactor can be optimized using
tions. In addition, due to cellular adaptation, steady-state optimization techniques based on
selection, enrichment, as well as variations in steady-state data. However, these techniques
the crude feed, optimum conditions cannot be are usually extremely time-consuming since the
determined off-line. Thus, it is important to time required for the continuous cultures to
track closely changes in the culture and operat- reach a new steady state after step changes in
ing conditions and reflect them in the optimi- the manipulated variables may be over 20
zation. Therefore, an adaptive feature is re- hours, resulting in optimization times of the
quired in the optimization of bioreactors. A order of days and weeks. Therefore, it is essen-
typical adaptive optimization scheme is shown tial to develop a speedy optimization algo-
in Fig. 19. The process optimizer consists of a rithm based on short transient data rather than
process identifier that identifies changes in the to wait for steady-state data.
culture and operating conditions through a An adaptive optimization algorithm based
model with adjustable parameters and an op- on dynamic model identification has been de-
timization scheme for the setpoint, which uses veloped for continuous cultures (ROLF and
the model with updated parameter values. We LIM, 1985; CHANGet al., 1988; SEMONES and
LIM, 1989; CHANG and LIM, 1989). This
method requires only minimal a priori infor-
Disturbances mation, such as model structure, parameter
1 values, and feed conditions. A simple empiri-
cal dynamic input-output model is used, and
its parameters are recursively estimated from
-:- transient data to track the process on-line. A
. .
........ recursive least-square method was used for pa-
.......................... .: rameter estimation. In determining the control
action to improve the performance of the con-
Fig. 19. Adaptive optimization. tinuous culture, the steady-state portion of the
dynamic model was used instead of waiting for
actual steady-state information.
shall discuss in some detail the adaptive optim-
ization of a continuous culture to maximize
the productivity of a baker’s yeast culture by 5.1.1 Problem Formulation
manipulating the dilution rate and tempera-
ture. The performance index to be maximized is
the productivity of a continuous culture of
baker’s yeast, i.e., the product of the dilution
rate, D , and the cell concentration x:
5.1 Adaptive Optimization
of Continuous Bioreactors P=Dx (145)
In this section adaptive optimization strate- The manipulated variables for the optimiza-
gy is implemented to maximize the productivi- tion are temperature, T, and dilution rate, D:
ty of a continuous culture. Specifically,
steady-state cellular productivity of a contin- u T = [ T ,D]
Adaptive Optimization and Control 55 1
z = [ x1
x2 + x7
I+" model parameters converging on the true val-
ues toward the end of the run.
As stated earlier, this general scheme can be
applied in many situations to estimate contin-
where x = [xspmaxKskd YsblT and z = [ x , ~ , ] ~ . uously not only the states but also the asso-
Now we have system and measurement mod- ciated kinetic parameters on-line, since they
els. Using the extended Kalman filter equa- are observable. This scheme is especially useful
tions, Eqs. (130)-( 138), we can simultaneously for estimating highly noisy derivative variables
estimate cell mass and substrate concentrations such as the specific growth rate. A general
and model parameters. Experimental data for drawback of these filters is that judicious
a batch culture of Trichoderma viride grown choices of the noise covariance matrices, Q
and R , and the initial error covariance matrix,
Po, are required to achieve proper filtering.
This may sometimes be a difficult task.
0.05
5 Adaptive Optimization
4" ;+=-=
15
and Control
'"1...........................................................................
5 Whereas in adaptive control the operating
points (setpoints) are assumed to remain con-
stant, the objective of adaptive optimization is
to determine optimal operating conditions for
bioreactors that may be unknown or may
change with time. For example, for a contin-
uous culture one needs to determine on-line
such optimal operating conditions as dilution
rate, temperature, and pH. This is done by
perturbing the culture, observing the output
(perhaps the productivity), and continually im-
............................................................................. proving the output. For batch and fedbatch
cultures there is no reason to hold any operat-
ing conditions constant, since optimal condi-
a5 tions for growth may not necessarily be opti-
mal for product formation. Consequently it is
desirable to determine on-line the changing
1ime.days
characteristics of cultures and reoptimize bio-
Fig. 18. Model parameter estimates by a Kalman fil- reactor operating conditions. Due to the com-
ter. True values (. . . . . .) estimates (-) plexity of microbial cultures, it is difficult to
550 16 Control of Bioreactor Systems
In these Kalman filtering schemes, not only the saturation constant, the yield coefficient,
state variables but also unknown constant pa- and the death rate coefficient, respectively. Be-
rameters such as cell yields can be estimated by sides the system model, the measurement mod-
treating them as additional “state variables” el must be specified. If the dry weight of the
whose time derivatives are zero. For example, biomass can be measured directly while the
consider a batch bioreactor characterized by substrate concentration measurement is
two state variables: biomass and substrate. If biased, one can propose the following meas-
10 , I 30
lime, days
Fig. 17. Kalman filter estimates of cell mass and substrate concentration. Model (. . . . . .) Kalman
filter (-) (0 cell mass, 0 substrate)
Experimental data: NIHTILAand VIRKKUNEN (1977).
the growth kinetics follow the Monod type and urement model (NIHTILA and VIRKKUNEN,
the yield coefficient is constant, the following 1977):
system models are generally used for growth
and substrate consumption: x, = x+ v1 (141)
sm= s + Sb + v2 (142)
(139)
where x, and ,s are the measured values of
and dry weight and substrate concentration, sb is a
constant bias, and v, and v2 denote measure-
ment errors. The unknown parameters prnax,
(140) K,, k d , Y , and sb are contained in the system
and measurement equations, Eqs. (139)-( 142).
where x and s are the concentrations of bio- These parameters can be estimated with x and
K,, Y ,
mass and substrate, respectively; pmax, s by setting up an augmented dynamic equa-
and kd are the maximum specific growth rate, tion:
Adaptive Optimization and Control 553
where aD and aT are the gains (step sizes) for The remaining problem is to determine the
the dilution rate and temperature optimiza- nine constants in Eq. (151). Knowing these,
tion, respectively. the steady-state gains can be calculated from
The optimization scheme can be easily im- Eq. (156). Eqs. (149) and (150) are then used
plemented if one can determine the steady- to calculate appropriate changes in the dilution
state gains, a x / a D and ax/aT. These can be rate and temperature to optimize continuous
calculated on-line using an input-output dy- culture.
namic model of the process and updated pa-
rameter values.
5.1.3 Parameter Estimation
5.1.2 Process Model The initial estimates of the nine constants in
Eq. (151) are established by using a standard
Proper choice of a process model is critical least-square technique. During initialization
in applying adaptive process identification al- the culture is excited by introducing test signals
gorithms (ROLF and LIM, 1985; SEMONESand in the manipulated variables to generate an ini-
LIM, 1989). Based on the accuracy of the stea- tial data set that will yield good first estimates
dy-state gain estimates, matrix invertibility, of the parameters. The test signal chosen is a
and sensitivity to data-sampling time, the weighted m-sequence (KASHIWAGI,1974) that
model selected is a second-order linear model is a low frequency signal with an impulse-like
given by: autocorrelation function. Staggered m-se-
quence signals are used for temperature and
x(k)= -a,x(k-1)-~2x(k-2)+b,oD(k)+ dilution rate as shown in Fig. 20. A one-shot
+ bll D ( k - 1 ) + b,,D(k-2) + b,oT(k) + least-square algorithm is then applied to the
+ b21 T ( k - 1) + b22 T(k-2) + c (151) data set to obtain the initial parameter esti-
mates:
where k is the discrete sampling-time index and
the constants ai,b,, and c are the model pa- q = ( X T X )- l X T y (157)
rameters to be determined on-line using tran-
sient data. Eq. (151) may be also written as: where
where and
1.2-
1.1-
1.0-
c
2I 0.9- -
-y* 0.8-
U
-
0
.
a
- 30
-- 2 9 -c
0 .
.-.->
e a
-28
- 27 2
e e
u
a
U a
&
e c
0
--2 6
a
0 -25
"i I - {
0.25-
0.21 -
0.23 -,
0.22 -
0.21 -
0.20 I
55 60 65 70 75 80 85 90 95 100 105 110 115
Time, h
Fig. 20. The m-sequence profile with staggered step changes in manipulated variables ( n = 15,
A t = 96 min, SEMONESand LIM, 1989).
making the first step changes in the manipu- Our experience has been best with the so-called
lated variables, the parameters are updated to bilevel forgetting factor method in which the
improve the accuracy of the linear model. This forgetting factor is placed either at a maximum
is done by using a recursive least-square algo- level, Amax, when the error between the mea-
rithm (HSIA, 1977): sured and predicted output values is smaller
than a preset value, e2( k )= Iv ( k )-y (k)]
+ +
g(k+ 1) = q ( k ) y(k 1) P ( k )x ( k + 1)- <vartol, or at a minimum value, Amin, when
* b ( k +l)-xT(k+ 1) q ( k ) ] (160) the error is greater than the preset value
(CHANGet al., 1988; CHANGand LIM, 1989).
y ( k + 1)= 1/[1 + x T ( k +l ) P ( k ) x ( k + l)] (161) The details and the exact experimental condi-
tions are available elsewhere and are therefore
P ( k + 1) = [ P ( k )- y(k+ 1) P ( k )X*X*T. omitted here, where we instead give a brief de-
* ( k +1) P ( k ) ] / A ( 162) scription of its implementation.
X * ( k ) = [x(k),0, . . ., 01 ( 163)
5.1.4 Implementation
P(0)= (X'x) - (164)
The procedure consists of a sequence of
where A is the forgetting factor, O < A 5 1 and P initialization, optimization, and supervision/
is the covariance matrix of the input-output reoptimization. In the initialization stage, stag-
data. This algorithm is sensitive to the choice gered rn-sequences in dilution rate and temper-
of the forgetting factor. The forgetting factor ature are introduced to generate an initial set
can be set at a constant value, adaptively tuned of data, from which the initial estimates of the
(HSIA, 1977), or adjusted by other criteria. input-output model parameters and the
Adaptive Optimization and Control 555
steady-state gains are made. These initial esti- Do = 0.27 h and To= 28.0 "C. The initializa-
mates are used to initiate recursive parameter tion phase took 45 hours, which was then fol-
estimation by sampling the data and introduc- lowed by an optimization phase of about 100
ing step changes in the manipulated variables hours. At 210.5 hours a very large step change
at several sampling times. This process of pa- was introduced in the pH (from 5.5 to 7.0) to
rameter updating and making step changes in test the adaptability to extreme external distur-
the manipulated variables continues until the bance and the reoptimization capability of the
algorithm is judged to have converged by the algorithm (Fig. 22). The gains in the steepest
smallness of the gradients in the performance ascent were increased for faster optimization
index. results. This large change in pH caused the
At convergence the step changes are halted, growth rate to decrease substantially; as a re-
but parameter updating is continued (supervi- sult, the dilution rate decreased rapidly at the
sion). In the event a process variation or dis- beginning to save the culture from washout.
turbance occurs, the gradients will increase After about 40 hours the culture converged to
and trigger a restart of the optimization (reop- a new steady state: D=0.25 h-', T=28.6"C,
timization). and D x = 0.79 g/L/h. The algorithm can ap-
parently detect changes, and it quickly finds
the corresponding optimum conditions and
5.1.5 Results and Discussion maintains the operation at this optimum. To
see if more effective perturbation of the cul-
Fig. 21 shows the initialization and optimi- ture in the initiation stage would improve the
zation stages, and Fig. 22 shows the reoptimi- initial estimate and therefore shorten the op-
zation phase. The initialization was started at timization time, another run was made using
43.7 hours into a steady-state operation at larger step changes in temperature. In addi-
Time, h
Fig. 21. Multivariable on-line optimization (initialization and optimization periods) (CHANGand LIM,
1989).
A D = 0 . 0 2 6 h - ' , AT=l.O"C, m(z)= -1, -1, -1, -1, -1 for D and T , Ami,=0.94, &,,,,=0.995,
CYD= 0.0005, CYT= 4.0.
556 16 Control of Bioreactor Systems
.
.
d
- 1.1-
e" 1.2-
- -
*
- -
s* 1.0-
.->
.-z 0.8-
a 0.31
w
t 0.6-
-
--0.29 -0
I0.30 .'
-
-
-0.28 ;
10.27
10.26
0.25
u 35:
f 31:
5 33-
E 32:
5 31- 5.5 -7.0
-
30:
g 29:
28 I I I 1 I I I I I I I u I I ~~~
.
:1.1
m
--4
I
1.3
*
.-
' 5 1.2
-
I
a
-0
1.1
35-
-
u
31:
-;
O .
-
L 33-
32-
K
I
e 30-
31-
T
-
5
29:
28-
27:
ox
:.-
5 1.3-
1.2-
2 1.1-
g 1.0- 0 -0.32f
-0.31 2
-0.30 2-
-0.21
Fig. 24. Multivariable on-line op- e
-0.261
timization (CHANGand LIM, u
-c
0 . T
1989). No initialization stage, 3b-
tion, unlike the first run in which the same m- rameter updating and optimization calcula-
sequence but staggered signals were used, dif- tions. The calculated optimization results are
ferent m-sequences were employed. The results then implemented for the next interval. During
are shown in Fig. 23. The initial estimates were this interval the necessary information is col-
more accurate, but the overall results were dif- lected again for parameter updating and op-
ficult to judge due to the many tuning con- timization calculations. The results are imple-
stants in the algorithm. Since considerable mented in the subsequent subinterval. This
time is spent in making the initial estimate, an- procedure may be repeated until the end of the
other run was made in which the initialization bioreactor operation period. Therefore, as
stage was eliminated by using the initial esti- long as one can monitor on-line the key state
mates from a different run. The results (Fig. variables that make up the mass and energy
24) show no ill effect, and this implies that the balance equations, one can update the bioreac-
algorithm can accommodate considerable er- tor model parameters through an identifica-
ror in the initial estimate. tion scheme and then reoptimize the operating
conditions. If one or more of the key state var-
iables are measurable on-line, other readily ac-
5.2 Adaptive Optimization of Batch cessible parameters should be measured, which
are then used in an estimation scheme to pro-
and Fedbatch Bioreactors vide the necessary information to update the
bioreactor model parameters. Alternatively,
Adaptive optimization of batch and fed- the extended Kalman filter may be used to ob-
batch bioreactors is theoretically feasible, es- tain simultaneously an estimate of the missing
pecially for those situations in which the peri- key state variables and the bioreactor model
od of bioreactor operation is long, for exam- parameters, as demonstrated in Sect. 4.
ple, a penicillin fermentation that may last 5 On the other hand, the operating period of
days or more. In these situations the bioreac- bioreactors may be rather short, sometimes
tor operation perioct is broken into a number only several hours, relative to the repeated on-
of subintervals, the number depending on the line parameter updating and optimization
process dynamics. The bioreactor operation computation required for adaptive optimiza-
during the first subinterval is used to collect in- tion. In these situations the computational
formation necessary for parameter updating burden may be too great to subdivide the oper-
and subsequent optimization calculations. If ation period into several subintervals, and on-
the computational burden is great, the bioreac- line adaptive optimization may not be realistic.
tor operation may continue over a short inter- One approach based upon the bioreactor oper-
val using operational conditions prescribed ation time and computational requirement is
perhaps by the previous run, during the pa- the so-called cycle-to-cycle optimization.
558 16 Control of Bioreactor Systems
The basic principle behind cycle-to-cycleop- bly limited by the scarcity of reliable on-line
timization is simple and is often practiced in sensors for monitoring key parameters in a fer-
industry. The idea is to make a batch or fed- mentor. Availability of more on-line sensors
batch run and generate the data required to will help in utilizing computers more fully for
update the bioreactor parameters. Parameter optimization and control of fermentation
estimation and optimization calculation are processes.
done off-line, and the results are implemented Empirical correlations capable of predicting
for the next cycle. This cycle then generates the effects of small changes in control variables on
necessary information for off-line parameter the response of microbial systems are very use-
update and optimization calculation. The cal- ful, and they can be effectively utilized in
culated results are then implemented in the adaptive control. The next logical challenge is
next cycle. This procedure is repeated until a broadly applicable adaptive optimization al-
there is negligible change predicted by the op- gorithm that is capable of self-optimization.
timization calculation. This approach has ap- We also look forward to seeing more effective
parently been successfully implemented for applications of artificial intelligence and expert
penicillin (CHITTUR, 1989) and baker’s yeast systems.
fermentations (MODAK, 1988). It was found The intent of this chapter has been to pro-
that the procedure would converge on the opti- vide necessary information concerning the
mal results in two to three iterations. Howev- types of control techniques available and their
er, if the structural form of the model is incor- significance to fermentation processes. Specif-
rect the result may not converge. Interested ic examples of the application of some of these
readers are advised to consult the original techniques can be found in the reviews men-
sources. tioned earlier. Finally, it must be stressed that
control considerations should be taken into ac-
count early in the design stage of a fermenta-
tion process. This will assure both controllabil-
ity of the fermentation process and overall op-
6 Future Prospects timal performance of the entire system.
BRISTOL,E. H. (1977), Control Eng. 24(1), 24. HSIA, T. C. (1977), System Identification, Toronto:
BULL, D. N. (1983), Annu. Rep. Ferment. Proc- Lexington.
esses 6, 359-375. HUMPHREY,A. E. (1971), Proc. LABEX Symp.
CHANG, Y. K., LIM, H. C. (1989), Biotechnol. Computer Control of Fermentation Processes,
Bioeng. 34, 577. London, p. 1.
CHANG,Y. K., PYUN,Y. R., LIM, H. C. (1988), JEFFERIS,R. P . (1975), Process Biochem. 10(3),
Biotechnol. Bioeng. 31, 944. 15.
CHITTUR,V. (1989), Ph. D. Thesis, Purdue Univer- JEFFERIS,R. P., I11 (1979), Ann. N . Y. Acad. Sci.
sity, West Lafayette, IN. 326, 241.
COHEN,G. H., COON,G. A. (1953), Trans. ASME JEFFERIS,R. P., 111, WINTER,H., VOGELMAN,H .
75, 827. (1977), in Workshop Computer Applications in
CONSTANTINIDES, A., SHAO,P. (1981), Ann. N . Y. Fermentation Technology, GBF Monogr. Ser. 3,
Acad. Sci. 369, 167. 141.
CONSTANTINIDES, A., SPENCER,J . L., GADEN,E. JUTAN, A., RODRIGUEZ,E. S. (1984), Can. J.
L. (1970a), Biotechnol. Bioeng. 12, 803. Chem. Eng. 62, 802.
KALMAN,R. E. (1960), J. Basic Eng., 35.
CONSTANTINIDES, A., SPENCER,J. L., GADEN,E.
KASHIWAGI, H. (1974), Proc. Fifteenth Joint Auto-
L. (1970b), Biotechnol. Bioeng. 12, 1081.
matic Control Conference, p. 285, AIChE Publi-
COONEY,C. L., Mou, D. G. (1982), in Computer cation.
Applications in Fermentation Technology, p. KOIZUMI,J., MODEN,Y., AIBA,S. (1985), Biotech-
219, London: Society of Chemical Industry. nol. Bioeng. 27, 721.
COONEY,C. L., WANG, H. Y., WANG, D. I. C. KOPPEL,L. B. (1968), Introduction to Control The-
(1977), Biotechnol. Bioeng. 19, 55. ory, Englewood Cliffs, NJ: Prentice-Hall.
COUGHANOWR, D. R., KOPPEL,L. B. (1965), Proc- LIM, H. C., TAYEB,Y. J., MODAK,J. M., BONTE,
ess Systems Analysis and Control, New York: P. (1986), Biotechnol. Bioeng. 28, 1408.
