. A.

RADICALLY DIFFERENT FORMULATION AND

SOLUTION OF THE SINGLE-STAGE FLASH PROBLEM

J. F. BOSTON*~
Department of Chemical Engineering, University of Toledo, Toledo, OH 43666,U.S.A.

and
H. I. BJtrrrt
Union Carbide Corporation, South Charleston, WV 25303,U.S.A.

(Received 29 June 1978)

Ah&a&-A new algorithm has bean developed for performing muhi-compondnt, single-stage flash calculations. The
algorithm is based on a combination of concepts which makes it capable of solving wide-boii. narrow-boiing,
and highly non-ideal systems and which reduce the computational load associated with enthalpy and equilibrium
ratio evaluations. These concepts involve the choice of iteration variables, the form of the describing equations,
and the convergence procedure. The performance of the algorithm is demanstrated using a number of example
problems involving both moderately and highly non-ideal systems, as well as nearly ideal wide- and narrow-boiling
systems.

Seo~Multicomponent vapor-liquid equilibrium flash al9oritbms are widely used in computer process cal-
culations. The purpose of this work is to introduce a new flash algorithm based on several related concepts which
make it superior to, and radically different from, conventional algorithms. These concepts may be summarked as
follows:
(I) Complex phase equilibrium and enthafpy models are used only to generate parameters for simple models of
equilibrium ratio and vapor and liquid enthalpy departure.
(2) These simple model parameters replace temperature, pressure, vapor and fquid composition, and phase
fraction as the problem independent variables. We call the conventional independent variables the primitive
variables and our new independent variables the iteration variables of the system. The iteration variables are
chosen to be as independent as possible of the primitive variabks aad of each otkzr. Because of their
independence. precise initialixation of the iteration variables is not critical to good algotithm performance.
(3) The system describing equations are expressed in terms of the simple models for equi&ium ratio and
enthalpy departure and are rearranged in a novel manner so that a complete solution for the primitive .variabks
may be obtained by converging a sin&.-variable iteration loop. The single iteration variable is a unique combination
of temperature and phase ratio which completely eliiinates the need to make a distinction between wide- and
narrow-boiling systems.
(4) The primitive variables resulting from the solution of the describing equations for a given set of iteration
variables are used to calculate a new set of iteration variabks. A direct substitution system of equations results.
Because of the principal of independence used to choose the iteration variabks the Jacobian matrix of this system
is strongly d&oually dominant and usually not greatly diBerent from the identity matrix. Such systems are
particularly amenabk to solution by a quasi-Newton technique, such as Broyden’s method, becawe the identity
matrix is a good initial Jacobi matrix and because the off-diagonal ekments that must be approximated by the
Broyden method are small and do not greatly affect the new estimates of the iteration variables.
(5) The main effect of system non-idea& is to reduce the independence of some of the iteration variables.
Non-ideality is seldom so great as to eliminate diinal dominance so that these effects may be handled efficiently
by the Broyden method.
(6) If the algorithm is used in an iterative environment, such as a process simulator, not only the converged
values of the. iteration variables, but also the Broyden approximation to the system Jacobian matrix, may be saved
for subsequent solution at perturbed conditions.
Whik not all of the features introduced above are new, taken together they are used to develop a complete
algorithm that efficiently considers not only convergence but also differing types of systems, initialixation, and
restart capability.
Before Use compkte algorithm is presented the problem considered by this paper will be formally defined,
conventional algo&ms for its solution discussed, and the. concept of ush physical proper&s or physical
property rekted parameters as independent variables introduced.
An important assumption throughout the paper is that equilibrium ratios and enthalpks are computed by some
physical property pa&age not a part of the actual Bash algorithm. This means that we do not know what
thermodynamic models may be used and that, in all probability, composition derivatives are not available from the
paclnge.
The performance of tbe algorithm is demonstrated using a number of problems involving both moderately and
highly non-ideal systems, as well as nearly ideal, extremely wide- and narrow-boiling systems.