McGraw-Hill. MODAK,J. M. (1988), Ph. D. Thesis, Purdue Uni-
COURT,J. R. (1988), Progress in Industrial Micro- versity, West Lafayette, IN.
biology (BUSHELL,M. E., Ed.), pp. 1-45. New MODAK,J. M., LIM, H. C. (1989), Chem. Eng. J.
York: Elsevier. 42, B15.
DEKKERS,J. G. J., DE KOK, H . E., ROELS, J. A. MODAK,J. M., LIM, H. C., TAYEB,Y. L. (1986),
(1981), Biotechnol. Bioeng. 23, 1023. Biotechnol. Bioeng. 28, 1396.
DORBY,D. D., JOST, J. L. (1977), Annu. Rep. Fer- MOORE, C. F. (1986), Proc. A m . Control Conf.,
ment. Processes 1, 95-114. Seattle, p. 643.
DOUGLAS, J. M. (1972), Process Dynamics and Mou, D. G. (1983), Ph. D. Thesis, Massachusetts
Control, Englewood Cliffs, NJ: Prentice-Hall. Institute of Technology, Cambridge, MA.
DUNN,I. J., DANG,N. D. P. (1976), in Abstr. Pa- Mou, D. G., COONEY,C. L. (1983a), Biotechnol.
pers 5th Int. Fermentation Symp. (DELLWEG, Bioeng. 25, 225.
H., Ed.), p. 59, Berlin. Mou, D. G., COONEY,C. L. (1983b), Biotechnol.
EDGAR,T. F., HIMMELBLAU, D. M. (1988), Optim- Bioeng. 25, 257.
ization of Chemical Processes, New York: NIHTILA,M., VIRKKUNEN, J. (1977), Biotechnol.
McGraw-Hill. Bioeng. 19, 1831.
ERICKSON, L. E. (1979), Biotechnol. Bioeng. Symp. NYIRI,L. K. (1971), Proc. LABEXSymp. Comput-
9, 49. er Control of Fermentation Processes, London,
GOFF, K. W. (1966), ISA J., December, 44. p. 16.
HAMPEL,W. A. (1979), Adv. Biochem. Eng. 13, 1- NYIRI,L. K. (1972), Adv. Biochem. Eng. 2, Chap. 2.
33 *
HATCH, R. T. (1982), Annu. Rep. Ferment. Proc- NYIRI,L. K., TOTH, G. M., CHARLES,M. (1975),
esses 5, 291-311. Biotechnol. Bioeng. 17, 1663.
HERBERT,D. (1976), in Continuous Culture 6: Ap- OHASHI,M., WATABE,T., ISHIKAWA,T., WATA-
plications and New Fields (DEAN,A. C. R., ELL- NABE, Y., MIWA, K., SHODA, M., ISHIKAWA,
WOOD,D. C., EVANS,C. G. T., MELLING,J., Y., ANDO, T., SHIBATA,T., KITSUNAI,T.,
Eds.), London: Ellis Horwood. KAMIYAMA,N., OIKAWA,Y. (1979), Biotechnol.
HILL, G. A., ROBINSON,C. W. (1974), Biotechnol, Bioeng. Symp. 9, 103.
Bioeng. 16, 53 1. PAZLAROVA, J., BAIG,M. A., VOTRUBA,J . (1984),
HIRAMA,T., HUMPHREY,A. E. (1980), Biotech- Appl. Microbiol. Biotechnol. 20, 33 1.
nol. Bioeng. 22, 821. PERRY,R. H., GREEN,D. (Eds.) (1984), Perry’s
HOSLER,P., JOHNSON,M. J. (1953), I & EC 45, Chemical Engineers’ Handbook, 6th Ed., pp. 22-
871. 26, New York: McGraw-Hill.
560 16 Control of Bioreactor Systems
PIRT, S. J. (1965), Proc. R. SOC.London Ser. B SWARTZ,J. R., COONEY,C. L. (1979), Biotechnol.
163, 224. Bioeng. Symp. 9, 95.
PIRT, S . J., RIGHELATO,R. C. (1967), Appl. Mi- SWARTZ,J. R., WANG, H . Y., COONEY,C. L.,
crobiol. 15, 1284. WANG,D. I. C. (1976), in Abstr. Papers, 5th Znt.
PONTRYAGIN, L. S . , BOLTYANSKII, Y. G . , GAM- Fermentation Symp. (DELLWEG,H., Ed.), p. 29,
KRELIDZE, R. v., MISHCHENKO, E. F. (1962), Berlin.
The Mathematical Theory of Optimal Processes, TAGUCHI,H., HUMPHREY,A. E. (1966), J. Fer-
New York: Wiley-Interscience. ment. Technol. 44, 881.
RADJAI,M. K., HATCH, R. T., CADMAN,T. W. VERES,A., NYESTE,L., KURUCZ, I., KIRCHKNOPF,
(1984), Biotech. Bioenp. Symp. 14, 657. L., SZIGETI,L., HOLLO, J. (1981), Biotechnol.
REUS, M., JEFFERIS,R. P., 111, LEHMANN,J . Bioeng. 23, 391.
(1977), in Workshop Computer Applications in WAGHMARE,R. S . , LIM, H. C. (1981), Znd. Eng.
Fermentation Technology, GBF Monogr. Ser. 3, 107. Chem. Fundam. 20, 361.
ROELS, J. A. (1980), Biotechnol. Bioeng. 22, 2457. WANG, H. Y., COONEY,C. L., WANG, D. I. C.
ROELS,J. A. (1981), Ann. N. Y. Acad. Sci. 369, 113. (1977), Biotechnol. Bioeng. 19, 69.
ROLF, M. J., LIM, H. C. (1982), Enzym. Microb. WANG, H. Y., COONEY,C. L., WANG, D. I. C.
Technol. 4, 370. (1979a), Biotechnol. Bioeng. Symp. 9, 13.
ROLF, M. J., LIM, H. C. (1985), Biotechnol. WANG, H. Y., COONEY,C. L., WANG, D. I. C.
Bioeng. 27, 1236. (1979b), Biotechnol, Bioeng. 21, 975.
SAN,K. Y., STEPHANOPOULOS, G. (1984), Biotech- WEBB,J. C., DDC Hardware and Software, Gener-
nol. Bioeng. 26, 1209. al Electric Company.
SEBORG,D. E., EDGAR,T. F., MELLICHAMP,D. WEIGAND,W. A. (1978), Annu. Rep. Ferment.
A. (1989), Process Dynamics and Control, New Processes 2, 43-72.
York: John Wiley & Sons. WHITAKER,A. (1980), Process Biochem. 15, 10.
SEINFELD,J. H., GAVALAS,G. R., HWANG,M. WILLIAMS,T. J., MOWLE,F. L., WEIGAND,W.
(1969), Znd. Eng. Chem. Fundam. 8, 257. A , , REKLAITIS,G. V., GOODSON,R. E., LIM,
SEINFELD,J. H. (1970), AZChE J. 16, 1016. H. C. (1973), Digital Computer Applications to
SEMONES,G. B., LIM, H. C. (1989), Biotechnol. Process Control, Purdue University, West La-
Bioeng. 33, 16. fayette, IN.
SHINSKEY,F. G. (1979), Process-Control Systems, YUWANA,M., SEBORG,D. E. (1982), AZChE J. 28,
Application, Design, Adjustment, New York: 434.
McGraw-Hill. ZABRISKIE,D. W. (1979), Ann. N. Y. Acad. Sci.
SIEGELL,S. D., GADEN,E. L. (1962), Biotechnol. 326, 223.
Bioeng. 4, 345. ZABRISKIE,D. W., HUMPHREY,A. E. (1978),
SMITH,C. R., MOORE,C. F., BRUNS,D. D. (1981), AZChE J. 24, 138.
Paper TA-7 in Joint Autom. Control Conf., ZABRISKIE,D. W., ARMIGER,W. B., HUMPHREY,
Charlottesville, VA. A. E. (1976), in Abstr. Papers, 5th Znt. Fermen-
SPRIET,J . A., BOTTERMAN,J., DE BUYSER,D. R., tation Symp. (DELLWEG,H., Ed.), p. 91, Ber-
DE VISSCHER,P . L., VANDAMME, E. J. (1982), lin.
Biotechnol. Bioeng. 24, 1065. ZABRISKIE, D. W., ARMIGER,W. B., HUMPHREY,
STEPHANOPOULOS, G., SAN, K.-Y. (1982), in A. E. (1977), in Workshop Computer Applica-
Chemical Engineering-Boston, ACS Symp. Ser. tions in Fermentation Technology, GBF Monogr.
196, 155. Ser. 3, 59.
SVRCEK,W. Y., ELLIOTT, R. F., ZAJIC, J. E. ZIEGLER,J. G., NICHOLS,N. B. (1942), Trans.
(1974), Biotechnol, Bioeng. 16, 827. ASME 64, 759.
17 Automation in Biotechnology
ANDREAS
LUBBERT
Hannover, Federal Republic of Germany
1 Introduction 563
2 Experiences from Initial Applications 564
2.1 Multi-Control Applications 564
2.2 Computer-Aided Measuring and Control 565
2.3 Conclusions from Initial Computer Applications 566
3 Actual Industrial Installations 567
3.1 Quantitative Measures of Reliability 567
3.2 Distributed Control 568
3.3 Redundancy Approach 569
3.4 Software Analogs 570
3.5 Concluding Remarks 571
4 Actually Developed Systems 572
4.1 General Requirements 572
4.2 Architectural Considerations of Modern Automation Systems 575
4.2.1 Nodes in the Local Area Network 577
4.2.2 User Interfacing 578
4.2.3 PCs as Front-End Processors to the User 582
4.3 Special Functionality 582
4.3.1 Processing of Mathematical Relationships between Real-Time Data 582
4.3.2 Off-Line Data 584
4.3.3 Database Operation 584
4.3.4 Control Profile Generation 585
4.3.5 Implementation of Models 585
4.4 New Approaches to Implement a priori Knowledge 588
4.4.1 Expert Systems 588
4.4.1.1 General Aspects 589
4.4.1.2 Software Base 589
4.4.1.3 Examples of Process Automation Discussed in the Literature 591
4.4.1.4 Knowledge Acquisition 593
4.4.1.5 Completeness Problem 594
4.4.1.6 Further Applications of Knowledge-Based Systems in Automation 594
562 17 Automation in Biotechnology
Introduction 563
gies, one cannot find an analog. Hence, in or- bility of control hardware, and to reduce its
der to develop process control, a tool is neces- costs.
sary which is so easy to use that the biotechno- One of the first goals was to exploit the
logist himself can use it directly to implement most prominent advantage of computers, their
his ideas on process control. It must be a main computational power. The basic idea behind
goal in the development of computer applica- this was to improve actual information about
tions in fermentation technology to make a the process by model-supported measure-
universal system of this type available that has ments.
easily mastered user interfaces similar to PC- Controllers were formerly based on analog
based text processing software. Simultaneous- electronics. With the advent of devices based
ly, the aim must be to reduce the fermentation on software running in a computer it seemed
operator’s work load, not to replace him. The furthermore possible to install a nearly unlim-
latter will not be possible in the foreseeable fu- ited number of controllers simply by copying
ture. the program code. With a sufficiently large
The term automation in biotechnology de- number of controllers the expensive computer
serves some comment. Due to the very nature was thought to pay for itself. Consequently,
of the processes discussed, one usually cannot one tried to apply this idea by drastically en-
automate chemical production as fully as a car hancing the flexibility of the fermentation con-
manufacturing assembly. Most computer ac- trol system.
tivities are therefore directed toward support- Initial implementations in industry and
ing the operators in keeping track of their sys- scientific research laboratories, however, re-
tems, providing them with all information nec- vealed numerous difficulties with the new tech-
essary for decision-making, and acquiring the nique. Most of the difficulties seem now to be
capabilities for manipulating the process if forgotten, but they are still of importance for
something is going wrong. today’s applications.
Biotechnology has its own problems specific
to its tasks. For essentially the same reason
that one cannot call for mathematicians to 2.1 Multi-Control Applications
model biochemical kinetics, it is not possible
to look for an information specialist to solve The development of computer control began
problems of computer-aided bioprocess con- during the mid-sixties. MURANOand YAMA-
trol. Consequently, biotechnology must build SHITA (1967) were the first to discuss some
its own measuring and control systems, but kind of supervisory monitoring of fermenta-
based on general developments investigated in tions by means of computers. Distra Products
specialized fields and with the help of measur- Ltd. was a pioneer in its decision to install a
ing theory, physics, and control engineering. computer providing automatic control of an
Hence, it is necessary for a biotechnologist to existing plant for antibiotics production and of
gain some general knowledge in these fields. some new large-capacity fermenters. A total of
114 control loops were implemented at 26 pro-
duction fermenters and 10 seed vessels. GRAY-
SON (1969) reported that, in comparison with
conventional controllers, the main criteria for
2 Experiences from Initial the choice of this novel technique, were:
Applications 0 greater reliability
0 more precise control
0 greater flexibility with computer control
The general aims in the initial application of algorithms
computers to biotechnological practice were to 0 additional data logging and alarm facili-
improve the quality of information on the ac- ties provided by computers
tual process state, and on that basis to improve 0 the possibility of model-supported con-
the fermentation control, to enhance the flexi- trol.
Experiences from Initial Applications 565
The company started with a 13-bit ARCH- rectly. The most interesting was bio-
102 process computer from ELLIOTTAUTO- mass.
MATION (1967) with a basic ferrite core memo- 0 Model-supported control on the basis of
ry of only 8 kWords. Calculation speed, store indirectly measured quantities.
access times, and memory size were very small,
even by today’s home computer standards. Two pioneering groups that began this re-
Moreover, neither higher language compilers, search were the group around A. E. HUMPH-
nor an operating system was available. The REY and his collaborators (e.g., JEFFERISand
machine was programmed in ARCH-102 ma- HUMPHREY,1972; ZABRISKIE,1976) and the
chine code. Even an assembler was not availa- one around D. I. C. WANG(COONEYet al.,
ble. Since at that time computers were known 1968, 1977; WANG et al., 1977; SWARTZ
to be unreliable, one did not rely on computer and COONEY,1978). These groups saw the
control alone, but installed additional conven- most prominent advantage in computer appli-
tional instrumentation and valves so that man- cations to be the extension of measuring and
ual control could be used to run the process control potential by using the computational
during hardware or software breakdowns. power of the machines. They tried to obtain
It was found that control deviations were data on key parameters not measurable on-
significantly smaller than those of convention- line. Since the early efforts to design instru-
al controllers, and data logging as well as ac- ments capable of measuring some of the most
companying alarm facilities were reported to fundamental cultivation process variables such
be advantageous. However, there also ap- as biomass concentration and growth rate, at-
peared some significant disadvantages. The tention has been focused on numerical proce-
main severe problems of this system were: dures that indirectly estimate these parameters
by means of process variable data which can
0 Insufficient reliability of some compo- be measured on line and which are associated
nents of the computer, especially the with culture growth. The latter include viscosi-
electromechanical ones. This pertained ty (SWARTZet al., 1971), heat evaluation
to the main I/O-interfaces - i.e., the (COONEYet al., 1968), carbon dioxide evolu-
multiplexers - which were relay-based at tion (STOUTHAMER,1971), oxygen uptake
that time. (JEFFERISand HUMPHREY,1972), and sub-
0 The program coding took a great deal strate consumption and product synthesis
of time on account of very unsatisfacto- (COONEYet al., 1977). The primary relation-
ry software tools. ships were based on stoichiometric balances of
0 Moreover, since computer memory was all principal chemical components that compose
limited, software could not be structured biomass and products (COONEYet al., 1977),
appropriately, but had to be optimized or on relationships between a single chemical
to minimal code length. Software tools component and culture growth (ZABRISKIE,
could not be produced. Consequently, 1976; ZABRISKIE and HUMPHREY, 1978).
the machine-coded programs became The same applies to control algorithms
very complicated and their modification (WANGet al., 1979). To optimize control ac-
extremely difficult and time-consuming. tions, it is necessary to know how the system
will react. This can only be predicted by means
of a suitable Drocess model. which necessarilv
2.2 Computer-Aided Measuring must be a special one for each individual proc-
ess. An adequate evaluation of such a model
and Control can only be done by means of a computer.
2.3 Conclusions from Initial possible for several reasons. The main disad-
vantage of the early systems was unreliability.
Computer Applications The hoped-for reduction in the number of
process operators could therefore not be
Interestingly, nearly all the aims still on the reached. On the contrary, additional, better
agenda of workers active in automation in bio- educated personnel were required to keep ma-
technology were already formulated by the chines running and to adapt the machines to
first investigators (HUMPHREY,1977a, b): the processes. Since computer downtime
meant breakdown of all controllers, it was not
0 Data acquisition, storage, analysis, and possible to keep product quality as uniform as
reduction was originally expected.
0 Monitoring of fermentation plants by Today, most controllers work internally
multichannel data logging connected with digital electronic components, and they
with a flexible automatic alarm system are once again installed in separate units,
0 Extension of the limited on-line measur- which are incorporated into a system of front-
ing possibilities by model-supported end processors. Internally, however, they work
methods, especially for the on-line meas- on the same basis as analog PID-controllers.
urement of biomass This computer control often degenerated to an
0 Direct digital control with many control at best more economical substitute for the ear-
loops working simultaneously, coupled ly analog controllers. Only very recently have
or running independently more advanced controllers (as compared to
0 Model-supported control to enhance the simple PID-controllers) become available for
performance of conventional controllers process automation systems.
0 On-line process optimization. The limited control obtained in the first ap-
plications was obtained at high cost in money
From the fact that some of these goals have as well as in manpower. The cost was much
not yet been reached in a generally satisfactory too high for most companies in the fermenta-
way, it is evident that there are many practical tion industry. They also saw that the use of a
obstacles. On the other hand, even during first single computer to control a whole plant
attempts it proved possible to operate a large would constitute a considerable risk of pro-
number of control loops simultaneously in a duction breakdown. It was essentially this low
single computer system. GRAYSON (1969) reliability together with the high cost in man-
stated that control accuracy had already been power that led to fierce resistance in industry
improved significantly compared to conven- against a wide application of computer control
tional analog computers. in fermentation plants. Another significant
Nowadays, the early assumption that digital reason for the slow acceptance of the rapidly
control could be installed more cheaply than developing computers in fermentation plants
analog techniques is accepted as true, but it was that on-line sensors were missing for many
was not so at that time, not if one took into essential process variables such as biomass,
account the overall cost of single central proc- substrate and product concentrations etc.
ess computers, including their installation, (CHATTAWAYand STEPHANOPOULOS, 1989).
software development, and maintenance. As Moreover, difficulties in analytically modelling
time progressed, the probability of failure of fermentation processes prevented playing the
microelectronic components dropped exponen- highest trump in computers for any process
tially by several orders of magnitude. Thus, automation: model-supported measuring and
this risk has become significantly smaller, but control.
it still exists and must be considered when de- In general, the more sophisticated aims were
veloping a new computer-based control sys- unreachable with the computer systems availa-
tem. As compared to today’s computer prices, ble at that time. For example, model-sup-
the early machines were extremely expensive ported measurements of biomass proved to be
tools; investments were thus very high. Com- much more complicated than originally
pensation by a corresponding return was not thought, since correct and sufficiently detailed
Actual Industrial Installations 561
models were not available, and if they were, having less software flexibility. Redun-
the necessary data or other process parameters dancy must additionally be applied in
could often not be supplied with sufficient ac- these systems.
curacy. Thus, smaller steps in the development
of bioreactor automation had to be chosen In the software of such systems the term
with aims that could be reached within reason- free programming was eliminated as far as
able time intervals and that were easy to sur- possible, since programming by users was
vey. To change the control strategy of a whole identified as a major source of failures. In-
plant with a new technique, as was done by stead, the software was distributed in a closed
Dista Products, was then and remains a high- form, containing routines necessary to handle
risk endeavor. From the industrial point of standard tasks. For a concrete application, the
view, the advanced methods of measuring and software modules must be linked together and
control seemed to be unsuccessful. supplied with actual process parameters. This
However, first experiences were not suffi- activity was called configuration and was care-
ciently discouraging to stop further attempts at fully distinguished from programming.
improving computer-based process control.