Cm&akmaNi -A new algorithm has been developed for performing muhicomponent, single-stage
hash calculations. The algorithm has the following desirable performance characteristics:
(1) The algorithm exhibits efficient and stable convergence behavior for a wide range of systems without the
*Author to whom cormspondence should be addressed.
tCutrent address: Energy Laboratory, Massachusetts Jnatitute of Technology, Cambridge. MA 62139.U.S.A.

109
110 J. F. BWON and H. I. Bturr

need for user manipulation of convergence parameters. These systems include wide-boiling, narrow-boiiing, highly
non-ideal systems, and non-ideal systems which result from using psuedocquilibrium ratios to account for
equilibrium reactions in either the vapor or liquid phase.
(2) The. algorithm is as efficient as conventknaI aigotiths for ‘easy problems and more efficient for
‘difhcuIt’probkms. EiIickncy Is measured by the number of enthalpy and quiliium ratio evaluations.
(3) The user is not required to furnish initial estimates of vapor and liquid composition and the algorithm is
relatively insensitive to initial estimates of temperatureor pressure.
(4) Convergence is achieved very rapidly when the fiash calculation is restarted from a previously converged
solution at different conditions. This behavior results from the fact that the iteration variables are chosen to be as
independent as possible of temperature, pressure, etc. and that the Broyden approximation to the system Jacobian
matrix may be saved from the”previous coovergeoce, and oot simply because the new conditioos are ‘close’ to the
previously converged conditions.
The algorithm is especially suited for use in a process simulation system in which case ah of the characteristics
listed above are desirable. The authors feel that the concept applied here can be applied to other unit operations
models as well. In fact it is similar in marry ways to the Boston-SuUivan[l] distillation algorithm. The two
algorithms differ in many sign&ant details but most sig&antly the distillation algorithm does not use a
quasi-Newton convergence method to accelerate- convergence of the main iteration variables.

~~DFscIIpI1oN Table 1. Types of single-stage vapor-liquid flash

The multicomponent,single-stage,vapor-liquid equilib-
rium flash, depicted schematically in Pig. 1, is one of Flash Specified Calculated
the most frequently used elements in computer-aided type quantities quantities
process calculations.In the most common situation, the
I T, V
composition and condition of the feed to the flash are ;pT K JJ
known, and it is desrired to determine the compositions, :E Q, V
conditionsand flow rates of the vapor and liquide!Ruent ’ IV K ‘I- :;
streams.When the heat duty is not specified,it is usually V v, P T, Q
desired to calculate it from an enthalpy balance. VI T, P KQ
If we adopt the premise that the feed quantities, z, F All cases:
and HF are known,and that x and y are to be calculated, z, F, HF specikd
then there remain five quantities that must be fixed in 1,P.L calculated
order to completelydescribe the system shown in Fig. 1:
T, P, Q, V and L. Since V and L are not independent,it
is further assumedthat L is alwaysa calculatedquantity. indirectly through x and y and, dependingon lIash type,
Any two of the remain& four variablesmay be specified T or P. Of the flash types listed, Type VI is the least
arbitrarily, and the other two determined to satisfy the ditktlt since both T and P are tlxed. Types I, III and V
describingequations. are more dilkult than the other three types because the
The six types of flash calculations that represent all temperature is a calculatedquantity. Types I and III are
possible combinationsof specitkd quantitiesare listed in more difhcult than V because they require that the
Table 1. Types I, II and III are usually referred to as ‘enthalpybalance equation be solved simultaneouslywith
‘adiabatic’flashes, since the total system enthalpy is the other describing equations. For an N-component
fixed. Types IV and V include the common dew point system, there are a total of 2N +3 quantities to be
and bubble point problemsas special cases. calculatedin all cases.
Of the four possible calculated quantities, only The five types of equations that are required to
temperature and pressure have a direct effect on the completelydescribe a single-stagetlash are:
equilibrium ratios and enthalpies. Of these two, the Phase equilibrium:
pressure effect is less complex, and in the case of
enthalpies,weaker. The remainingtwo calculated quan- yi = KiXi (i = 1,2,. . . , N). (1)
tities affect the equilibrium ratios and enthalpies only
Componentmass balance:
L,V,T,P
T t
xiL+y,v= ZiF (i = 1,2,. . . , N). (2)
Total mass balance:

L+V=F. (3)

Enthalpy balance:

H&+HvV=HpF+Q.

Fii. 1. Schematic diagram of sia&-s@c Rash. We refer to this set of equations as the ‘primitive’
 A radicallydifferentformulationand solutionof the single-stageflash problem 111

equations and the set of calculated quantities in Table 1 Table 2. Type I adiiabaiicflash conventionaliterationalgorithm
as the ‘primitive’ variables. Thus, for an N-component
system, there are 2N+3 primitive equations and, for’ (1) Assume x, y, T, B
each flash type, 2N t 3 primitive variables. It is easily
shown that when Eqs. (2), (3) and (5) are satisfied, the
conditions Xxi = 1 and Z yi = 1 are also satisfied.
I I fi (wide-boiling)
Solve @(I(,@)= 0 for
T (narrow-boiling)
coIvTFmmNAL samTIoNALmRmMs
In the following several sections, the discussion will be ’ (4) Calculatei, 9 using Eqs. (6) and (7)
restricted to the Type I case for simplicity. Type I was
(5) Assume new I, p and returnto (2) untillb- 41,E-AI <
chosen because it is the most common ‘dilIicult’ case
ll
and because it implicitly includes the significant features
of the other flash types as well. (6) CalculateHL,Hv.Y
Most algorithms for solving the Type I flash problem
do not use the describing equations directly in their
primitive forms. The most common rearrangement is
L T (wide-boiling) and returnto (2)
(7) Assume new @(narrow-boiling) I until IY(5 e? _
obtained by combining the primitive equations in the
following way. The dimensionless, normalized quantity paired with their dominant variables. For a wide-boiling
/3 = V/F is introduced in place of V, and the Eqs. (l)-(3) system, the value of /I is best determined in the inner
are combined to obtain expressions for the liquid and loop to satisfy @(K, /3) = 0. The converged fi is largely
vapor mole fractions of the form: determined by the distribution of light and heavy
components in the feed, and will therefore be relatively
insensitive to changes in T in the outer loop. A binary
system consisting of a very light component and a very
heavy component would be an extreme example of such
a system. For a narrow-boiling system, on the other hand,
the converged value of T in the inner loop is largely
These two expressions are next substituted into E?q.(5), controlled by the feed composition, and it will be rela-
and the function Q, is introduced: tively insensitive to changes in B in the outer loop. An
extreme example would be a pure component system.
Note that the solution of O(K, j3) = 0 for T with fixed B
requires evaluation of the equilibrium ratios at the trial
values of T.
There are a number of non-derivative algorithms that
Using q = QIF, and introducing the function Y, the
may be used to generate the new estimates of x and y in
energy balance equation then becomes:
step (5) of Table 2. Among the simpler and most common
ones are direct iteration with or without arbitrary
Y~BHv+(l-B)HL-HF-q=o. (9) damping, and the bounded Wegstein method proposed by
Kliesch[rl] applied to the liquid or vapor component tlow
Finally, in terms of B, the total mass balance equation is:
rates. The generalized dominant eigenvalue method of
t Crowe & NishioU], or any of the quasi-Newton
L=(l-j?)F. (10) methods, such as the one proposed by Broyden[6], could
also be used. For highly nonideal systems, in which the
For an N-component system, there are still 2N +3
component flow rates are highly cross-correlated, the
equations, and the 2N t 3 unknowns x, y, T, &3,and L are
bounded Wegstein method could perform poorly since it
in essence the same as the primitive variable set.
is based on the assumption that the auto-correlation of
The equilibrium ratios and the vapor and liquid
the variables is strong compared with the cross-cor-
enthalpies are functions of the primitive variables, x, y
relation.
and T. The describing equations must then be solved by
For nearly ideal systems, it is usually more efficient to
some type of iterative procedure, a variety of which have
proceed directly from step (5) to step (6) than to return to