Especially desirable was the possibility of run-
ning a large production system from a central 3.1 Quantitative M~~~~~~~
monitoring system with improved control
quality. of Reliability
Reliability and performance of the comput-
ers turned out to be the main obstacle to wide
3 Actual Industrial acceptance of computers in industry. Control
engineering companies developed new systems
Installations to respond to the bad experiences suffered in
industrial practice. These ideas, subsequently
developed, still rule today’s computer installa-
Actual working industrial installations have tions in fermentation plant control.
been constructed by responding to the main Since reliability is the decisive demand of an
failures suffered by the pioneers. Since the industrial control system, this topic deserved
drawbacks of the first installations were se- to be discussed first. The breakdown of an au-
vere, drastic measures were chosen to reduce tomation system is a random event. Hence, a
cost and to enhance reliability. Essentially two quantitative treatment of reliability can only
paths were taken in the hardware field. The be made from a statistical base. As generally
first was to distribute several parts of the sys- accepted, one takes the mean relative time, A ,
tems’ load to different processors, and the sec- during which the particular system is running
ond was to install system redundancy wherever correctly, to quantify reliability. This is termed
it could be paid for. availability A :
From these facts, the following two hard-
ware philosophies emerged: A = MTBF/(MTBF+ MTTR)
0 The first saw the advantages of powerful where MTBF represents the initial letters of
process computers in their computing “mean time between failures”, and MTTR the
power and stressed the software argu- “mean time to repair” the system. It is obvious
ments. A simple redundancy approach that MTBF must be maximized and MTTR
was preferred, which was realized minimized in order to obtain optimal availabil-
through tandem computers. ity.
0 The second preferred widely distributed T o reduce MTTR, one not only needs well-
systems consisting of smaller computers. educated field service specialists who find and
These have the advantages of hardware repair failures very quickly, but also a stock
building-block systems, but suffer from that contains all required components that
568 17 Automation in Biotechnology
The major justification for large invest- is that the bus systems within most distributed
ments in process control systems, including the control systems d o not meet international
required on-line field measuring devices, is re- standards. Moreover, the software does not
duction of overall costs by dispensing with support the integration of components from
shift-workers, reducing substrates and energy, different vendors. This situation came about
and avoiding costly plant downtimes as far as because most current standards did not exist at
possible (LITZ, 1989). The reduction of per- the time the existing control systems were de-
sonnel costs is often the most directly account- veloped. The development of classical automa-
able and often the biggest point. Advantages tion systems dates back to the seventies and
in energy and substrate consumption are also the early eighties.
predictable. Some advantages are not easy to Much money has been invested by many
quantify in concrete numbers, i.e., the greater companies in developing their own special
flexibility with respect to changes to other standards which are not compatible with oth-
products, enhancement of product quality, ers (e.g., DOHMEN,1990). It is evident that
and the achievement of more uniform quality, many do not like to cancel their developments
since they influence the overall result only indi- and change direction to meet international
rectly. standards. This is a disadvantage to custom-
One very prominent feature of the control ers, since these systems are more or less closed
systems described is that they rely predomi- forcing consumers to buy components of one
nantly on ready-to-use software. Configura- vendor’s system only. Moreover, often only a
tion has replaced programming in order to re- very limited number of basic components is
flect the often-claimed simplicity with which available. This restricts the customer’s flexibi-
system activities, provided for in the software, lity in building an optimal system, since he
could be made available. Many standard con- cannot combine the components best suited to
trol actions can be implemented in this way. his special needs. In this way there arises a de-
Inspection of systems in actual use shows that pendency on the initial vendor.
closed-loop control is more often supported, For new installations it is therefore recom-
sometimes at the cost of open-loop control. mended to look for systems based as far as
Free programming possibilities, however, were possible on universal hardware and software
restricted or disallowed. This makes the inte- components which meet international indus-
gration of advanced measuring and control ap- trial standards. There are several such choices
plications difficult. The same applies to the in- now on the market.
tegration of all other user-defined software
components. Basically, most plants are con-
trolled as they were some 40 years ago using
PID algorithms adjusted according to ZIEG-
LER and NICHOLS(1943) or variants. 4 Actually Developed
The application of the described systems is
suited for productions that are already optim-
Systems
ized and are not designed for continuous im-
provement. However, in biotechnology, espe-
cially in research and development depart- 4.1 General Requirements
ments, things are always changing. Thus, flexi-
bility of the automation systems is at least as Automation systems currently installed in
important as availability. the industry are optimized for high reliability.
One of the main drawbacks of the systems This was mainly achieved through two design
discussed above is that they cannot easily be principles. The first is modular design, which
combined with hardware and software compo- led to distributed automation systems. The
nents of other manufacturers in process auto- second, supported by this modularity, is the
mation. A case in point might be the coupling redundancy concept. The advantages of modu-
of a general purpose computer for process si- larity are not restricted to the hardware, how-
mulation or optimization. The main problem ever, but can also be extended to the software
Actually Developed Systems 513
domain. Based on the idea of a building block is derived from the information gained on dif-
system, modular software systems contain a ferent levels.
fixed number of modules, which can be confi- The spectrum of different tasks processed in
gured and parametrized in order to adapt the modern automation systems is still expanding.
software to a given automation task. Such sys- At the same time, the computational require-
tems can fulfill the requirements of system re- ments for individual tasks, e.g., process con-
liability. trol, do not become smaller. This expansion in
In actual process automation systems, much quality and quantity of tasks to be handled
more data about the processes are accumu- brought the conventional process automation
lated and generated than was formerly possi- systems to their limits. The tasks within auto-
ble. In order to use these data adequately, in- mation systems require more computing power
telligent management is required. Information than current computer nodes can supply. A
processing also rapidly developed in various straightforward way to extend the perform-
other parts of industrial companies, e.g., in ance of these systems is to use more power-
the sales departments. There are many data of ful nodes in the distributed systems (e.g.,
importance to more than one department in a FEWKES,1988) and to increase the number of
company. Thus, it is advantageous to integrate “intelligent” components. This way, however,
computers into company-wide information is circumscribed, since the nodes of most cur-
processing systems. The structure of such com- rent automation systems are constructed by
prehensive systems is everywhere similar to the vendors and cannot be freely exchanged by
that proposed by NAMUR, as sketched in Fig. units of higher performance. Even if this
2. Process automation, then, is only one part would be possible on the hardware level, most
of the whole. Since the individual systems vendor companies would not be able to extend
within the different levels are all distributed their software bases, which usually are also in-
systems, development proceeds in the direction dividually written.
of creating networks within networks. The It can be seen from the software develop-
connection of the local area networks installed ments of the last ten years that even very large
within different levels of a company’s hierar- software development companies are not able
chy requires efficient interfaces between the in- to develop basic software systems at a rate
dividual systems in order to ensure that benefit comparable to the development of new-genera-
tion processors in microelectronics. Hardware
developments thus orient themselves to stand-
Function Quality Level ard software bases, e.g., new generation proc-
essors must be operated with established oper-
Control of Dispositive Company Control
I ating systems. Consequently, one must base
new automation systems on hardware and
Control of
Control of
rn
Production Control
software components that are involved in these
development flows. The development of proc-
ess automation systems will probably move in
the direction of using more and more standard
components, which may be developed in com-
pletely different environments and are sup-
ported by large interest groups rather than by
Measuring and comparatively small control engineering com-
Control panies.
~ Operational Performance is the primary criterion for the
Sensor-Actor
Functions
IField Levcl
developed by one company, while an advanced ing the biotechnologists. Integration plays a
closed-loop controller of another firm may be special role in this system. In other words,
more suitable. In most instances it is better to data from different sources are integrated by
acquire a database system from a third sup- CIF, measuring and control subsystems of dif-
plier. Hence, automation systems are required ferent vendors can be integrated, and data
in which a customer is to put those compo- analysis software from different developers
nents together which optimally meet his de- can be readily adapted. Finally, CIF itself can
mands. Such systems are referred to as “open be integrated into a company-wide informa-
systems”. tion processing system.
It is obvious that such open systems can Although several biotechnological cultiva-
only work if certain standards are observed. tions can be controlled completely indepen-
This does not necessarily mean that the differ- dently of each other, it is also possible to inter-
ent nodes must be standard components. That connect different applications. CIF is a strictly
would restrict development too severely. It modular software package running under the
merely denotes that communication between operating system VMS. Its basic hardware
the system’s components must be standar- components are VAX computers, which are
dized, a prerequisite if cooperation is to be ef- the recognized standards among more power-
ficient. ful 32-bit computers. These computers are
As compared with modern software compo- widely distributed in chemical and biochemical
nents available to users of PCs and worksta- industries. CIF software runs on single com-
tions, the active user support in currently used puters and can be distributed onto different
process automation systems is highly unsatis- nodes in a local area network (LAN) as well.
factory. The ease with which automation sys- For initial installations, one can use a single
tems can be handled by process personnel must workstation, and one can expand the system
be drastically increased. It is essential that the later by setting up additional computers.
user be supported by the system and not hin- The different components in a system con-
dered by complex operating procedures. Much taining more than one computer are arranged
can be learned from developments in the P C along a serial bus system based on Ethernet
field concerning convenient user interfaces, es- technology (IEEE 802.3). This network system
pecially graphical visualization techniques, to is open to computers of different vendors,
mention only one aspect. Summarizing, one since Ethernet has emerged as the most widely
can state that systems must be optimized to accepted standard for such local area net-
support the user. works. Many vendors supply devices with
Current development will be discussed by Ethernet interfaces and appropriate network
means of some concrete examples, in which software, which can be used to integrate the
the process management system CIF will be devices into VAX computer networks.
the main accompanying example to guide the In an actual installation of CIF, the VAX
discussion. computers must be complemented by “front-
end processors” optimized to data acquisition
and basic control tasks. Front-end processors
Accompanying example: CIF developed by individual fermenter manufac-
turers can be used, e.g., the DCU of B. Braun
CIF is a modern process automation man- Melsungen AG. Often, programmable con-
agement system that can be used to automate trollers such as the Simatic S5, supplied by
medium-scale biotechnical pilot and produc- Siemens, are still installed at the plants, and it
tion units. CIF stands for “computer inte- is important to be able to integrate them into
grated fermentation”. It is set up to suit the the automation system. Many other types of
special requirements of biotechnology; thus front-end processors used in biotechnical
the idea behind it is to enhance the perform- plants can be adapted.
ance of universal automation systems by con- Fig. 3 depicts the general software structure
centrating on special problem-oriented proc- of CIF. Central components are the real-time
essing concepts of biotechnology and support- database called POOL, together with its data
Actually Developed Systems 515
strument systems are often automated in them- Information exchange between different
selves. It is quite clear that these small-scale items requires an agreement about a common
automation systems must be incorporated into base of communication. For the communica-
the larger process automation system as a tion between automation system components,
whole, instead of trying to develop new auto- several standards have emerged. They are of
mation software for these devices within the widely differing performance, and so cannot
larger system. Hence, smaller systems must be be applied equally efficiently throughout the
capable of integration into larger ones. automation system. For data transmission be-
The same reasoning applies to the develop- tween a measuring device and its host comput-
ment of software for other basic automation er, a 4-20 mA serial connection according to
system components, e.g., for programmable RS232 may be sufficient in many cases, where-
controllers. System integration is of increasing as such a connection is completely inappro-
importance, since there is a general trend of priate if used to connect different 32-bit com-
migration of microprocessors into measuring puters. Connections between computers are
and control devices, i.e., closer to the process. not only different in their performance, but
It is necessary to make the process automation also in cost. Numerous solutions have been
management systems open for the incorpora- discussed within the standards organizations,
tion of many different internally automated but no general agreement is in sight. In the
components from different vendors. The meantime, customers could not suspend the
choice of measuring and control components development of their automation systems, so
must above all be guided by their performance they created de-fucto standards simply by their
in relation to the task to be performed, and choices. For communication between comput-
not primarily by other arguments. Hence, a ers in local area networks, only a very small
modern process automation management sys- number of network systems in practice proved
tem must supply a common base for integra- to work with sufficient reliability; of these,
tion of a wide variety of different devices. This Ethernet (IEEE 802.3) and SNA (IEEE 802.5)
is essentially a question of communication be- emerged as de-fucto standards (CLEAVELAND,
tween the base and its components. 1989), although they were once severely criti-
Actually Developed Systems 571
cized by network experts. Such a de-facto 4.2.1 Nodes in the Local Area
standardization is not new on the computer
scene. The IBM-PC is another example, Network
which, as is well known, became an industrial
standard owing to the fact that numerous cus- The choice of the basic computer compo-
tomers chose it as their favorite system. nents of an automation system must be guided
Although actual development moves strictly by practical considerations. The application
in the direction of open systems, the number software providing the functionality of com-
of different components in a network must be puter automation systems must be the first cri-
kept minimal for economic reasons, since terion in the choice of a computer system. This
maintenance costs rise with the number of dif- software determines what computers are re-
ferent system components. Thus, there are quired.
conflicting requirements, for each of which a In smaller applications that can be solved
compromise must be found. with small software packages, PC-systems are
The hierarchy in modern automation sys- suitable. There are many small automation
tems is of a logical nature. Concerning net- systems on the market written for single PC-
work hardware, democratic networks such as systems. These systems can be classified by the
Ethernet are preferred. In Ethernet, all nodes operating systems under which they run. The
are connected to a bus system and have the smallest systems are based on the well-known
same chance to access it. The hierarchy is es- PC-operating system MS-DOS, which is a sin-
tablished by an authorizational levels system gle-task operating system. More flexible sys-
for persons logged in at the workstation nodes tems use multitask operating systems of vary-
and by a priority system for different software ing performance. Concurrent DOS is the smal-
modules. To establish a well-structured prior- lest of these, OS/2 is a more extended one,
ity system within a distributed process man- and some PCs even use UNIX. Since conve-
agement system, powerful operating systems nience in operating systems usually can be
are required on the nodes that support such achieved only at the expense of computing
structure. power, the latter systems require more power-
The large centralized control rooms, charac- ful PC-versions, e.g., ones based on Intel’s
teristic of current automation systems, are be- 80386/80486 or Motorola’s 68020/68030 proc-
coming obsolete, since in the new network sys- essors.
tems all information and manipulation capa- There is a continuous transition between
bilities are available everywhere within the net- UNIX-based PCs and workstations. Worksta-
work. Control rooms are only necessary where tions can be viewed simply as more powerful
the operators must be protected from extreme single-user multitask systems with advanced
environmental conditions, such as humidity, graphic capabilities. In this area the UNIX-
excessive noise, or extreme temperatures. systems provided by different vendors are in
Modern local area networks permit the process strong competition with Apple’s Macintosh
operator or the bioprocess engineer to plug and especially DEC’s VAXstations, both run-
into workstations wherever needed. One day ning under their own powerful operating sys-
they become necessary near the reactor, the tems. The different workstation operating sys-
next near a downstream processing unit; at tems differ greatly in performance, so differ-
other times they must be used in a lab, or in ent criteria are applied in choosing among
the process engineer’s office. This flexibility is them. The system of the Macintosh has been
a primary criterion that saves considerable optimized for user communication. Apple set
time for the personnel. an important landmark on this subject, much
In the accompanying example, CIF, Ether- discussed and further developed by many oth-
net was chosen as the basic network technolo- er software developers. The VMS-operating
gy. It actually provides the broadest base for system of DEC is the most convenient of the
different subsystems available on the market. general operating systems and has the shortest
real-time response time of the systems men-
tioned. UNIX, however, has the advantage of
578 17 Automation in Biotechnology
being the system with the best portability fea- network of VAX computers it is possible si-
tures among the multitask systems discussed multaneously to use some nodes running under
above. VMS and others under DEC's version of
In CIF, the VMS operating system was UNIX, known as ULTRIX, or even MS-DOS
chosen, since comfort and real-time perform- machines (cf. Fig. 5). The general direction of
ance of the operating system were rated higher development is to further reduce the operating
than portability. Furthermore, for VAX com- system differences in local area networks by
puters a well-performing network software, supplying universal software environments
DECnet, is available. Moreover, VAX com- that support the exchange of information be-
puters have been particularly favored in indus- tween different tasks running on different
trial applications. The computer family con- nodes under different operating systems. An
cept of DEC is of great practical advantage. important example is X-Windows (SCHEIF-
From workstations up to mainframe comput- LER, 1987), which is becoming the standard
ers, many VAX computers of different per- base for user interfaces within local area net-
formance levels are available, and all can be works.
operated under the same operating system,
VMS, so they do not require changes in the ap-
plication software. Another criterion is that 4.2.2 User Interfacing
these computers have been proved to be com-
patible with successive developments for the The fact that the system operator is indis-
last 15 years. Such a continuous development pensible in the control of biotechnical cultiva-
is of considerable economic importance, since tion processes leads to the conclusion that he
the system can be extended or adapted step- must be supported by the computer system.
wise to the actual development over many Most of today's digital control systems (DCS)
years. do not have adequate operator interface capa-
The choice of a special operating system bilities. T o facilitate the contact between man
loses its significance as a result of the latest de- and machine, the so-called user surface or in-
velopment in local area network software. In a terface must be optimized to reduce the infor-
PC DECstation VAXstation
MS-DOS UNIX VMS
ETHERNET
DECnet
P R O C E S S
Fig. 5. Example of the incorporation of nodes running under different operating
systems in the CIF-network in the context of an open network. The nodes be-
long to the nodes on the process data management level as well as to the front-
end computer level.
Actually Developed Systems 579
protocol, which separates the virtual interface ing memory, forcing the windows to be stored
of the application program from the server im- on the system disk. The degree of user friendli-
plementation. The latter does not necessarily ness, depends heavily on the speed with which
run on the same computer, and it alone must the process information can be displayed on
be specialized for the concrete display hard- the video screens (e.g., BAUMAN,1989).
ware. With such a separation of the (virtual) The ability to display on the user’s screen
I/O of an application program from the soft- only the information necessary for judging the
ware serving a special display hardware, the actual process state is a matter of experience
application becomes more portable. In the with the process in question. A context-sensi-
case of process management systems, this fea- tive selection of information requires at least
ture can be exploited in two directions. First, a heuristic knowledge of the process. Several
special application task can be interchanged groups are working on the support of man-
between the various computers within the net- process interfaces by expert-knowledge-based
work, while the operator stays at the same vi- systems (BAR and ZEITZ, 1989; NAKAMORI,
deo front end. Secondly, a process can be ob- 1989), as mentioned later in this chapter.
served from different nodes within the net- Such standardized user interfaces eliminate
work using a man-process interface, which has many differences in the appearance of operat-
the same appearance for the operator. ing systems to the users, and they reduce dif-
Windowing allows operators to shuttle be- ferences in application systems to functional
tween different application programs. Win- ones. In this way different software compo-
dows can be depicted on a video screen in nents are virtually integrated from the user’s
overlapping or tiled fashion. This allows for viewpoint. Since there is no operating system
more transparent process monitoring. But X- available that can be used at all levels of an
Windows does more than windowing. X-Win- automation system, it is more important to
dows is a tool that can be used as the basis for have a standardized man-machine interface
the development of network-wide user inter- than to have a standard operating system.
faces regardless of operating system compati- Thus, the days of lengthy discussions on com-
bility among the nodes in the network. It patible operating systems seem to be num-
therefore promotes uniformity of user inter- bered. Only those will survive which support
face functionality in different application pro- standard user interface systems. It now ap-
grams in order to present them via similar pears that X-Windows or OSF Motif will be
screens, menus, and graphics. the standard for the next few years.