been proposed. Most of them begin with assumed values
step (2) until the error criterion l1 is satisfied. In effect,
of x, y, and T, and proceed through a cascade of cal-
then, x and y are converged simultaneously with T (or 8)
culations in which Eqs. (8) and (9) are solved for #I and
in the outer loop, and the ‘numberof evaluations of the
T, while x and y are updated using Eqs. (6) and (7).
K’s in step (2) is reduced. For moderately nonideal
A notable feature of this set of describing equations is
systems, it may be necessary to allow 2 or 3 iterations of
the form of the function a, lirst given by Rachford &
the middle loop to achieve the most efficient perfor-
Rice[2], when it is interpreted as a function of K and @.
mance. For highly nonideal systems, in which K depends
It is well known that when Kq. (8) is iteratively solved
very strongly on x, it is usually necessary to converge the
for B with the K’s held constant, its convergence
middle loop, at the same time possibly increasing the
characteristics are superior to those of the alternative
degree of damping if direct iteration is used, thereby
functions derived from T x!_; 1 or K vi = 1.
greatly increasing the number of K evaluations.
A typical iteration scheme is shown in the stepwise Any of a number of methods may be used to generate
outline of Table 2. The distinction between wide- and a new T or @in step (7). For a narrow-boiling system, it
narrow-boiling cases in steps (3) and (7) is based on the is frequently adequate to use direct iteration, Since fl can
idea proposed by Friday & SmithI that, when solving be calculated directly from Eq. (9).
a system of equations, individual equations should be If T is the variable of the outer loop, one approach is
112 J. F. BOSTONand H. I. Barrr
to generate the new T using Newton’s method at
constant composition. The values of (#ZdaTX,,
(d&/U), and (@?/aZ’)., are required for this. The latter
derivative may be obtained through implicit differen-
tiation of E!q. (81, using (aK/an,. This approach tends
to suffer from neglecting the dependencies of x and J on
T and of HL and Hv on x and J. That it works well for
nearly ideal and moderately nonideal wide-boiling
systems can be attributed to the facts that for such
systems the converged value of j3 does not vary widely,
and the calculated x and y will therefore not depend
strongly on T. However, it is encumbered by the fact
that the determination of (aK/Ul., is usually feasible
only by numerical differentiation for nonideal systems.
An alternative approach is to use a secant type
method, based on the results of consecutive passes of
the outer loop. This approach, however, may not work
when the middle loop is not iterated to convergence,
especially for highly nonideal systems.
Before leaving this discussion of conventional al-
gorithms, it is worth mentioning two additional possi-
bilities. The hrst is to solve for jl and T simultaneously
in the outer loop, using either a Newton or quasi-Newton
method. One advantage of this approach would be logical
simplicity, since the question of dominance is automa-
tically taken care of. The previous discussions regarding
the convergence of the x and J l&p apply directly to this
case. The second possibility would be to solve all of the
describing equations simultaneously, again using either a
Newton or quasi-Newton method. This approach does
not seem to be attractive for the single-stage problem, in
view of the requirement for partial derivatives with
respect to x and y. These partial derivatives are required
for every iteration of the Newton method and for the
initialization stage of the quasi-Newton method.
While the preceding discussion is a reasonably
comprehensive review of the conventional approach to
flash problems, it is not the intent of this paper to
evaluate or compare conventional tlash algorithms. The
main purpose is to present a radical departure in which
the primitive variables x, y, T and @ become dependent
rather than independent variables. The new independent
variables are physical properties or parameters related to
physical properties. The advantage of this ‘inside-out’
approach will become apparent as the discussion pro-
ceeds.
In order to introduce the concept of inside-out al-
gorithms, two simple modifications of the conventional
wide-boiling algorithm in Table 2 will first be considered.
Then in the next section, the new algorithm of this paper
will be presented.
The first modification of the conventional algorithm is
based on simply replacing x and y as the iteration vari-
ables of the middle loop with K. An algorithm embodying
this change is outlined in Table 3. It is possible to do this
because the function Q, depends directly on j3 and the
K’s, and only indirectly on the x’s and y’s. Any of the
methods mentioned previously can be used to drive the
K’s to convergence. In such an algorithm cross-cor-
relation between the K’s wouldresult only from system
non-ideality, and the problems associated with cor-
relation of x and y would no longer exist.
This first d&gun-astrates & -msidesut ap_
preach, but the new iteration variables still depend
strongly on T and, depending on the extent of non-
Table 3. Type I adiiatic Rash-modifiedwide boii algorithm
using K in place of x and J as iterationvariables
(1) AssumeK, T
(2) Solve I$= 0 for fi
(3) Calculatex, J using Eqs. (6) and (7)
(4) Calculate%
T; (5) Assume new K and returnto (2) until @ - K[< c,
(6) CalculateHL,Hv,Y
- (7) Assume new T and reutm to (2) until IYl<e
ideality, x and y. The initial estimate of K must be
generated using initial estimates of x, y, and T, and poor
initial estimates of them may lead to very poor initial
estimates of the K’s. Hence, as successive iterates of the
K’s are generated, they may be required to change by
orders of magnitude from the initial estimates.
In view of the undesirable features of the first
modification, the iteration variables of the next
modification are chosen with the idea of reducing as much
as possible their dependence on the primitive variables.
The dependence on temperature is virtually eliminated
by making use of the relative volatilities
air E KJ K, (11)
where r is some reference component. Since in most
systems the temperature dependence of the K’s is of the
same order of magnitude for all the components, the a’s
are much less sensitive to T than the K’s. As iteration
variables, then, they would obviously have much better
convergence characteristics than the K’s. Neither their
initial estimates nor their successive iterates would
depend strongly on the values of T used to generate
them.
The a’s, however, clearly do not constitute a sufficient
variable .&t, since the K’s cannot be determined uniquely
from the a’s alone. Hence, if the K’s are to be replaced as
iteration variables by the a’s, additional quantities must be
included, along with the a’s, in the new variable set. The
reference equilibrium ratio, K,., would be sufficient but it
depends strongly on T. To introduce additional variables
that do not depend strongly on a primitive variable, use is
made of the well-known fact that the dependence of
equilibrium ratios on temperature is represented very well
by a model of the form:
(121
where. P is some reference temperature, and Bi is
largely independent of T. It follows, then, that A, and B,
are logical chdices for the addiinal variables. For any
set of values of A,, B, and the a’s, all of the K’s may be
determined directly through Eqs. (11) and (12). A
mod&&on of the algorithm of Table 3 embodying these
ideas is shown in Table 4.
In step (6) of Table 4, the calculated value, 8, may be
obtained either by analytical differentiation of the actual
K model:
 A radicallydifferentformulationand solution of the single-stageflashproblem
Table 4. Type I adiabatic flash-modifiedwide boiling algorithm
using a, A, and B, in place of x and J as iteration variables
(1) Assume a, A,, B, T
+(2) Calculate K, and Ki (i# r) using Eqs. (11)and (12)