All these convenient features and the flexi- The advantage of X-Windows-based user
bility of operation naturally come at a price. interface stations has been recognized by many
Workstations for X-Windows applications companies, and they are actively working to
need large, high-resolution video displays for incorporate them into their systems. One cur-
the simultaneous representation of many win- rent example is the series 40 system of Bailey
dows. Since the on-screen windows can, if nec- Controls (BAILEYCONTROLS,1989).
essary, be superimposed upon other windows, The user interface of CIF has been develop-
their contents must be backed up. Thus, large ed directly with X-Windows and thus exploits
memories are required to hold simultaneously its full functionality. It allows the user to de-
the overlapped parts of the display in the back- sign his own individual man-process interface
ground in order to allow for a rapid restora- without restrictions as to special terminals and
tion of hidden parts if the foreground windows PC- or workstation screens. Moreover, the
are removed. If one uses several windows on a management of network-transparent access to
high resolution video display, several MBytes the different application programs is taken
of working memories are required to hold the over by CIF.
graphic displays in RAM. Hence, more than As a classic example of a non-trivial X-Win-
10 MBytes of memory is not unusual within a dows application in CIF, a typical display is
modern workstation. It should be mentioned shown in Fig. 7. The operator can acquire a
that such window-based systems become very quick overview of the system by arranging a
slow if the server does not have enough work- small window for every key process signal on
Actually Developed Systems 58 1
feast fermentation
I
Fig. 7. Typical example of an X-Windows operator surface in CIF. The upper win-
dow contains smaller windows through which the operator can monitor key varia-
bles. These windows contain so-called push buttons, i.e., miniaturized plots of the
signal records taken over short, medium, or long time segments. These are contin-
uously updated. These small windows also contain the current values of the varia-
bles. If the push buttons are clicked with the mouse, the plots can be drawn to an
arbitrary size as shown in the lower part of the figure. The plots are then fully
scaled and captioned.
the video screen. These windows in turn con- detail, he can, with a mouse click at the small
tain smaller windows, each of which contains a window push-button icon, immediately zoom
history of the corresponding signal in the form the signal representation to any size represent-
of a chart recorder strip observed from a dis- able on his video screen. These larger represen-
tance. These miniaturized histories are contin- tations are then fully scaled, and by means of
uously updated. In this way trends can easily cross-hairs the operator can examine any parti-
be recognized from the curves. The windows cular signal value within the display.
also contain the current value of the measuring X-Windows supports the flexibility of CIF
variable. If the operator is interested in more to run on different hardware configurations,
582 17 Automation in Biotechnology
both on a single VAX computer and on differ- gently required, since many state variables
ent systems within a local area network. Of cannot be measured directly. Furthermore,
course, all the software (including the X serv- many key variables can only be measured off-
er) can run on a single VAXstation smaller ap- line; the measuring results then must be re-
plications. The second possibility is that the turned into the system and used for process
CIF system runs on one VAX and the server control as promptly as possible.
on another station. Finally, it is also possible
to distribute the CIF software on different
nodes in a local area network independently of 4.3.1 Processing of Mathematical
the distribution of the X servers. Relationships between Real-Time
Data
4.2.3 PCs as Front-End Processors
The performance of a process management
to the User system is determined not by its ability to ac-
quire, display, and store the process data ap-
It is a general trend in the development of propriately, but primarily by its capability of
distributed process control systems to integrate manipulating the data in order to extract more
PCs as front-end processors for the user (e.g., valuable information about the process state,
Zoz, 1990). One reason is to make their well- and to react to deviations from the predefined
developed graphical user interfaces available path of operation by calculating control ac-
to the larger systems, e.g., window surfaces. tions and performing them. This principally
High-resolution video screens like the VGA- requires a means to relate data from one, two,
standard in PCs are unusual in currently work- or several measured signals by mathematical
ing process control systems (LITZ, 1990). Be- functions or algorithms.
cause of the large market, much more man- Such on-line calculations between process
power can usually be put into the development signals must be possible without a change in
of PC-software. In the future, many software the basic software of the application system.
components familiar to the growing commu- The relationships necessary in a special appli-
nity of PC-users will be available in process cation cannot be contained in a general auto-
control systems, e.g., spread sheet programs mation system. They may change from appli-
such as Lotus 1-2-3 or Symphony, either as cation to application. Hence, they must be de-
software packages on the normal nodes of the fined by the user and be made available to the
systems or via a software interface on PCs system. There are essentially two principal
within the local area network. means by which relationships can be imple-
For the VMS system, on which CIF is mented: The first is to attach a separate user-
based, a fully PC-compatible Lotus 1-2-3 is supplied task to the system, which is able to
available. This is compatible with DEC win- access the process data, make the manipula-
dows. Other software packages such as the oft- tions on them, and transfer the results back
en-used text editing system MS-Word can be into the base-line system.
run on VAX/VMS-systems with the help of an The second means is a math-function proc-
MS-DOS simulator. Using the tool PCSA, essing component in the automation system,
PC-files can be transformed transparently into which can be supplied with the functional rela-
VMS-files, which then can be transferred tionships to be processed at run time. It is ad-
through the network. vantageous, especially in pilot-scale applica-
tions, to install, change, and remove such rela-
tionships as a run-time activity of the system.
4.3 Special Functionality Such real-time math-function signal-proc-
essing facilities are available in several sys-
A high degree of functionality is required of tems. Usually they work as interpreters, proc-
automation systems in biotechnical applica- essing the instructions introduced by the oper-
tions. Model-supported measurements are ur- ator into a prepared buffer. Unfortunately, in-
Actually Developed Systems 583
OF ._._
-. un
Ilanagsment
a L)h:
F
161
II II I /I
I
I
I
Fig. 8. User interface provided in the process con- level language such as FORTRAN. The other win-
trol system CIF to easily program math-functions. dows around provide additional information for the
They are directly written into the window at the cen- user.
ter of the video screen depicted here using a higher-
terpreters d o not have good real-time proper- the interpretation speed for such source code
ties. Thus, to decrease the execution time for segments, precompilers have been developed
the statements by the formula interpreter, the to translate higher-level language code into a
syntax of the code must be optimized. In order compact, quickly interpretable auxiliary code.
to obtain short real-time response times, it was An additional way was developed for CIF, the
most common to choose a reversed Polish no- cited example. Exploiting the convenient oper-
tation in process control systems, recalling the ating system VMS, it was possible to find a way
programming of H P pocket calculators. Such by which modules containing the mathematical
coding may be carried out quickly, but it is not functional relationships between signals and all
simple to use if larger code lengths are re- other data available could be written in higher
quired. Thus, higher-level programming lan- programming languages, supported by a com-
guages were preferred by the users. To increase fortable user interface as shown in Fig. 8. All lan-
584 17 Automation in Biotechnology
guages available in VAX computers are us- Of course, the process management system
able. The source code segments are automati- must be able to utilize the off-line data imme-
cally complemented by code to produce com- diately. First of all, the data must be stored in
plete subroutines, which can be compiled with essentially the same way as on-line data taken
standard compilers and then linked into the at larger sampling intervals and arriving at the
running automation system. Such compiled computer with larger time delays. Consequent-
code is the fastest way to process statements, ly, they must be treated just as the latter are.
permitting the relationships to be processed Data with low sampling frequencies and long
much faster than any interpreter can work. time delays often arise in biotechnology. Near-
All on-line and off-line data, as well as all ly all the more sophisticated measuring tech-
parameter values, can be used in CIF’s func- niques, e.g., HPLC, FIA, autoanalyzer, etc.,
tion processing module. Even complicated, exhibit this disadvantage.
special higher level control algorithms can be In CIF, the off-line data are stored in the
installed. They not only can be caused to gen- same way as those measuring values that stay
erate key auxiliary variables that more clearly fairly stable during the fermentation run.
depict the system’s state, but also can reduce Their values can be used directly by the math-
the redundancy of data and compare the sys- function processing system of CIF, which in
tem’s actual state with the required one. In this the evaluation of formulas takes the most cur-
way the math-function processing facility of rent measuring value available. Where this is
CIF can be used as a decision-support tool. not possible, extrapolations can be performed
on the basis of user-defined model equations.
4.3.2 Off-Line Data
4.3.3 Database Operation
Since many key variables in fermentations
are measured off-line, it is necessary to feed In addition to the real-time databases in
them back into the automation system as process data management systems, considera-
promptly as possible to ensure their proper ble benefit can be derived by installing conven-
use. In smaller laboratories, the off-line data tional, large mass-storage-based relational da-
must be entered into the computer manually tabases to make proper use of much more data
via the computer terminal; in larger ones, it is than can be stored in the working memory of
necessary to couple the laboratory information the automation system.
management systems (LIMS), usually opera- Since process control systems generate in-
tive in analytical laboratories, to the process creasing quantities of data, a problem arises in
automation management systems. storing data’in such a way that they can be re-
For manual off-line data input it is first of covered quickly when needed in subsequent
all necessary to have a suitable user interface fermentations or during an off-line analysis. It
with which the plant personnel can enter the is important to be able to use standard data-
data in a way that avoids input failures as base systems, which can guarantee a high de-
completely as possible. A simple-to-use menu gree of data security with respect to access and
should be displayed on the video screen, and modification of data. The informational value
the software behind it should be failure-toler- of the process data from fermentations accu-
ant. This means that all inputs must be mulated over some years should not be under-
checked immediately for completeness and estimated.
plausibility, so that the operator can directly It is often desirable to compare data from
recognize possible input failures and correct earlier fermentations with those arising during
them. Moreover, it is desirable to inform the a current run. This requires the display of his-
control system about every sampling event torical data onto the video screens during the
during the off-line analysis, to record the exact monitoring of the actually sampled data. His-
sampling time, and to remind the operator af- tories must thus be accessible in a simple
ter a certain period of time to look for the re- way via an appropriate easy-to-use operator
sult of the off-line analyses. surface.
Actually Developed Systems 585
To facilitate access to the data, it is impera- justable parameters such as temperature, pH,
tive to use a standard interface that is indepen- and pOz along the profiles of the sample.
dent of the underlying database systems. This Noisy data from older fermentations obviously
permits the specific system employed in the cannot be used directly. Thus, the control pro-
user’s company to be easily coupled. Such an files must be appropriately prepared by means
interface must, therefore, meet common of filters.
standards. The “standard query language” CIF, therefore, supplies a filter bank that
(SQL) provides such a standard for the most allows the biotechnologist to simply refurbish
frequently used relational database systems historical fermentation data taken from the
such as Oracle, RDB, d-Base, etc. (e.g., database via the SQL-interface. As indicated
JONES,1989). It decreases the sensitivity of in Fig. 9, several filter algorithms are supplied.
application programs to details of the data If none of the standard filters satisfies the
storage techniques and vice versa, since it acts user, it is possible to define the control profile
as a buffer between the two. The present ver- manually by setting certain estimated data
sion of SQL is not very convenient, but new points into the plot using the mouse and subse-
versions are to be expected in the near future. quently starting a spline interpolation to finish
Since process control applications place with a smooth control profile. This can be sent
high demands with respect to the mass of data to the POOL, from which it can be accessed
to be stored and the required speed of data ac- by the controllers.
cess, specific data-storage techniques must be
used. In CIF the database is assumed to con-
tain only that information necessary to find 4.3.5 Implementation of Models
the real-time data and to supply additional in-
formation required for their analyses, while Advanced control is usually thought to be
the original sampled measuring values are model-supported. Appropriate control algo-
stored in files accessible at a much higher rithms must be specially programmed routines,
speed. since they cannot be supplied by a general
CIF provides a functional module for access process automation system. Since currently in-
to the data within a relational database by stalled automation systems do not provide
means of an SQL-based interface (cf. Fig. 3). general programming environments, addition-
Hence, this module is not restricted to a specif- al computers are required which must be fitted
ic database system. This user interface is main- into the automation system. Their task is to
ly used to quickly find all relevant data on the supply a base for the installation of extended
currently running fermentations as well as control programs, since such algorithms usual-
from all completed fermentations for compari- ly require a universal software base as pro-
son purposes. The data in the database can be vided by general computer operating systems
accessed and extended, even beyond the end of such as UNIX, VMS, or comparable systems.
the fermentation. There are essentially two ways in which
models can be attached to process manage-
ment systems. The first is to supply an open
4.3.4 Control Profile Generation software interface to a programming environ-
ment linked to the user-written application
One of the tasks most often required of the program. In this way, a fast connection be-
bioprocess engineer in pilot plants is to im- tween the user program and the automation
prove the control sequence for the next fer-, system can be established. In distributed sys-
mentation in order to optimize the process. In tems the interfacing routines are required to be
most cases the control profiles are determined network-transparent.
empirically according to experience from pre- The possibility of using the math-function
vious cultivations. Often a model fermentation interpreters already discussed is very limited,
is chosen from preceding experiments as a since interpreters are too slow for real-time
guide for the subsequent runs. In this case it is model simulations. This is because the original
obvious that one must try to control the ad- equations must be evaluated too often in com-
17 Automation in Biotechnology
7 B
fl
1
I
Actually Developed Systems 581
Controls
b
Help
SUBrn-
Modal Parunator
1: BIOMASS-
kd : I 0.364~aE-OlI
YS : I 1.1999 1
Fig. 10. Operator interface to the model bank in the measured process variables. The order of both
CIF. In the upper left window the user can click a must be matched by mouse clicks. Beneath that ap-
desired model and a solver. Below the window the pear the proposed start parameters for the estima-
process data are displayed. At the right-hand side tion procedure.
the model variables appear, which must be fitted to
plex applications. Another possibility is to These models are stored in the computer to-
provide a means of incorporating model equa- gether with information on appropriate solv-
tions directly into the automation system via a ing algorithms and a set of start values for the
prepared user interface; this is only economi- parameter estimation procedure. Model banks
cal if a compiled code is used in the function can be used on-line to obtain information on
processing instead of the interpretation of the the state of the process as well as off-line for
equation. process data analysis.
A notable effort to make simple models, In CIF, a similar model bank is available to
which is discussed in the literature available to the biotechnologist (MATHISZIKet al., 1990).
the biotechnologist in industry, is the construc- It is restricted to simple models that are uni-
tion of a model base running on PCs. I n t h i s versal enough to be of general importance.
development, many models which may be of The model base can be changed and extended
interest for cultivations in a special plant can by experienced users. Fig. 10 shows an exam-
be put together in a model bank and selected ple of its user interface to give an idea of the
for a special application by a simple interface. simplicity of such a tool.
588 I7 Automation in Biotechnology
POOL
CIF also permits the other two methods of formation processing, e.g., by means of ad-
model evaluation. Since, as mentioned, its vanced model-supported methods in measur-
math-function processing module processes ing and control. Apart from better sensors,
compiled code, it is possible to implement nearly all the developments in measuring and
smaller models. It is also possible to attach control have been based on the use of a priori
freely programmed modelling routines of arbi- knowledge of the processes in question. Since
trary complexity to the system. The access of mathematical models are scarce in production-
such programs to the POOL is regulated in the scale systems of biotechnology, it is necessary
same way as that of any person. The program to exploit alternative sources of knowledge.
must be logged in with the authorization privi- Therefore, it will be appropriate to leave the
leges necessary to acquire the required access beaten track taken in conventional process
to the data (Fig. 11). control engineering. Experienced fermenter
operators can control their fermentation proc-
esses manually if they deviate from the ex-
pected paths without using deep mathematical
4.4 New Approaches to Implement knowledge. Until now, however, their knowl-
a priori Knowledge edge could not be used directly in computers to
aid fermenter automation, since it could not be
coded in an algorithmic way. A new software
The limitations of automation systems d o technique for coping with such linguistically
not concern only the software and hardware, formulated knowledge has been developed in
but also the knowledge of the systems to be information science under the heading “expert
automated. In order to optimize an automa- systems”. These provide new programming
tion system, as much a priori knowledge as techniques that make it easier to manage prob-
possible about the biotechnological system lems involving large amounts of symbolic
must be exploited to improve its control. Since (non-numerical) data than with classical pro-
most knowledge available is not representable gramming languages.
in the form of mathematically formulated During the last years, much interest has
models, one must also tap other resources for been focused on expert systems. Many authors
a priori information. have reported sometimes too enthusiastically
about possible advantages of expert systems.
Unfortunately, much nonsense has been writ-
4.4.1 Expert Systems ten on expert systems by people who have nev-
er worked actively in this field. In too many
As already mentioned several times, one of articles insufficient distinction has been made
the main deficiencies in bioprocess automation between mere ideas and implemented soft-
is the lack of analytical models for the fermen- ware, causing high expectations awakened in
tation processes. As a result, the most advan- potential users, more promise than can be sa-
tageous features of computers will not prop- tisfied within the foreseeable future. Only very
erly come into play, i.e., their capacity for in- few concrete results have followed the papers;
Actually Developed Systems 589
on the contrary, many of the early advocates In many cases even this type of representa-
even beat a retreat, and the whole field has be- tion is not possible, even though enough exper-
come somewhat discredited. ience may be available with respect to the par-
In the following section the basic ideas of ticular biotechnological process in question. In
expert systems in automation are discussed many practical cases this experience can best
and illustrated by a working example. We re- be formulated by rules, e.g.: “Ifcertain condi-
strict the discussion to real-time expert sys- tions are met, then specific conclusions can be
tems, which can be used during automation of drawn.” Such conclusions may consist of the
bioreactors. Other interesting developments in assumption that the system will achieve a spe-
the design of processes and equipment, of im- cial state, or that the system will next behave
portance to bioprocess automation as well, in a predefined way. This type of rule is often
cannot be enlarged upon here. called “production” or “production rule”.
Apart from such state descriptions, the con-
clusions may also refer to actions to be per-
4.4.1.1 General Aspects formed by the computer; should that be the
case, they are applied, provided the conditions
As mentioned frequently in this chapter, are met. Such rules are usually imitations of
process automation requires a great deal of reactions of human operators to special proc-
knowledge of the biotechnological process, es- ess states. Thus, one often merely attempts an
pecially of biochemical kinetics, bioreactors, impersonation of an experienced operator in-
and other branches of bioprocess engineering. stead of devising a linguistic analog to a math-
This is necessary not only to predetermine the ematical model of the process.
process responses to possible actuator changes
during control, but also to determine the state
actually encountered by the real process at a 4.4.1.2 Software Base
given time.