C(3) Solve Q = 0 for 6
(4) Calculatex, y usingEqs. (6) and (7)
(5) Calculate K
(6) Calculate 8, a, 8,
-(7) Assumenew (I, A, B, and return to (2) until
(loi-a/(< er, IA, - A,( < ~2, 12, - &I
(8) Calculate HL, Hv. Y
- (9) Assume new T and return to (2) until jYl< e,
or as:
_ In (K/K:)
3
, --
1
z-z
t
T T’
where K, and K: are evaluated using the actual K model
at the two temperature levels T and T’, with x and y held
constant. In either case, the value of A, must be up-
dated, since it may be composition dependent:
THENEwsoLuTIoN~
The new method is based on an extension of the ideas
just presented, in which T is eliminated as an iteration
variable. New variables are introduced that depend only
weakly on any primitive variable. The number of itera-
tion loops is reduced from three to two, with the outer
loop having only N +6 variables in the most general
case, and the inner loop only one. In addition, the need
to make a distinction between wide- and narrow-boiling
systems is entirely avoided.
In describing the ne4 method, the new variables will
tirst be introduced, the rearranged forms of the equations
will then be discussed, and finally a stepwise outline of
the algorithm will be presented.
The main variables of the outer iteration loop are
referred to as ‘volatility parameters’, and are defined as
ui = In (KiI&)
where Kb is a reference equilibrium ratio defined as a
weighted average of the form:
In Kb z 2 WiIn Ki
where the w’s are weighting factors, expressions for
which are derived in the Appendix.
The u’s are clearly similar to the a’s of the preceding
section, with the role of K, now being played by Kb. The
dependence of & on T is then represented by a model
113
similar to that of Eq. (12) for Kr:
1
?‘i > (1%
where T* is a reference temperature.
Just as the K’s are strongly dependent on T, and for a
highly nonideal system, on x and y as well, Kb will also
exhibit a strong dependence on these primitive variables.
However, the coefficients A and B willbe only weakly
dependent on T. Furthermore, since the T dependence
of the K’s is usually not strongly affected by composi-
tion, A and B willalso be only weakly dependent on x
and y. As a result of these weak dependencies, A and B
are excellent iteration variables to include along with u in
< ~3
the variable set of the outer loop.
The values of the coefficients A and B are determined
by evaluating the K’s at two temperature levels, T and
T’, while holding x and y constant. The corresponding
values K,, and Kb are obtained from the defining relation
Eq. (14), and A and B are given by:
B = ln (KblKb)
1 1 (16)
---
T’ T
A=lnK*-B
1 1
~-?;i . (17)
( >
It should be noted that, while the u’s are nearly
independent of T, the distribution of the u’s is influenced
by composition to essentially the same extent as the K’s.
The algorithm used to drive the u’s to convergence in the
outer loop should develop some information about this
composition dependence, and use it to improve the con-
vergence characteristics. It is notable that, while the
value of an individual Ui may increase or decrease in
successive iterates, the weighted average B WiUimust
remain unchanged. This fact tends to increasi the stabil-
ity of the outer loop iteration algorithm.
It will next be shown how the describing equations
may be rearranged to eliminate T as an iteration vari-
able, so that a complete solution may be obtained by
converging a single-variable inner iteration loop for a
given set of values of A, B, and II. The component mass
balance and phase equilibrium equations are first written
in terms of component and phase flow rates,
li t Ui= fi (18)
vd V = Kb e”%lL (19
where K,, eui has replaced Ki. Next, the new variables R,
defined as .
(13)
&V
REKbVtK:L (20)
where Ka”is a reference val~.of Kb, and pi, dehed as
(14)
z-4__ (21)
Pi-l-R
are introduced. If Eqs. (18) and (19) are now combined
and the new variables introduced, on rearranging and
solving for pi, the result is:
114 J. F. BOSTON
and H. I. Ban-r