In the natural or engineering sciences, a The essential difference between software
working mathematical model of the process dealing with heuristics and normal algorithmic
constitutes the best way to represent such programs is that in the first case one processes
knowledge. One is interested primarily in exact written words or, more generally, character
mathematical models that can be solved and strings rather than numbers. Conventional
used to describe the time development of the computer languages used so far in the natural
key parameters at given start and boundary sciences and engineering, FORTRAN, PAS-
conditions. Otherwise it would be too compli- CAL, C, and others, are not optimized for this
cated to quantitatively describe the process type of representation. Hence, special lan-
state and its dynamic development. Unfortu- guages have been constructed to deal with
nately, this case is normal in biotechnology, strings, e.g., LISP (from “list processing”), the
where process states are described more quali- oldest one, or PROLOG.
tatively, usually in words rather than in mathe- Early expert systems were programmed
matical formulas. exclusively in these languages. Today, only a
Previously, descriptions based on linguistic few are based on such elementary languages.
representations could not be used with com- Most are now based on higher-level languages,
puters in a straightforward manner, and peo- which are themselves grounded on an elemen-
ple tried to translate their knowledge as far as tary language. One example is the language
possible into formal mathematical relation- OPS5, a LISP-based development of the Car-
ships. The most familiar method in chemical negie-Mellon University (e.g. BROWNSTONet
and biochemical engineering is the use of so- al., 1985). Most applications of expert systems
called correlations, which suffer from not be- use ready-to-use expert system shells, i.e.,
ing based on crisp dynamic models, but de- complete user programs, into which are en-
scribe experimental experiences by artificial tered knowledge in the form of data and rules,
formulas that can easily be exploited numeri- which are then transferred via special user in-
cally. terfaces.
590 17 Automation in Biotechnology
Not only are the common programming lan- most expert systems developed in information
guages not optimal for list processing, but uni- science were designed as consulting systems,
versal computers themselves are not well- this question was not of primary importance
suited to string processing. The by far most during the development of most existing soft-
frequent process (and thus really time-limiting ware tools.
procedure in list processing) is the comparison One practical way to obtain higher proc-
of strings for equivalence in so-called match- essing speeds is to reduce as far as possible the
ing routines. To enhance the processing speed number of items in the working memory in or-
of these performance-limiting routines, it was der to reduce the number of matching opera-
decided to relegate them to special computers tions. Another possibility, based on the prac-
constructed solely for this purpose. Such spe- tice in process management systems, is to
cial LISP-machines have been built, e.g., by structure the whole set of rules constituting the
Symbolics. Today, there are also special LISP- expert system in order to reduce the number of
processors on the market that can be used as rules that must be considered during every cy-
coprocessors in universal computers. This cle of the system. This can be done by context-
might represent the more economical path for sensitive activating or deactivating of parts of
the future, since it reduces problems of inter- the rule volume according to different process
facing with machines coupled to the proc- phases.
esses. Concerning real-time expert systems in
All rules and facts, i.e., information on the process automation, an essential point is that
actual state of the system being treated, are heuristic knowledge is by no means the only
usually stored within a knowledge base. Its source of knowledge that must be taken into
two parts are called the “rule memory” and the account to control a process. Many subunits of
“working memory”. Besides this knowledge a process can be more adequately described by
base, there must also be an executive program. mathematical models. T o combine both as-
This is usually an interpreter for the rules, oft- pects, hybrid software is necessary (Fig. 12).
en called an “inference engine”, which applies Thus, expert systems should invoke algorith-
one rule after the other to the facts sampled in mic and heuristic software modules. The way
the working memory, seeking the conditions in which these parts communicate with each
that are to be met (pattern matching). other to exchange information provides a cri-
A remarkable feature of rule-based software terion with which to judge their performance.
is that the order of rules within the rule memo- In expert systems built as consulting sys-
ry does not have a strong influence on the tems, the decision-making process is usually
processing sequence. This is of immediate made transparent to the user by some kind of
practical effect on the maintenance of such explanatory component, which depicts the line
programs as compared to classical control pro- of argumentation that leads to a particular
grams, where the order of statements is ex- special decision. Such a feature, however, is
tremely critical. Such flexibility in the order of much too time-consuming to be incorporated
the rules allows for simple extensions of the into a real-time expert system. The basic idea
software by simply adding additional rules or of informing the user about the system’s ac-
discarding others which are no longer neces- tions, however, is of such importance to the
sary. In this way, maintenance of the software acceptance of expert systems that one must
is significantly simplified. Both are advanta- look for alternative arrangements. One possi-
geous in large software systems as compared to bility is to continuously inform the user by
conventionally programmed automation sys- means of a properly arranged graphical user
tems. However, a cyclic search through all the interface onto which the process state and its
rules for matching the required conditions is possible developments are displayed. This con-
time-consuming. In real-time applications, cept is shown in the examples to be discussed.
which are indispensible in automation, proc- Different routes in the development of ex-
essing time is a primary criterion. Thus, a cen- pert systems are of interest with respect to au-
tral question becomes how to minimize the re- tomation in biotechnology. The first is the ap-
sponse times of real-time expert systems. Since plication to control. Under the heading “ex-
Actually Developed Systems 591
Model
1 use the methodology of expert systems for this
central task if there is no adequate mathemati-
Decisions Calculations cal model available for a state identification
(ISERMANN,1988). State identification is a
prerequisite for the question of whether the
system behaves in a predefined manner or not.
EIIP r o c e s s
technology are designed for supervisory con- closed loop with biotechnological processes,
trol of the process. In most cases a convention- SUPERVISOR (LUBBERT and HITZMANN,
al process control system was employed to 1987; HITZMANN, 1988; LOBBERT et al.,
serve as a connection with the fermentation 1989), a system for supervisory control of a
process. cultivation of genetically modified Escherichia
KARIMand HALME(1988) are developing a coli bacteria to produce fusion proteins, is dis-
real-time expert system for monitoring and cussed in more detail. SUPERVISOR was
analysis, diagnosis of faults and malfunctions, built to transfer operator activities stepwise
and on-line optimization of fermentation proc- into the computer controlling the cultivation
esses. They have applied their results to a process.
batch fermentation in which the enzyme a- Since it does not make sense to again pro-
amylase was produced with Bacillus subtilis in gram into an expert system the data acquisi-
a Braun Biostat fermenter. tion and basic control routines, which are al-
The expert system was written in the special ready well-developed in conventional process
language OPS83, which supports the construc- control systems, there is a need to use an ap-
tion of rule-based expert systems on a PDPl1- propriate underlying automation system. The
computer running under the operating system only question of interest, for all practical pur-
RSXllM. Connections to the fermenter were poses, is how to obtain rapid access to the
made via RINTEKNO’Sprocess data manage- process measuring data. In SUPERVISOR,
ment system MFCS. A closed-loop control the basic computer control of the cultivation
with the expert system within the loop has not process was accomplished separately by the
yet been accomplished. process management system CASFA (FRUH et
COONEYet al. (1988) developed an expert al., 1986).
system which they called IFCONS (intelligent SUPERVISOR runs on a VAX computer
fermentation control system). It is designed to operated with the VMS operating system, and
act as a reliable and robust supervisor to proc- it is written in the knowledge-representation
ess controllers within a biotechnological proc- language OPSS (DEC, 1988). CASFA runs on
ess in order to surmount the limitations of a P D P l l process computer under RSXl 1M + ,
conventional control. Especially stressed are and is coded primarily in FORTRAN77. The
the dependency of the control algorithms on two computers are interconnected by a fast lo-
the quality of information from sensors and cal area network using Ethernet technology.
the inflexibility of most mathematical models DECnet, the network software working on
to alterations in metabolic pathways. They ap- both computers, was utilized to connect the
plied heuristics in the form of rules. The basic two software systems. It supplies a fast inter-
control strategy was taken from early papers task communication facility.
of WANG et al. (1977, 1979) and material bal- OPS5 proved well-suited to realizing the in-
ances from COONEYet al. (1977). terplay of heuristic and conventional algorith-
COONEYet al. (1988) took the control of a mic software modules. In the OPS5-versions
baker’s yeast cultivation as a concrete example for VAX computers used in SUPERVISOR,
for which they developed their software. Their this was achieved by the usual subroutine tech-
conventional laboratory bioreactor was con- niques. Subroutines written in any language
trolled by means of a direct digital controller for which a compiler is available for the VAX
running on a PDP11/23 process computer at- can be linked to OPSS programs and called by
tached to a special LISP-machine (Symbolics simple subroutine calls within the conclusion
3640) used for the expert system. IFCONS was part of any rule. The parameter transfer is ac-
written in the shell ART (automated reasoning complished accordingly.
tool), developed by Inference Corporation, A typical rule written in this special lan-
which also supplied elegant user interfaces. guage, OPS5, is shown in Fig. 13. This exam-
Access to the data was achieved via the ple not only shows the usual method of pro-
PDP11. A feedback to the process occurred gramming rules in SUPERVISOR, but it also
via the human operator of the system. As an demonstrates one of the general design aspects
example of an expert system operating in a of this expert system: by relatively simple ob-
Actually Developed Systems 593
Example for Coding a Rule cially in processes with higher time constants,
in order to become aware of unwanted process
IF states prior to the time at which they become
Off-gas C02-measuring vzlues critical. All extrapolations are analyzed and
become smaller, formulated as warnings, which are written in
short texts into the graphic representation if
THEN they indicate severe deviations from the nor-
mal process state. It is obvious that this meth-
Suspect a limitation od of analysis utilizes many algorithmic fea-
tures of the software.
Such user interfaces will be given much
more attention in the future of process man-
in S U P E R V I S O R agement generally. They must be optimized
for the needs and requirements of the human
(p rule-CO2-warning limitation operator (which will still be necessary in the
(evaluation future) and not for the characteristics of a giv-
"measgrobe CO2offgas en computer. Such a sophisticated user inter-
"determination slope face must also be based on heuristic knowl-
lralue c 0.0) edge, on the ways in which a human operator
-- > can survey a given process state as quickly as
(MAKE limitation suspected possible. Similar developments are also appar-
ent in related disciplines. BAR and ZEITZ
CO2offgas)) (1989) demonstrated one concept of a knowl-
Fig. 13. Typical rule written in OPS5. The example edge-based user interface, originally created
shows the usual way of programming rules in SU- for a dynamic simulator for chemical proc-
PERVISOR and also demonstrates one of the gener- esses and plants, which meets many of the re-
al design aspects of this expert system, as explained quirements of process automation as well.
in the text.
system, however, they must actively work with ists are the papers from chemical engineering
the cultivation process over a long period in (e.g., YAMASHITAet al., 1988a, b).
order to learn the difficulties of the special
process. Another measure would be to develop
user interfaces to expert systems that would al- 4.4.1.6 Further Applications
low biotechnologists to implement their rules of Knowledge-Based Systems
directly into a given system, i.e., to develop
special shells optimized for biotechnologists. in Automation
Since the human operator will remain the
decisive element in fermentation control, it is
4.4.1.5 Completeness Problem necessary to support him with as much infor-
mation on the process state as possible, but
A point often criticized in expert system de- not more than that. Thus, compared with the
velopments is that one does not have a real data displayed on today’s video screens, the in-
chance to sample all necessary knowledge formation must be reduced to what is neces-
within a given domain, e.g., the automation of sary to decide how to proceed in the actual sit-
bioreactors. Omniscience is not achievable. uation. This information is equivalent to a
But there are two ways to cope with the prob- well-defined process state representation,
lem. which is only possible with a model of the sys-
The first is to reduce drastically the goals. tem. By the same arguments stated above for
In the accompanying example of SUPERVI- model-supported control, such information
SOR, this has been achieved simply by restrict- can sometimes be given only in heuristic terms,
ing the support given to the operator regarding often only in an incomplete and uncertain
process supervision. This was accomplished by way. Thus, it is obvious that heuristic knowl-
dividing the task into small slices of operator edge-based methods are of advantage in sup-
activities, e.g., the detection of very special porting the man-process interface if a compre-
process states such as the detection of diauxic hensive mathematical model is not available.
growth of the organism. Thus, step by step, Another field in which knowledge-based
more and more activities can be added to the systems become important is that of “measur-
system. Consequently, the system is constantly ing techniques”. Here, sensor fusion is gaining
in a working state, and it can be used advanta- increasing importance in combining informa-
geously by the operator. tion obtained through various types of sen-
The second possibility for coping with the sors, which can be realized most appropriately
problem is to use methods developed to handle using unified knowledge-based methods
incomplete, inexact, or even contradictory in- through multilevel representation combining
formation. In such cases, rule-based systems quantitative and symbolic attributes (PAu,
may encounter conflicts, which must be re- 1989). The primary goal is to extract better
solved as well as possible by a robust method. features and to achieve sensor diversity. This is
Such a technique is the fuzzy theory or fuzzy an approach similar to model-supported meas-
reasoning. Fuzzy reasoning is the process by urement on a heuristic base, which is necessary
which a possibly imprecise conclusion is de- where neither a sensor for certain key process
duced from a collection of imprecise premises. variables nor mathematical models which re-
The first model of fuzzy reasoning was devel- late them to easily measurable quantities are
oped by ZADEH beginning in 1965 (ZADEH, available.
1965, 1984). This technique has been at-
tempted several times in biotechnology, espe-
cially in measurement (POSTLETHWAITE, 4.4.1.7 Conclusion
1989), process control (TONGet al., 1980; NA-
KAMURA et al., 1985; CZOGALAand RAWLIK, The idea of using expert systems in automa-
1989), and process simulation (e.g., TURUNEN tion biotechnology is very attractive to many
et al., 1985). Also important for biotechnolog- people. Expert systems are under development
Actually Developed Systems 595
at many places. However, their broad applica- ling complex processes for which the existing
tion will take considerable time. models are insufficient (TONG, 1984). It is
Many people in industry would be satisfied based on fuzzy sets and fuzzy reasoning.
if only the operator behavior could be assumed The general idea of fuzzy reasoning has
by a computer. The goal is not only to spare been developed by ZADEH (1973). It can be
manpower, but also to ensure well-defined understood as an extension of set theory, in
reactions to specific process faults. In practice which the essential idea is that membership of
it has been recognized that crews in different an element in a set is not a strict binary yes-
shifts react markedly differently to identical or-no relationship as in conventional set the-
deviations of the process from a predefined ory, but is defined by a membership function
path. This leads to unwanted variations in that takes continuous values within the inter-
product quality. Such fluctuations could be re- val [0, 11. These membership functions express
duced or eliminated if a unified behavior could to what degree the element is assumed to be a
be implemented and incorporated into a member of the fuzzy set. Since membership
knowledge-based computer system. Another functions are more or less subjectively defined
point raised by industrial people is that such by experience, this is another way of incorpo-
an implementation would conserve at least rating heuristic knowledge into an automation
some knowledge of older, highly experienced system.
workers against loss after their retirement. The development of fuzzy rule-based con-
As the term “expert system” indicates, the trollers of interest to automation systems dates
system tries to reason heuristically, like people back to ASSILIANand MAMDANI(1974), who
in situations in which no exact methods are tried to establish a non-mathematical tech-
available, or when the evaluation of an ade- nique to control a pilot-scale steam engine.
quate solution is not possible during the time The difference between fuzzy rule-based
period within which a decision must be systems and the normal expert system ap-
reached. Biotechnology is, therefore, an ideal proach is that the conditions are formulated by
application field for expert systems, since it fuzzy expression or relationships, such as:
represents a very complex field in which only a
few aspects can be covered by exact mathemat- if the deviation of the measuring variable
ical models. from the setpoint is of positive medium
With expert systems in automation of bio- size:
reactors, the way ahead will be analogous to then change the valve setting by a small
the progress already apparent in the underly- amount in the positive direction.
ing digital process control systems. Distributed
systems will arise, and priority structures will Here the fuzzy variables “measuring varia-
be incorporated into the expert systems. Final- ble” and “valve setting” take fuzzy values,
ly both universal control systems and real-time e.g., “positive small”, which represent their
expert systems will merge. linguistic meaning. The relationships between
Many critical comments on expert systems the variables are calculated by means of fuzzy
from people living in the world of exact mod- reasoning. In a concrete application the rela-
els have missed the mark. It is well accepted tions are then made discrete and put into a
that exact modelling is better in nearly any large matrix from which they are retrieved dur-
case; in most cases, however, they are not im- ing the operation of the controller. This table
mediately practicable. Thus, to solve problems look-up procedure guarantees fast processing
now, one must take “the bird in the hand rath- of the fuzzy controllers.
er than two in the bush”. There have been several attempts to imple-
ment such fuzzy controllers in biotechnology.
TONGet al. (1980) used a fuzzy controller in
4.4.2 Fuzzy Methods an activated sludge wastewater treatment proc-
ess. NAKAMURA et al. (1985) applied fuzzy
In the fuzzy approach, one attempts to mod- control to a glutamic acid fermentation. Tu-
el the strategies of a process operator in hand- RUNEN et al. (1985) applied fuzzy techniques
596 17 Automation in Biotechnology
This requires the individual modules to be self- SNSS acts as the only public interface for the
contained and matchable. Furthermore, uni- core, which allows for communication with
versal usabiltiy requires that objects be equip- other objects around the CIF-core. This en-
ped with standard communication interfaces capsulation technique acts as a protective
by which they can easily be connected with shield for the internal data against wallowed
others. access, since the SNSS-routines check every re-
One difference between object-oriented and quest for POOL data in terms of access au-
conventional procedural programming arises thorization. The SNSS interface is network-
in the handling of data. In an object-oriented transparent; this object can be used by other
environment an object’s internal data struc- object-oriented programs running on any com-
tures and current values are accessible only to puter within the local area network.
the methods within the object. These methods This core object holds all data and provides
are activated through messages passed from the basic functionality of the automation sys-
other objects which specify the methods to be tem. It is used extensively by all other objects.
activated and the data or parameters required. It decouples programs for more sophisticated
Whole programs are then constituted by a process management tasks, e.g., advanced
hierarchy of objects and their interrelation- model-supported supervisory process control
ships (CHEN, 1977). algorithms, as well as knowledge-based proc-
Traditional programs rely on function- ess control software modules from process vis-
oriented approaches that organize the process ualization and basic process control.
through a hierarchy of functions or data-
oriented approaches, which in turn emphasize
a hierarchy of data structures. Object-oriented 4.5.2 CASE Tools
methods, on the other hand, impose the natu-
ral modularization of the process in question A widely heard criticism regarding software
through an emphasis on objects with a one-to- development is that it is not done systematical-
one correspondence with the actual compo- ly enough in terms of engineering and organi-
nents in the process. This provides for easy zation. As compared to hardware develop-
mapping between the software and the proc- ment, where all stops are pulled for planning
ess. and production control, the way in which pro-
This software technique, thus, points in the grams are developed has been compared to ar-
same direction as the modularization of hard- tistic work. It has been argued that this signifi-
ware components within distributed control cantly retards development.
systems, which has demonstrated high reliabil- Consequently, to tighten development and
ity with respect to errors because of its high permit the planning of larger software pro-
transparency. Object-oriented programming jects, engineering methods have been proposed
thus exploits the advantages of distributed di- and tested in many different software projects.
gital control systems. It is obvious that one can use the computer it-
By definition, a software object is a set of self to aid software development by assigning
data together with the software components to it all tasks in software development that can
required to manage the set and provide access be formalized. Software tools designed to sup-
to it. This is essentially the way in which the port systematic program development have
central core of CIF works, including the cen- been developed by all of the larger computer
tral database POOL, its managing program manufacturers and software houses under the
EXECUTIVE, and the standard software in- name CASE tools (computer aided software
terface SNSS (see Fig. 3). engineering).
This CIF-core is the fundamental software Software engineering can be divided into
building block employed to perform the basic three main parts. First there must be a model
automation activity at the biotechnological of how to proceed during the development of a
plant. Its data and the basic data manipulation software project. It must describe the sequence
routines are completely separated from all oth- of steps to be performed. The second part con-
er modules and, thus constitute an object. The tains the description of the methods to be ap-
598 17 Automation in Biotechnology
plied within the individual steps. Finally one To utilize the concept of distributed proc-
needs the software tools necessary to perform essing of system activities comprehensively, it
individual tasks. Software tools are now evolv- is necessary that generally-accepted open
ing into complete tools for problem solving in- standards be obeyed. These are, first of all,
stead of tools that can be used only to con- necessary to assemble a distributed system of
struct a solving mechanism. the components from different vendors that
CIF has been developed using CASE tool- are regarded as optimal for an applicant’s
box VAXset supplied by DEC to support pro- needs. This increases the flexibility of adapting
gram development in all its phases, starting a real system to changing conditions from the
with system design, moving to advanced lan- viewpoint of the goals of production system as
guage-sensitive editors, and culminating in well as those of the available system compo-
software version management tools. nents.
The realization of this concept is not so
much limited by the hardware components as
by the availability of appropriate software.
The software is responsible for the functionali-
5 Conclusions and ty of an automation system and is thus more
important. Therefore, it is necessary to choose
Recommendations software that meets the requirements and then
for Future Developments the hardware on which the software can be
run.
This argument becomes even more impor-
Distributed digital automation systems dom- tant if one realizes that the software develop-
inate the automation of chemical and bio- ment time constant is much longer than the in-
chemical production systems. They are com- novative periods in hardware development.
posed of microprocessor-based computers Today, software cost is roughly twice that of
linked together in local area networks. Systems the computer on which it runs. This means
will be expanded to networks within larger that a chosen software package often must sur-
company-wide networks. Even in a chemical vive several hardware generations. Thus, the
or biotechnological production unit, several customer is advised to look for hardware
levels of networks will soon appear, e.g., net- manufacturers with a strong family concept of
works of field instruments and networks of la- computers, offering confidence that the chos-
boratory equipment in a network for process en software will also run for years to come on
management. machines adapted to a more transient state of
Their main advantage is that of a building the art.
block system. Powerful systems clearly organ- The requirements for automation software
ized and adapted to the special needs within an are steadily and rapidly growing. More and
application can be flexibly built (Fig. 14). more functions are becoming available to
Maintenance is supported by working with a process management systems, and additional
limited set of small standard components. The economical and ecological boundary condi-
computing power of such distributed systems tions must be met and integrated into process
is naturally increased by true parallel comput- automation. These all require software exten-
ing of the individual nodes. sions. Hence, the systems become larger and
However, the structuring of distributed au- more difficult to control. Measures are re-
tomation systems cannot be arbitrarily. The quired to keep larger software systems trans-
structure must resemble the logical structure of parent. One promising approach is software
the process to be automated. Individual com- structuring by means of object-orientated pro-
ponents must be chosen to be as internally self- gramming. This point of view is important not
contained as possible to avoid unnessessary only for software development, but also to the
communication and synchronization efforts user, since it significantly influences the soft-
that would overcompensate for the advantages ware’s maintenance and extendibility proper-
of distributed systems. ties.
Conclusions and Recommendations for Future Developments 599
Fig. 14. Architecture of an ideal process automation system hardware. Several differ-
ent components are arranged in a local area network utilizing Ethernet technology.
FEP stands for front-end processors, which serve as gates to field busses to which the
measuring (M), open-loop control (S), and closed-loop control (R) devices are at-
tached. MPC represents the man-process interface workstations, PMS the process-
management computers, XPS the expert system hardware, and GWC the gateway
processors that provide connections to the data highways of the higher-level compa-
ny-wide information processing systems.
It has become increasingly evident in recent “expert systems”. Both parts will be used in a
years that the biotechnologist at the plant site complementary way, with a clear preference
is not replaceable by automation systems in the for mathematically exact models, wherever
foreseeable future. He remains the decisive they exist and are solvable within the time lim-
component in the bioprocess. Software devel- its set by the requirements of the real-time
opment must therefore be focused on support- environment. Knowledge-based systems, for
ing him in his decisions as much as possible. example, can build a platform for dealing with
This requires putting more of the software and uncertain and incomplete information, one of
hardware money into user interfaces. It is nec- the properties of human beings that in long-
essary to provide the biotechnologist with term thinking deserves to be incorporated into
tools for incorporating his ideas on process automation systems.
control directly into the automation software. One point usually not emphasized in discus-
All types of a priori information must be sions of automation is the requirement that
used to improve measurement and control: sufficient information on the actual state of
knowledge represented by mathematical mod- the process be available. This, above all, is a
els as well as heuristic knowledge that can only matter of reliable sensors for the key process
be formulated linguistically in the form of variables. A remaining general problem in bio-
rules or other heuristic means. Conventional technology is the scarcity of powerful sensors
process control will merge with the knowledge- for the basic quantities.
based automation techniques that we now call
600 17 Automation in Biotechnology
component balancing and culture fluorescence, the analysis of complex systems and decision
PhD Thesis, University of Pennsylvania, Phila- processes, IEEE Trans. SMC-3, 28-44.
delphia. ZADEH,L. A. (1984), Making computers think like
ZABRISKIE, D. W . , HUMPHREY, A. E. (1978), Real- people, IEEE-Spectrum, Aug., 26-32.
time estimation of aerobic batch fermentation - ZIEGLER,J. G., NICHOLS,N. B. (1943), Optimum
biomass concentration by component balancing, settings for automatic controllers, Trans. ASME
AIChE J. 24, 138-146. 65, 433-444.
ZADEH, L. A.(1965), Fuzzy sets, Inf. Control 8, Zoz, B. (1990), Einsatz von Industrie-PCs als lokale
338. Kopfstation flexibler verfahrenstechnischer An-
ZADEH,L. A. (1973), Outline of a new approach to lagen, Verfahrenstechnik 24, 52-58.
18 Modelling, Design, and Control
of Downstream Processing
SUTEAKISHIOYA
KEN-ICHI SUGA
Osaka, Japan
List of Symbols
cm2 cross-sectional area of column
- fouling constant
g cm-3 concentration of particle in the suspension
mg cm-3 concentration in bottom phase
g cm-3 bulk concentration
g cm-3 maximum concentration
mg cm-3 concentration in mobile phase
mg cm-3 equilibrium concentration in mobile phase
mg cm - 3 concentration in stationary phase
mg cm-3 concentration in top phase
cm fluid channel height above the membrane
cm particle diameter
8 small amount of sample
cm2 s - ' diffusivity
cm2 s - ' axial dispersion coefficient
cm2 s - ' diffusion coefficient into the particle
fraction of population in rth cavity
transfer rate of solute
gravitational constant
distribution coefficient
ratio of effective space = (1 - E ) / E
cm height equivalent to a theoretical plate
cm3 cm-2 s - ' flux
cm3 cm-2 s - ' flux at a time t
cm3 cm-2 s - ' flux at t = 1 min
mass transfer coefficient
partition coefficient
overall mass transfer coefficient
length of cylindrical centrifuge
length of the channel of cross-flow filtration
number of space between disks in the stack
number of fractions
number of ideal stages
fraction of materials in top phase
concentration of the product in ith purification stage
amount of the product in ith purification stage
gauge inlet pressure
gauge outlet pressure
average transmembrane pressure drop
amount of absorption per unit weight
throughput of the continuous centrifuge
absorption rate per unit weight of packed bed
radial distance from the center of the centrifuge
radius of liquid surface in the bowl of the centrifuge
radius of the bowl of the centrifuge
peak position of objective material
ratio of the solute in the mobile phase to the total amount of the solute
relative centrifugal force
Reynolds number (dp, v/,u,)
R, cm-' resistance of the retained components
List of Symbols 605
Greek symbols
ai
- purification index
P - parameter of absorption equilibrium
& - void ratio
0 - dimensionless time
OR, i S retention time of the peak i (i= 1,2)
Pn
- n th moment of the elution curve (n= 1,2, .. ., n)
PS g cm-'s-' viscosity of the fluid
Pr,s g cm-'s-' viscosity of the fluid at T "C
P20, u
-
g c y -1 s -' viscosity of the water at 20°C
V g - cm3 partial specific volume
Pb g~ r n - ~ density of packed bed
PP g cmP3 density of the particle
Ps g cmP3 density of the fluid
PT,s g cm-3 density of the fluid at T "C
P20, w g cm-3 density of the water at 20°C
c7 - standard deviation of the elution curve with respect to the peak r,
r S retention time
4 rad the conical half angle
0 rad s - ' rate of rotation
Biotechnology Second, Completely Revised Edition
Edited by H.-J. Rehm and G.Reed in cooperation with
A. Puhler and P. Stadler
copyright@WILEY-VCH Verlag GmbH, D-69469 Weinheim (Federal Republic of Germany). 2001
fugation and filtration with filter presses or ro- Tab. 2. Comparison of Hollow-Fiber Filtration and
tary vacuum filters are the processes ordinarily Centrifugation for Bacteria Harvesting (Annual)
used. Recently, the application of cross-flow
Hollow-Fiber Centri-
membrane filtration for the harvesting of cells Filtration fugation
has been investigated. In this filtration, the
suspension is recirculated across the filter sur- Depreciation $ 4500 $ 15000
face at high velocity to remove the retained Maintenance 1800 6 000
particles. This type of filter, therefore, is dif- Labor 7 500 7 500
ferent from the conventional type, in which Power 1500 6 200
the filtration rate decreases with the build-up Water, chemicals 1400 1200
of retained cells on the filter surface. In this Membranes 10500 -
system the filtration rate depends on (1) the Total $ 2 1 200 $ 35900
transmembrane pressure, (2) the bulk cell con- Capital cost $45000 $150000
centration, and (3) the mass transfer coeffi-
cient. PATEL et al. (1987) reported the feasibil- Assumptions:
ity of harvesting yeast cells using synthetic System output 5 000 L/h (water removal rate)
Operation 250 d/a, 16 h/d; plus 2 h/d clean up
membranes. They concluded that (1) filtration Depreciation 10 years S/L
rates decreased exponentially with time, (2) Maintenance 4% of capital cost annually
fouling rates became lower at lower cell con- Labor at $15.00/h
centration, lower transmembrane pressures, Power at $ O.OS/kWh
and higher velocity, and (3) media components Membrane life of 1 year, $ 157/m2
played a relatively greater role in the fouling Membrane av. flux rate 75 L/m2 h
phenomenon at low cell concentrations. Mean- Broth Escherichia coli simple medium, 20 g/L feed
while, porous cylindrical sintered stainless steel 180 g/L product
tubes of 2 pm and 3 pm nominal pore size
have been applied to the cross-flow separation
of yeast cells (KAVANAGH and BROWN,1987).
In this case average filtration rates as high as
1.25 m 3 m - 2 h - ' were recorded for low cell Tab. 3. Comparison of Hollow-Fiber Filtration and
concentrations (3 g L - I ) . As cell concentration Centrifugation for Yeast Harvesting (Annual)
was increased from 3 g L-' to 25 g L there
was a steady but gentle decline in average fil- Hollow-Fiber Centri-
tration rate. The flux of the tubular microfilter Filtration fugation
was found to be more than 100 L m-* h-' at Depreciation $ 7500 $ 5000
yeast cell concentrations of 100 g L - I , while Maintenance 3 000 2 000
the hollow-fibers had a flux of 40 L m-' h-' Labor 7 500 7 500
at 250 g L-' (PATELet al., 1987). Power 2 400 2 900
The coarse sintered stainless steel cylindrical Water, chemicals 2 300 1500
filter element (nominal pore size of 75 pm) as Membranes 17400 -
a sterilizable cross-flow filtration unit was Total $40 100 $18900
used for the separation of filtrate from a fun- Capital cost $75000 $50000
gal Trichoderrna reesei suspension in a sterile
manner (BROWNand SALAM,1984). This pa- Assumptions:
per also reported an average filtration rate of System output 10000 L/h (water removal rate)
1.0 m3 m P 2h-' and the fact that cell concen- Operation 250 d/a, 16 h/d; plus 2 h/d clean up
tration increased from 16.4 to 47 kgm-3. Fur- Depreciation 10 years S/L
Maintenance 4% of capital cost annually
ther, in the harvesting of Aspergillus niger Labor at $ 15.00/h
from a fermentation broth, the flux of the fil- Power at $ O.OS/kWh
trate was 80 L m w 2h - ' at a mycelial dry mass Membrane life of 1 year, $ 157/m2
concentration of 57 g L - I , a velocity of 4 Membrane av. flux rate 90 L/m2 h
m S K Iand, an initial transmembrane pressure Broth Saccharomyces cerevisiae, simple medium, 40
of 10 kPa using a tubular cross-flow microfil- g/L feed 180 g/L product
Principle of Unit Operation in Downstream Processing 609
ter (pore size of 0.2 pm) (SIMSand CHERYAN, 2.2 Separation of Bioproduct
1986). From the economic viewpoint a com-
parison of hollow-fiber filtration (0.1 ym mi- 2.2.1 Centrifugation and Membrane
croporous-type) and centrifugation (a contin-
uous disc-type centrifuge) was made for both Filtration
bacteria and yeast harvesting (TUTUNJIAN,
1984) as shown in Tabs. 2 and 3. It is evident After cell harvesting and disruption, the re-
that the hollow-fiber system is preferable to covery of intracellular products from cell de-
centrifugation when dealing with smaller cells, bris is carried out by centrifugation and/or fil-
in which case centrifuge outputs drop signifi- tration. Cell debris is made up of particles with
cantly. a size of the order of 0.1 pm, while protein
molecules are two orders of magnitude small-
er. Cell debris removal is the most difficult
2.1.2 Cell Breakage and expensive solid-liquid separation. Various
techniques are available, and the choice de-
When the product is accumulated inside pends on cell properties and on scale of opera-
cells, the first step of separation is the break- tion. This has most commonly been performed
down of the cells, and the desired product is by centrifugation. Semi-continuous type cen-
usually extracted into an aqueous or other trifuges are generally used in large-scale ex-
phase. Methods of cell breakage are: destruc- traction. The insoluble cell debris is pumped
tion by osmotic pressure, repetitive freezing continuously into the tubular-bowl centrifuge,
and thawing, utilization of enzymes, self-di- where it is quickly accelerated to bowl speed.
gestion or autolysis, and mechanical disrup- Solids are deposited, while the clarified liquid
tion by Dyno-Mill, sonication, French press, is discharged continuously over a dam at the
and other methods. Also, surface-active agents top of the centrifuge bowl. Since there is no
such as polyoxyethylen(n) octylphenyl ether automatic solids removal system, only small
-
(n = 9 40) are frequently effective in breaking concentrations of solids can be handled. Disc
down cells and viruses. centrifuges have provided an excellent means
of extract clarification. These types of centri-
fuges, however, may rotate at slow speeds, de-
2.1.3 Precipitation Partitioning veloping centrifugal forces only up to 8000 x g.
They have a capacity of up to 20 kg of sedi-
In order to obtain concentrated samples and ment. For large-scale separation, continuous
also to reduce sample volume, precipitation by flow centrifuges are expensive in capital and
denaturation of protein is utilized for separa- running costs. When the density difference be-
tion from bulk solutions. The most popular tween the particles and medium is small, cen-
method is salt dialysis, which depends on salt trifugal throughputs may need to be reduced in
concentration and surface properties of target order to maintain high separation efficiency.
proteins. Isoelectric precipitation is also used, Recently, tangential flow filtration has been
because precipitation occurs readily under the investigated as an alternative to centrifugation
condition of low ionic strength at isoelectric for the separation of soluble intracellular
pH. Precipitation by organic solvents is also products from cell debris. The performance of
used for protein partitioning. tangential flow filtration in the removal of cell
debris is affected by a large number of factors,
such as the organisms utilized, the molecular
weight of the desired enzyme, the method of
cell breakdown, the physicochemical proper-
ties of the membranes and their pore sizes,
pressure, flow hydrodynamics, temperature
and cleaning methods. These factors may be
investigated by measuring the effect of flux
rate, process time, and turbidity of the filtrate
610 18 Modelling, Design, and Control of Downstream Processing
on the yield of enzyme as the desired biopro- (300 min), and mercaptoethanol following ly-
duct. The molecular weight of the enzyme sozyme breakage (180 min) for P. fluorescens
could be an important parameter during mi- homogenates, the specific activity of filtrates
crofiltration. QUIRKand WOODROW (1984) re- decreased from 0.97 to 0.50 to 0.43, respec-
ported that enzyme yields were in inverse order tively (QUIRK and WOODROW,1984). This
of their molecular weights using asparaginase suggests that the characteristics of the cell de-
(M.W. 132000), carboxypeptidase (M.W. bris may affect the separation performance
83 500), and arylamidase (M.W. 52000). Simi- with respect to change of flux rate and specific
lar results were obtained (DATAR,1985) for p- activity. It also indicates the formation of a
galactosidase (M.W. 540000), carboxypepti- secondary filtration layer by accumulation of
dase Y (M.W. 65000), and human growth hor- either insoluble matter or soluble protein.
mone (M.W. 21 000). The Domnick-Hunter LE and ATKINSON (1985) reported the effect
membranes with an asymmetric structure (pore of ionic strength on separation efficiency using
size of 0.45 pm on the tight side and approxi- a lysate of P. fluorescens cells. The effect of
mately 1.5-2.0 pm on the open side) gave the ionic strength on enzyme transmission was
same recovery of arylamidase as millipore said to be independent of the method of cell
poly(viny1idene fluoride) membranes (0.5 pm). treatment. However, the enzyme transmission
Increasing the pore size of an Amicon mem- level increased with an increase in buffer
brane to 0.6 pm increased carboxypeptidase strength.
yield (from 52 to 78%), but resulted in de- Tab. 4 provides the data for the separation
creased quality (QUIRK and WOODROW, of formate dehydrogenase, using different
1984). Yields may be increased by using mem- methods, from the cell debris of Candida boi-
branes with larger pore sizes. However, the dinii (KRONERet al., 1984). The data show
passage of particles of larger sizes through the that energy consumption was low for cross-
membrane blocks the column for its use in co- flow filtration in comparison with tubular-type
lumn chromatography in subsequent down- centrifugation. The clarification of the filtrate
stream unit operations. Cross-flow velocity was 100% for cross-flow filtration. As men-
and pressure have a significant influence on tioned previously, this may be important for
flux rates. Flux was proportional to the feed further purification procedures, such as chro-
velocity raised to the power of 0.5 with aryl matography. The separation efficiency for sol-
acyl amidohydrolase lysates obtained by lyso- uble enzymes from cell debris by a cross-flow
zyme treatment (LE and ATKINSON,1985). filter system is presently not satisfactory in
The effect of pressure on flux of the cell lysate terms of enzyme yield.
for various feed velocities was investigated by In order to improve flux rate, hollow-fiber
LE and ATKINSON (1985). Flux increased with ultrafiltration membranes were examined for
increasing pressure up to a limiting level. Fur- the processing of a precipitate suspension of
ther, the transmission activity, i.e., the ratio of isoelectric soya protein (DEVEREUXand
enzyme activity in the filtrate to that in the HOARE,1986). These authors determined that
feed, was proportional to the feed velocity protein precipitation was a method that might
raised to the power of 0.18, and it increased be used to improve permeate flux when polari-
rapidly with increasing pressure. Physical and zation is a major limitation to membrane sepa-
chemical (or enzymatic) methods of cell break- ration.
age had little effect on enzyme yield in the sep-
aration of arylamidase from Pseudomonasfluo-
rescens (LE and ATKINSON,1985; QUIRKand 2.2.2 Extraction
WOODROW,1984) and on asparaginase from
Erwinia carotovora (QUIRKand WOODROW, Differences in the partition of solutes in a
1984). However, process time for filtration sig- liquid-liquid mixture can be utilized for ex-
nificantly varied with changes in the methods traction and separation of product. The two
of cell breakage. As process time decreased phases may be water-organic solvent, water-
with the different treatments, such as Dyno- water, or liquid-gas under supercritical condi-
Mill breakage (480 min), lysozyme breakage tions.
Principle of Unit Operation in Downstream Processing 61 1
Tab. 4. Comparison of Methods for Separation of Cell Debris from Enzymes
1- 5 P 2 0 , w
3.1 Centrifugal Separation
Centrifugation is a useful method for the where 5 is partial specific volume. The value of
separation of cells, subcellular organelles, or s20, decreases with increasing concentration,
large molecules. The basic parameters govern- C, of particles according to the following
ing the separation of particles in a centrifugal equation
field are mass, density, and frictional coeffi-
0
cients of the particles present in the suspen- s20,w
sion. Considering a spherical particle falling in s20, w = ~
1 +K,C
a gravitational field such that other particles
present do not hinder its fall, its velocity will where s;o,w and Ks are constants. Therefore, if
reach a constant value determined by a balance s;o,wand 5 are known, the terminal velocity u,
between the net force acting on the particle can be estimated by Eq. ( 5 ) using the charac-
and the frictional resistance. From Stokes' law teristics &,,, of the liquid.
the terminal velocity v, in a gravitational field The relative centrifugal force (RCF), de-
is given by fined by the ratio of u, to u, from Eqs. (3) and
(4), is commonly used to express the centrifu-
gal field as follows
.Z is a parameter of the centrifuge, which is K, and the distribution coefficient, G. For mo-
equivalent in area to a gravity settling tank lecular distributions the partition coefficient is
that is theoretically capable of doing the same defined as the concentration ratio of the top
amount of work. phase to the bottom phase. The distribution
For the disk-bowl centrifuge coefficient is the ratio of the total amount of
material in the top phase to that in the bottom
2 n m (r: - r:) o2 phase. The three parameters are related as
z=
3gtanQ
K=Ct/Cb (13)
where m is the number of spaces between
disks in the stack and
Q the conical half angle
also be obtained (TREFFRYand SHARPE, plished with the aid of a solution that is com-
1985) pletely different from the adsorbent. This is
called a selective adsorbent.
W,,, = 1.18 [r, (1 - rm/n)]’” (19)
Therefore, if the width at the half-height of an 3.3.1 Separation Characteristics
experimental CCD peak is significantly differ-
ent from W,,, as calculated from Eq. (19), the by Liquid Chromatography
assumption that the population is homogene-
ous must be questioned. Note that the peak To predict separation potential using LC, it
width divided by the peak location, Eq. (20), is necessary to know the elution volume or
decreases as the number of transfers is in- time required for attaining the peak of elution
creased according to of each component as well as the width of the
elution curve. Elution time depends on the rate
of motion of the peak as established by the
partitioning properties of the solute with re-
spect to the mobile and stationary phases. Fur-
where the curve is approximated by the normal ther, the width of the elution curve is governed
distribution with a = n P ( l -P)1’2.
Hence, the by the diffusivity of the solute into both
resolution between two peaks will increase as phases, backmixing of the liquid, and unifor-
1/2
mity of flow. These properties can be repre-
sented by the ideal plate model or mass bal-
ance, as shown later. The former uses the ideal
3.3 Chromatographic Separation plate number N to explain the observed sepa-
ration characteristics. The latter utilizes differ-
Since the driving force of the separation de- ential equations describing the mass transfer
pends on differences in the partitioning of ma- rate and axial diffusion of the solute. The con-
terials between mobile and stationary phases, cepts of adsorption and mass balance are also
various types of liquid chromatography can be applicable to affinity chromatography, but in
categorized according to the principal mecha- a slightly different way.
nisms of partitioning, as shown in Tab. 5 .
Many target bioproducts to be separated have
amphipathic properties, so the ampholyte, 3.3.2 Equilibrium Model
size, shape, electrostatic, and polar properties
and structural specificity of molecules have all V, and V, denote the total volume of the col-
been utilized for separation. umn and volume of the void, respectively, and
Elution procedures can be divided into three the mobile phase liquid flows at a space veloci-
kinds of operations, depending on the nature ty u. The rate of motion of a solute peak is
of the interaction between the stationary phase proportional to the ratio of the amount of so-
and the solution used for elution. First, a solu- lute in the mobile phase to that in the stationa-
tion with the same composition as the sample ry phase. The velocity of the peak then equals
solution is used for elution in the case of gel or the product of u and the ratio. The equili-
hydrophobic chromatography (isocratic elu- brium governing the concentration of the so-
tion). Second, when partition is complete or lute in the mobile phase C, and that in the sta-
clearly defined, the composition of the elution tionary phase C, can be described as
solution is subjected to stepwise change (step-
wise elution), in order to cause a gradual
change in the strength of partition. Third,
when the partitioning of the target component where K is the partition coefficient.
is especially complete, as in affinity chromato- Now the ratio, R,of the solute in the mobile
graphy, adsorption is continued until a break- phase to the total amount of the solute can be
through occurs, after which elution is accom- written as
Modelling of Unit Operations 6 15
R=
c mE =-
1
(22)
where 2 is the column length. The elution vol-
C m & + C s (- lE ) 1 +HK ume V, is given as
When a small amount of the sample, dS, column is approximated by a Gaussian distri-
flows continuously and at a constant rate into bution;
the column, the mass balance equation for the
solute is shown to be coBo
Cm,n = 1/2 n (1 + HK)'/N
3.3.4 Diffusion Model rived more simply using the moment method.
The moment is defined as
mass transfer between mobile and stationary P; = tCm(z, t)dt Cm(Z,t)dt (38)
0
phases as well as axial mixing along the col-
umn, and these can also be used for the analy- corresponds to the average retention time,
sis of liquid chromatography (KUBIN, 1975). and the variance around ,u; is given by
Consider the case in which the column is filled
1:E
m
with uniform particles of diameter dp and the
sample is applied from time 0 to to, following ~2 = I
0
(t-P;)'Crn(Z, t)dt G(Z, t)dt (39)
which elution begins. The mass balance equa-
tion along the column length becomes Here we can use the following relations in the
Laplace transformation
ac,
-=D,--
a2cmu--HF,
ac, (34)
at az2 az ,ul: = (- 1)" lim (d/ds)"c(z, s)/lim c ( z ,s) (40)
s-0 s-0
where C, is the solute concentration in the where s is the Laplace operator defined as
mobile phase,
m
0, the axial diffusion coefficient, and
F, the transfer rate of solute per unit c ( z ,s) = C(z, t)e-s'dt (41)
0
volume of the stationary phase.
Using Eq. (40), the average retention time
The diffusion of solute C, into a particle can from the basic equation (34) can be derived
be written as as
to
,u{=r(l+Hm+- (42)
5 = D , (a%,
F+--
2 ac,)
(35)
2
at ar r ar [ D r ( l + H f l z d:HK] + ti
,u2=2r
U2
+- 6 0 0 , -
12
(43)
where 0, is the diffusion coefficient into the
particle; that is, Then
HETP = Z,uZ/(,uU;)'
HF,=H- - 2 0 , dEu HK
--
- +- (44)
u 300, ( 1 + H f 1 2
If the liquid phase mass-transfer resistance can
be neglected, then the initial and boundary From this equation it can be seen that there is
conditions become an optimum u that minimizes HETP.
t=O z>o C,=C,=O
t>O r = d p / 2 C,=KC,
r=O ac,/ar=o (37) 3.3.5 Mass Transfer in Affinity
O<tsto z=o cm=co Adsorption
t>to z=o c,=o
By utilizing these basic partial differential In this case the third term of Eq. (34) can be
equations, the correct elution curve can be de- rewritten as
termined, although doing so entails many
time-consuming calculations. However, the av- Pb -
HF,=-Q (45)
erage elution time and the HETP can be de- E
618 18 Modelling, Design, and Control of Downstream Processing
or
as in the Freundlich adsorption equation. In
the case of elution, on the other hand, Henry’s R, = H(K2 -K 1 ) p / 4 (1 -!- H K J i m ~( 5 1)
equation is used with /3= 1, which is similar to
Eq. (21). In general, K,, changes during the Once again, the smallest possible HETP and
course of absorption. However, if p is small, the largest possible difference in K values are
Kf, can be calculated using the shape of the required to obtain a large R,.
breakthrough curve:
Kfa =
PbqO . 3.4 Membrane Separation
c,(t, - t,)
Membrane separations commonly used in
the bioindustry include reverse osmosis (RO),
ultrafiltration (UF), microfiltration (MF), dial-
Equations (34) and (46), with Kf, obtained ysis (DS), electrodialysis (ED), and pervapora-
from Eq. (48), provide breakthrough curves tion (PV): Separations by RO, UF, and MF
and elution curves by numerical calculations, are accomplished by forcing a fluid through a
where t B and XBrepresent time and dimension- porous membrane. Solid particles and/or large
less concentration at the start of break- protein molecules build up as a layer on the
through, and t E and X, are time and dimen- surface of the membrane. Thus, to attain a
sionless concentration at the end of break- reasonable throughput, the pressure drop (the
through, respectively. Moreover, when the af- so-called transmembrane pressure) should be
finity constant between the stationary phase increased or the resistance to permeation de-
and solute is very high, as in antigen-antibody creased. Cross-flow filtration aims at the pre-
system (p+O.l), equilibrium can be approxi- vention of cake formation in order to decrease
mated by an irreversible equilibrium and a the resistance due to retained feed components
breakthrough curve can be easily calculated. on the membrane surface. The gel-polarization
or concentration-polarization model has been
commonly applied to the calculation of the
Design and Control of Separation Systems 619
performance for cross-flow filtration. The av- velocity is greater than the value calculated
erage transmembrane pressure drop, A P T M , is from Eqs. (5 5 ) and (56).
defined as The phenomenon of decline of flux rate
with time is unavoidable in cross-flow filtra-
Pif Po tion. PATEL et al. (1987) reported that the
APTM =-
2 fouling process could be assumed to follow a
first order mechanism:
where Pi is the inlet gauge pressure and Po is
the outlet gauge pressure of cross-flow filtra-
tion.
The permeate flux, J, can be expressed as where Jt is the flux at time t , J1 is the flux at
t = 1 min, and b is the fouling constant. Their
conclusions concerning the fouling process
(53) may be summarized as follows:
a) Higher flow rates reduce the rate of flux
where R , is the resistance of the mem- decline.
brane, b) Higher pressures increase the rate of
R, the resistance of the retained com- fouling.
ponents, and c) Higher cell concentration leads to a
ps the viscosity of the fluid. greater fouling rate.
tude in the second step, this flexibility tained from 100 papers on protein purification
should be exploited for process optimi- published during 1984 in eight journals. Ion-
zation or process safety; i.e., process op- exchange chromatography was the most com-
timization can be affected by changing mon reported method. Although there is no
various design variables and/or operat- strict sequence for application of the methods,
ing or manipulating variables. a distinct trend is obvious. Homogenization is
generally followed by precipitation, then ion-
Of course no step can be taken in isolation. exchange chromatography, affinity separa-
It is readily apparent that if one wants to tion, and finally gel filtration. Sometimes the
choose the best configuration, then optimal sequence of purification techniques is not
conditions are required based on correct in- made explicit. However, it is a logical one.
put-output relationships. Viewed in this way,
steps 2 and 3 are already included in step 1.
w
t
4.1.1 Process Synthesis (Filtration)
1 -Waste
Orum -Filter
Extensive experience and intuitive compre-
hension of current techniques are required to
properly determine process configuration. In
+
(Extration by butyl acetate 1 B A ) )
other words, no systematic way to solve such 4
Solute
Washing, decoloring -Waste
problems has yet been developed. One possible
approach is to choose one of the candidates 7-
(Extration by K2C03) -Recovery of BA
amenable to solution and treat it as a combina- 4
torial problem.
In any case, process configuration or alter-
native structures of purification systems
+
Raw material
( Condensation I Evaporation
I-Butanol
should be taken into account if systems are to
(Crystallization I
be investigated that are more effective and eco-
nomical than current ones. For example, peni- 1-Butanol washing
cillin G (PenG) has been industrially purified 1 Drying 1
by the system shown in Fig. 2. PenG is ex- 4
Product PenG
tracted from the broth into butyl acetate (BA)
and then re-extracted from butyl acetate into Fig. 2. Flow-sheet of a separation and purification
a water phase by chemical reaction with system of penicillin G .
KZC03.
The condensed PenG-K solute is again ex-
tracted into butanol (BuOH), with the PenG Precipitation is appropriate for dealing with
becoming purified by crystallization because large quantities of material, and it is less af-
the solubility of penicillin in BuOH is low. In fected by interfering non-protein material than
an alternative configuration for the separation adsorption and chromatographic procedures.
process, PenG can be separated from other Affinity methods can be applied at an earlier
amino acids by LC using absorption column stage, but the materials are expensive; it makes
(preparative) chromatography rather than sense to initially use less costly ion-exchange
through extraction by BA and concentration media to reduce protein loads and to remove
of PenG into PenG-K. remaining fouling substances. Gel filtration
Another example, the sequence of protein has the least capacity for loaded protein, but it
purification, is generally reported as involving serves an important function in removing self-
the series of steps: homogenization, cell preci- aggregates of otherwise purified proteins, and
pitation, ion-exchange LC affinity chromato- it also allows for changes of buffer. In any
graphy, and finally gel filtration (BONNERJEA case, experience or expertise has an important
et al., 1986). Corresponding data were ob- decision-making role at this stage. In many
Design and Control of Separation Systems 62 1
cases it cannot be strictly determined whether the 1st level other conditions can help to min-
or not the structure selected was really opti- imize the extractor cost when a, and Y, are
mal, because not all possible structures can be fixed, as in the case of pH for the extraction of
tested in practical calculation and design. PenG into butyl acetate during penicillin puri-
fication (Fig. 2). Note that at low pH PenG
will be inactivated more in the water phase
4.1.2 Process Flow-Sheeting than in the organic phase. However, the parti-
tion coefficient, K , of PenG between the butyl
and Optimization acetate and aqueous phases depends heavily on
the pH, and a high pH is preferable for in-
In the stage of input-output calculation, creasing K . Use of the appropriate pH will
where material and energy balances can be ta- thus maximize the overall recovery of PenG in
ken into account (flow-sheeting), it is prefera- a continuous extractor. The solution to the
ble to utilize commercial program packages. problem is therefore obtained at the first
Several packages are already available or un- stage.
der development (e.g., BPS by ASPEN
Tech).
As already mentioned, the optimization
stage cannot be done separately, and it re- 4.2 Control of Separation Processes
quires cooperative effort. Optimization can be
carried out based on flow-sheeting calcula-
tions. A control system is necessary for each unit
used in the separation during industrial opera-
tion. There are many items that should be con-
Total PI
2nd level
sidered in terms of the control of the separa-
(Performance Index) tion process. We will not touch upon this
problem here, however, because space is lim-
ited. Only the following point is stressed: if
..... 1st level one wants to control a system, the output
sub PI
Fig. 3. Multilevel technique for optimization.
Tab. 6. Principles of On-Line Sensors in Separation
Processes
Many IIUmeriCal iteration methods have Absorbance (in visible range)
been developed for the optimization tech- UV Absorbance
nique. In this separation system, however, the Refractive index
multilevel technique shown in Fig. 3 is prefera- Fluorescence
ble because the overall calculation to achieve NMR spectrum
optimization is readily understood. In this Mass spectrum
configuration, two indices have important Others
roles, the purification index (or partition coef-
ficient) ai,defined as
AARNEHALME
Helsinki, Finland
KARIM
NAZMUL
Fort Collins, Colorado 80523, U.S.A.
1 Introduction 626
2 A1 and Expert Systems 626
2.1 Basic Concepts 626
2.2 How to Build an Expert System 627
2.3 What Kinds of Problems Should be Solved with Expert Systems? 627
3 Expert Systems for Bioreactors 628
4 A Practical Example 628
4.1 Computer System and the Programming Environment 629
4.2 Example of Rule-Based Programming 630
4.3 Materials and Methods 631
4.4 Examples of Experimental Tests 632
4.5 Some Practical Aspects 635
5 References 635
Biotechnology Second, Completely Revised Edition
Edited by H.-J. Rehm and G.Reed in cooperation with
A. Puhler and P. Stadler
copyright@WILEY-VCH Verlag GmbH, D-69469 Weinheim (Federal Republic of Germany). 2001
by using if ... then . ..else-type rules. Other ble consequences of a statement supposed to
methods that are sometimes used are represen- be true or false.
tations on the basis of frames or schematic The last step in development is to validate
nets (HAYES-ROTHet al., 1983). A special the expert system. This is done basically by de-
class of expert system is the system that oper- termining how well the system knows its field
ates with real-time data. Most of the applica- of expertise and how logically it can “think”.
tions related to process control systems have In practice, the only way validation can be
this feature. achieved is to test the system thoroughly with
the expert. Although the knowledge base was
constructed with the expert’s assistance, it
does not necessarily imply that the inferred re-
2.2 How to Build an Expert System sults must be rational.
A special feature of expert systems used in
Since an expert system is essentially a com- connection with process control systems is that
puter program, the basic elements needed to they operate with real-time data such as sensor
construct it are a computer and proper soft- readings. The proper connection of data of
ware tools. Although special computer hard- this type with the knowledge base requires spe-
ware exists for A1 applications, it is current cial tools, such as event-handling mechanisms
practice to use standard hardware such as engi- and temporal logic to ensure that time-depend-
neering workstations and personal computers. encies of data and events are taken correctly
Special “development shells” are available on into account. Many expert systems applied to
the market for software development. bioreactor control are of this type, and thus re-
The part of an expert system most laborious quire these special features when imple-
to construct is the knowledge base. It usually mented.
requires careful analysis of the problem by an
expert and a knowledge engineer. The expert
(or experts) is a person who knows the prob- 2.3 What Kinds of Problems
lem thoroughly and can solve it by applying Should be Solved with Expert
his specialized knowledge. The knowledge is in
the expert’s head, often as a complex mixture Systems?
of schematic descriptions, rules, and impor-
tant values. The knowledge engineer is a per- It is important to realize that an expert sys-
son responsible for gathering this knowledge tem is not always the best solution to a prob-
and organizing it into a rational form for the lem. If valid mathematical or biochemical
computer system. The first step is usually to models representing the main phenomena of
interview the expert. This is done iteratively, the problem are available, it is wiser to use
proceeding in several steps, each consisting of them than to build a knowledge-based system
a refinement of the knowledge obtained in the that explains the same phenomena on the logi-
previous step. If the knowledge is represented cal level. In fact, such models represent high-
in the form of rules, it means that the rule set level and concentrated knowledge that is diffi-
must be modified several times in order to cult to replace with rule sets or similar knowl-
agree with the expert’s thinking. The second edge commonly used in expert systems. There-
step is to run the inference engine on the pro- fore, the problem is best treated by employing
totype knowledge base together with a proper conventional methods such as mathematical
database, and ascertain whether the conclu- calculation, simulation, and parameter identi-
sions drawn by the system are in accord with fication. If the models or their usages are com-
the expert’s opinions. If the knowledge base is plicated, however, an expert system may be
constructed of rules, the conclusions may be helpful in guiding a non-specialist in their
reached by chaining the rules backward or for- use.
ward. Backward chaining typically explains
why a statement is true or false. Forward
chaining, on the other hand, determines possi-
628 19 Expert Systems for Biotechnology
Fig. 1. Symptoms and rules, inference engine, con- nect subsystems for expert analysis. @ Rules for
nectivities, etc. The rules are generated to intercon- phase determination.
4.1 Computer System and the system is unable to make a firm decision based
on real-time data. Data-producing state and
Programming Environment parameter estimation algorithms may be used
to assist the expert system to come to the “cor-
In this example the expert system was de- rect” conclusion.
signed to run in a separate PC/386 connected 2. Data from sensors must be interpreted
to the computerized fermentation unit (BIO- and reasoned about to identify types of events
STAT with Rinteknos MFCS) with a serial and “objects” known to the knowledge base,
computer link (RS-232C). The fermentation e.g., different data and rules are used during
computer system provides on-line measure- different phases of fermentation. The data are
ment data as well as the laboratory analytical time-dependent; i.e., the information included
data available after sample analysis. The ex- may be a function of the time at which the
pert system PC picks up the data it needs for data are used in reasoning. If the data are am-
reasoning and informs the operator of its con- biguous or inaccurate, the reasoning system
clusions via an interactive user interface. All must make “reasonable” and “safe” deci-
process control operations are performed via sions.
the control computer interface. 3. Characteristic curves are to be included in
Program development of the expert system the knowledge base, with variances repre-
was done under OPS 83, which is a rule-based senting the most “deviant” behaviors observed
on-line programming tool to develop expert in previous experiments.
systems. The following features and require- 4. The knowledge base is dynamic rather
ments were specified for the system: than static, as is the case with most expert sys-
1. All the decisions are to be based on real- tems. This means that as more experience is
time data, although off-line sample analysis acquired about a fermentation, the knowledge
may enter the decision process if the expert base and the rule sets are modified to account
630 19 Expert Systems for Biotechnology
’ -
Acid
Fig. 2. Measurement
Ferrnentor and regulation of pH
Recorder systems.
6. Once the identification of a phase is com- condition of the measurement and control
pleted (this may be based on a small subset of functions in the fermentation system are de-
the on-line measurements), the total set of fined as “Level 1 Rules”. The expert system
measurements is analyzed for faults in the continuously checks these rules, and if a fault
process and for contamination. If contamina- is suspected in any element of the measure-
tion is suspected, the operator is guided in the ment and control loops, it will activate the
taking of a sample and having microscopic “Level 2 Rules”, which then analyze the specif-
analysis performed on the sample in order to ic fault and make recommendations to the op-
determine the nature of contamination. In this erator concerning possible courses of action to
sense the decision made by the expert system is alleviate the situation. The basic strategy for
not necessarily a binary one. finding faults in the pH controller is shown in
7. If and when the measurement signals are Fig. 3. Here it is assumed that any inaccuracy
noisy, a moving average filter is used before of f 0 . 3 is due to a malfunction of the con-
the measurements are processed. The expert troller. Fig. 4 shows some of the OPS 83 codes
system decides which variable needs filtering for this strategy. Similar rules can be formu-
and it does it in real time. lated for the pH sensor and acid/base pumps,
as well as for the other sensors and controllers.
The windowing technique is used in operator
PH
set point sensor
off off
I
7
in the controller
Fig. 3. Logic for find-
ing faults in the con-
I I
-
&dev &pH.value[l] - &pH.stand[l];
if (&base.value[l] - &base.value[2] > 0.0)
The prototype expert system was tested in
the laboratory environment with several paral-
-
&base-on= l b
e k e &base-on Ob
if (&acid.value[l] &acid.value[2] > 0.0)
lel fermentations of Bacillus subtilis producing
a-amylase. Materials and methods used in
~
&acid-on = 1b
else &acid-on = Ob; these experiments are summarized below.
if (( Bdev > 0.3) A (Bbase-onA not (&acid-on))) Microorganism: Bacillus subtilis BRB 423
make (symptom name = controller; circuit = pH); (glucose repression-negative) carrying the B.
if (( &dev < -0.3) A (&acid-on”, not (&base-on)))
make (symptom name = controller: circuit = pH); amyloliquefaciens a-amylase gene as a single-
copy chromosomal integrate.
1:
Media and growth conditions: five Braun
Fig. 4. OPS 83 rules for finding faults in the PH Biostat M fermenters with working volumes of
controller. 1.0 L were available for this study. Fermenta-
tion media consisted of 1.0% glucose, 1.0%
yeast extract, 2.0% tryptone, 85 mmol NaCl,
interfacing. Different windows are created to 17 mmol K2HP04, and 5 mg/mL of antibio-
show multiple menus simultaneously. Special tics. Fermentations were carried out at 37°C.
care is taken to give meaningful messages to The pH was maintained at 7.2, the aeration
the operators. Color-coded information is pro- rate at 1.2 vvm, and the stirrer speed at 800
vided to distinguish the quality of information rpm.
632 19 Expert Systems for Biotechnology
Instrumentation: Each fermenter was equip- ever, the operating conditions were slightly
ped with a dissolved oxygen probe (Ingold), different. Different fermentation behaviors
pH probe (Ingold), temperature element, and probably resulted, although the indicators of
foam detection probe. The control of pH, tem- main variables of fermentation conditions re-
perature, stirrer speed, and antifoam addition mained the same. It was therefore difficult for
was accomplished by set-point control with a the usual operator to invoke corrective control
Braun Biostat system. A gas analyzer (Sie- operations during the runs.
mens) measured carbon dioxide and oxygen However, a skilled operator might have
from the exhaust gas. found the following symptom that was de-
Process control and data acquisition were tected by the expert system. Figs. 5 and 6 show
performed by the MFCS software package acid and base additions and C 0 2 and dry
(Rintekno), which was especially designed for weight measurements for the two fermenta-
fermentation processes. The MFCS package tions. In run 870115 much more base was ad-
runs on the DEC (Digital Equipment Corp.) ded than in run 870116. Since the acid and
Micro PDP/1 1 computer using the Micro RSX base additions represented cumulative plots of
operating system. the duration that pumps were open (calibrated
Analytical methods: a-amylase activity was to mL of acid/base addition), it is possible
determined by the Phadebas (Pharmacia) amy- that the pumps were pumping air in run
lase test. Turbidity was measured with a Klett- 870115. Fig. 7 shows the results of fault analy-
Summerson photoelectric colorimeter. Dry-cell sis of run 870115. The expert system predicts a
weight was determined by freeze-drying centri- possible fault in the base pump at 3.33 hours
fuged cells to constant weight. Glucose con- into the fermentation. After detecting this pos-
centration was measured with a Beckman glu- sible fault, the operator should use the fault
cose analyzer. study menu to determine the reasoning em-
ployed by the expert system in its decision. He
may accept the decision, but he may also ig-
4.4 Examples of Experimental Tests nore it by acknowledging that he has seen the
message and elected to take no action.
Test 1. The following is an analysis of runs In this instance the expert system predicted
870115 and 870116, conducted simultaneously that the fault was probably due to pumping air
under supposedly identical operating condi- through the system, and, in fact, it was con-
tions. Due to a hidden equipment fault, how- firmed that the pump in real time was intermit-
I
I
Totai Total
----
(02 Drywt base acid
i%voilh~tg/mllfmll I m l l
64- 40-
32- 20-
Total Total
c02 Drvwt base acid
tently pumping air rather than base. It could mentation, thus preventing the operation from
be concluded, although not conclusively, that proceeding with faulty equipment.
the difference between the two runs was main- Test 2. In another test, run 880034, deliber-
ly in the faulty base and acid pumps in run ate faults were introduced at 3.5 hours in the
870115. The result was lower productivity of pH controller and at 6.35 hours in the COz
a-amylase in that run. The use of an on-line sensor. The pH fault was introduced by adding
expert system in an actual production process a bias in the acid pump, and the COz fault ori-
could have resulted in a faster response from ginated with the addition of a pure carbon
the operator at about 3.33 hours into the fer- dioxide pulse into the carrier stream. Figs. 8
634 19 Expert Systems for Biotechnology
Total Total
co2 base acid fd
i%voll
+--- pH imti imtl
4 000-8000-
3200- 7600-
1 600- 6 800- 40 0- 40 0
80 0- 3 MO- 800-
60 O- 2 400- 6 00-
4 Stationary phase 1 67 h
and 9 show the plots from the MFCS system. mind, because the same phenomena can be de-
Fig. 10 indicates the analysis by the expert sys- scribed in several different ways.
tem. The faults were detected about two min- The problem of updating the knowledge
utes after they were introduced. The introduc- base was not considered during the testing of
tion of COz into the carrier stream perturbed the prototype system. In practical cases the up-
the OUR measurements and upset the OUR dating problem should always be taken into
reading. This, however, created problems for account. Even if the general philosophy of
the expert system, as it did not know that the knowledge base construction is supposed to be
faults were simulated. The expert system did independent of changes in the organism, me-
predict the faults in the pH control system cor- dia, or the process equipment, experience
rectly, but it failed to detect the fault in the shows that updating the database is usually in-
C 0 2 system. Instead, using heuristic “two out sufficient, and that rules must also be
of three” rules, it found that both OUR and changed. This is a feature of expert systems
C 0 2 sensors had behaved in an identical man- that can cause severe problems in practice.
ner (sharp increase), indicating rapid growth. Changes in the knowledge base should, in
The DO probe did not show this kind of activ- principle, be made by the same expert who
ity; it therefore reasoned that the gas sensors contributed to its creation. In practice, howev-
were performing properly and that the r.p.m. er, this problem may be overcome in well-de-
control system was probably faulty. Of course signed software by extracting those rules and
this analysis was wrong. However, since the data values that appear to require it, and de-
faults were not real ones, and merely scanning scribing the necessary changes as clearly as
the figures could not result in a determination possible.
of the cause of the abnormalities, the decision
of the expert system was logical. This would
also have been the conclusion of a human ex-
pert. 5 References
4.5 Some Practical Aspects AARTS,R. J., SUVIRANTA, A., RAUMAN-AALTO,
P., LINKO,P. (1989), An expert system in en-
zyme production control, in: Proc. Int. Conf. on
The knowledge base in the example above Biotechnology and Food, February, Stuttgart,
consisted of about 300 rules, an average num- FRG, pp. 20-24.
ber. However, as a prototype system, the CHEN-QI,SHU-QING,W . , JI-CHENG,W. (1989),
knowledge base did not cover all practical Application of expert system to the operation
cases, but rather demonstrated the functionali- and control of industrial antibiotic fermentation
ty of the system. Even with a knowledge base process, in: Proc. 4th Int. Congr. on Computer
of this size, inferencing with a PC takes time. Applications in Fermentation Technology (FISH,
N. M., Fox, R. I., THORNHILL, N. F., Eds.),
It helps considerably to divide the rules into London: Elsevier.
various levels and to limit the rule set by HALME,A. (1989), Expert system approach to re-
“zooming” the inferencing in the lower level. cognize the state of fermentation and to diagnose
The number of rules for a practical situation faults in bioreactors, in: Proc. 4th Int. Congr. on
could easily by 500-700, depending on the Computer Applications in Fermentation Tech-
complexity of the fermentation facility. A sin- nology (FISH, N,M., Fox, R. I., THORNHILL,
gle PC may not be adequate in such a case. N. F., Eds.), London: Elsevier.
Experimental tests have shown that the rea- HAYES-ROTH, F., WATERMAN, D. A., LENAT,D.
soning of the system is quite reliable provided B. (1983), Building Expert Systems, Reading,
Mass.: Addison-Wesley
the underlying rules are robust in the sense KARIM,M. N., HALME,A. (1989), Reconciliation
that they explain logical rather than numerical of measurement data in fermentation using on-
behavior. It is not always easy to devise rules line expert system, in: Proc. 4th Int. Congr. on
in which this condition is valid. When inter- Computer Applications in Fermentation Tech-
preting the thinking of an expert and formulat- nology (FISH, N. M., FOX, R. I., THORNHILL,
ing the rules, this precept is good to keep in N . F., Eds.), London: Elsevier.
636 19 Expert Systems for Biotechnology
LUBBERT, A., HITZMANN, B. (1987), Are there any in Fermentation Technology (FISH,N. M., Fox,
prospects for expert systems in chemical engi- R. I., THORNHILL, N. F., Eds.), London: Else-
neering?, Hung. J. Znd. Chem. 15, 39-45. vier.
LUBBERT, A., HITZMANN, B., KRACKE-HELM, H.- STEPHANOPOULOS, G., STEPHANOPOULOS, G. N.
A,, SCHUGERL,K. (1989), On experiences with (1986), Artificial intelligence in the development
expert systems in the control of bioreactors, in: and design of biochemical processes, Trends Bio-
Proc. 4th Int. Congr. on Computer Applications technol., 241-249.
Index
A - procedures 554
Accuracy, definition of 45 - schematic diagram 552
Acetic acid, degradation in wastewater Adenine dinucleotides, fluorescence 185ff
purification, by Methanosarcina barkeri Adenosine triphosphate (ATP), formation by
451 methane bacteria 449f
- - energy balances 446 - monitoring, bioluminescence assays 21 1
- - mathematicalmodel 463ff - - 31P-NMR 211
- gas analysis in bioreactors 60 - production in aerobic metabolism 414
Acetobacterium, anaerobic wastewater process Aeration, in stirred tanks, effect on circulation
447 time 312
Acetogenium, anaerobic wastewater process - - effectonmixingtime 311ff
447 Agitation system, measurement of power input
Acetoin, gas analysis in bioreactors 60 126f
Acetone, gas analysis in bioreactors 60 Air bubble see Bubble
Acoustic methods, for biomass determination Air flow, measurement, accuracy of 21
195 Air lift loop reactor, axial dispersion coefficient
Acoustic resonance densotimetry (ARD) 195, 364f
43 1 - liquid velocities 363
Activated sludge system, mass balance Air lift reactor, distributed parameter gas model,
equations, for a stirred tank 43 Iff single-cell protein production 376f
- mass transfer resistance, in flocs 431 - drafttube 374
- modelling sludge settling 435 - simulation of baker’s yeast process 402
- wastewater treatment 430 - structuredmodel 372
Adaptive control 551ff Air lift reactor flow, models 359ff
- schematic diagram 521 - - driftfluxmodel 359f
- self-tuning control 521 - - energy balance approach 360ff
Adaptive optimization 55 Iff - - momentum balance approach 362f
- definition 551 Air lift reactor models, concentration profiles
- multivariable on-line optimization 556 317
- of batch and fedbatch bioreactors 557f - cyclic operation 378
- of continuous bioreactors, baker’s yeast - for oxygen transfer 374
culture 552ff - for phenol degradation 374
- - implementation 554f - gas phase circulation 374f
- - parameter estimation 553f - mixed cells in series 374f
- - phases 555ff - two-section loop reactor 364
- - processmodel 553 Alcaligenes latus 42ff
638 Index
Biofilm processes, effect of colloids 423 Bioprocess kinetics, gene product formation
- in wastewater treatment models 418ff 490f
Biofilm reactors, modelling 425 Bioprocess modelling 488ff
Biofilm thickness 420ff, 426ff, 429 - bioplantmodel 489
- optimum 429 - bioreactormodel 489
- steady-state 423f - molecular model 488f
- - sloughing 424 - population model 489
Bioflocculation 436ff - simulation 488ff
- effectofPHB 436 - single-cell model 488f
Biofuel cell 84f Bioproduct purification, principles 607
Biogas, anaerobic wastewater purification, Bioproduct separation 609ff
composition 454 - centrifugation 609
- - gasflow 453ff - chromatography 611,614f
Biological models 277ff - control 621
- types 270 - extraction 610f
Biological oxygen demand see BOD - membrane filtration 609f
Biological test systems, for bioreactor - on-line sensors 621
characterization 125 - principles 607
Biomass, assumptions for modelling 409f Bioreactor 109ff
Biomass concentration 181ff - biological test systems 125
- definition 181 - bulk mixing properties 114ff
- determination methods 182ff - characteristic properties 1lOff
- on-line calculation, anaerobic wastewater - fixed-bed reactor 450
process models 470 - - wastewater treatment 479
Biomass determination 18Iff - fluidized-bed reactor, wastewater treatment
- acoustic methods, acoustic resonance 427ff, 450
densitometry (ARD) 195 - foamcontrol 14f
- - piezoelectricsensor 195 - global measuring techniques 119ff
- calorimetric methods 188ff - heat transfer 119
- electrochemical methods 192ff - hydrodynamics 113
- - amperometric sensors 194f - local measuring techniques 128ff
- - fuel cell type sensors 193f - - bubbleparameters 128ff
- - impedimetricmethods 192f - - liquid flow properties 138ff
- - potentiometric sensors 193 - mass transfer 110ff
- filtration methods 190ff - mixing behavior 116ff
- in anaerobic wastewater processes 450 - modelling 269ff
- optical sensors, culture fluorescence - oxygentransfer 38ff
185ff - pH-auxostatic operation 474f
- - nephelometric methods l82ff - pH-chemostatic operation 474
- thermogram 190 - stirred tankmodels 301ff
- viscositymethods 192 - tower reactor models 352ff
Biomass estimation, oxygen balancing 8 Bioreactor control 512ff
Biomass sensor, amperometric 194f - design of a control system 512
- fluorescence sensor 209 - use of expert systems 628
- fuel cell type 193f Bioreactor identification 226ff
- impedimetric 192f - application of EKF 232ff
- on-line 181ff - definition 226
- optical 182ff - mathematical formalism 227ff
- piezoelectric 195 - model formulation, model equations 228ff
- potentiometric 193 - - statevariables 228ff
Biomass yield, as a function of COD 414 - models 227ff
Index 641
R Respirometer 36
Rate equations, assumptions for modelling Reynolds number l l f , 369,619
410f - turbulence in bioreactors 115
Reaction rates, dynamic error analysis 54ff RNAmeasurement, flow cytometry 204f
Recombinant fermentation, application of - sample-flow analysis 205
parameter estimation methods 495 Rosenbrock method 458
- bioprocess kinetics 487ff Rotameter 59
- bioprocess modelling 487ff Rotating biological contactors 429f
- determination of genetic parameters 491ff - hydrodynamics of the liquid film 430
- - gene expression efficiency 491 - wastewater treatment, modelling 429
- - growthratio 491ff Ruppel number 12
- - kinetic model 492ff
- - method 492ff
- - plasmidlossrate 491ff S
- determination of kinetic parameters 491ff Saccharomyces cerevisiae,adaptive
- effect of specific growth rate, on gene optimization 552ff
expression rate 499f - biological properties 387ff
- - on plasmid content 498f - continuous culture 340
- - on plasmidstability 500f - diauxic batch growth, metabolic regulator
- effect of temperature, on gene expression rate model 296
501f - growth conditions 150
- - on plasmid-harboring cell fraction 502f - growth simulation, metabolic regulator
- instability parameter 494 model 295
- microbiological process design parameters - productionof 539
496 Sample pretreatment 56ff
- optimization 496ff Sampled-data control system 252ff
- - dilutionrate 497ff - configuration of 253
- ratio of plasmid-free to plasmid-harboring - controllability region 253
cells 492ff - state-space vector 253
Recombinants, productivity, genetic parameters Sample-flow analysis 199f
491ff - autoanalyzer 199,207f
- - kineticparameters 491ff - cell-viability determination 202
Redox meter, sterilizable 19 - DNA measurement 205
Redox potential, definition 9 - FIA system 199,207f
- determination of 18f - pHi measurement 203f
- - in anaerobic wastewater purification 45 1 - protein monitoring 207f
- rHvalue 9 - quasi-on-line methods 199
- use in process control 9 - RNA measurement 205
Redundancy, in automation systems 569ff - thermistor 208
Reflectometry 96 Sampling, optimization of 252ff
Residence time distribution, biological test - - sensorallocation 262
systems 125 - theorem 252
- gas 121f - time 252ff
- liquid 123 Sampling module 152f
- measurement 121ff - Bio-filtrator 152
- - flow-follower techniques 124f - Biopem 152f
- - tracertechniques 121ff - discus-shaped 155
- of gas bubbles 11If - Millipore 152f
- time-of-flow (TOF) measurement 139ff - rod-shaped 153,155f
Resipiratory activity, long-term regulation Sampling period, determination of 253
395 Scale-up 301f
Index 655