fl (22) IQ. (3), and the error in the enthalpy balance, Eq. (4),
Pi= l_R+&OReui'
can be determined. When the assumed value of R used
to start this sequence of calculations is adjusted to drive
If the constitutive equation the enthalpy balance error to zero, the results will
represent a complete, consistent solution of the describ-
ing equations for the given values of A, I3, and u.
It is notable that the variable R is a unique combina-
tion of temperature (through Kb) and phase ratio which
is also rewritten in terms of the component and phase both reduces the inner loop to a single variable con-
flow rates, and the new variables introduced, the vergence problem, and precludes the necessity of
remarkable result is that Kb may be. solved for directly in switching roles depending on whether the system is
terms of the p’s: wide- or narrow-boiling. As R changes for a wide-boiling
system, the calculated L and V will remain nearly
constant, but the calculated T will be relatively sensitive.
In effect, then, the enthalpy balance error will be driven
Kt, = (23) to zero by the change in temperature. For a narrow-
7 e%i ’ boiling system, on the other hand, the calculated T will
be relatively constant, but the calculated L and V will
It is easily shown that this expression reduces properly respond almost linearly to changes in R. In this case, the
when R takes on its extreme values. For R = 0, the result phase ratio will dominate in driving the enthalpy balance
is: error to zero. For systems that are neither wide- nor
narrow-boiling, the relative dominance of the tempera-
ture and phase ratio will change automatically depending
on the nature of the system.
In our implementation of the new algorithm, a second
order procedure is employed to drive the R loop to
convergence. Using information retained from suc-
which correctly gives the value of Kb corresponding to cessive iterations of the outer loop, the R loop is usually
the bubble point of the feed. For R = 1, the expression satisfactorily converged in two or three iterations. The
reduces to: purpose of the reference quantity Kbo in the defining
relation for R is merely to avoid numerical difficulties
when Kb is either very large or very small. The value of
Kt,’ is set equal to Kb at the beginning of the R loop, and
held constant until convergence is achieved.
In the inner loop calculation sequence described
above, the liquid and vapor enthalpies are required for
each iteration. When the actual enthalpy models are
which corresponds to the dew point of the feed.
simple functions of temperature, such as polynomials of
The new variable R becomes the single iteration vari-
low degree, the inner loop computational load is not
able of the inner iteration loop. For a given value of R,
increased substantially. However, enthalpy calculations
the p’s can be calculated from IQ. (22), Kb can be
can be costly when complicated models are used, such as
calculated directly from Eq. (23), and T may be cal-
those based on equations of state or corresponding states
culated from the Kb model Eq. (15):
correlations. In order to reduce the computational load in
these cases, the approach that has been adopted is to use
1
-_=-+ 1 InKb-A
(24 very simple models to represent the enthalpies inside the
TP B’ inner loop, and to update the few parameters on which
they depend using the actual enthalpy models in the
The liquid and vapor mole fractions are available from outer loop. These parameters then become members of
the p’s: the set of iteration variables. This is in direct analogy to
the use of the simple model, Eq. (15), for Kb in the inner
loop. In keeping with the spirit of the new method, the
(25)
simple models are formulated so that the dependence of
these new variables on the primitive variables is either
weak or relatively insignificant compared with other
euipl
Yi = (26) effects.
I$Ie9Pj' Several auxiliary enthalpy quantities are introduced to
facilitate the formulation of the simple models. However,
it is worth noting that the only quantities that must be
Using these values of T, x and y, the liquid and vapor provived by the actual enthalpy models are the mixture.
enthaipies HL and Hv can be evaluated. By summing values HL and Hv and, except in certain cases, the
JZq. (21) over components, the phase rate L may be component ideal gas enthalpies. Neither partial molar
obtained: enthalpies nor other partial derivatives are required.
Thus, the new algorithm imposes no requirements on the
L = (1 -R) 7 pi. actual enthalpy models that might be dit5cult to satisfy
(n)
or reduce its generality.
It should be recognized that the use of actual enthalpy
V can then be calculated from the total mass balance, models consisting of simple functions of temperature
 A radically different formulation and solution of the single-stage flash problem
alone is a common practice. When this is the case, the
actual models are used inside the inner loop instead of
the simple models to be proposed. The details of this
case will be discussed after the more general case is
considered.
The following defining relationships form the basis of
the approach to be described:
AH” = Hv - Hv’ (28)
AH, = HL - H," (29)
where Hv(y, T, P) and HL(x, T, P) are the actual
mixture enthalpies. AH” and AH‘ are vapor and liquid
enthalpy departure functions, and Hv” and HLo are vapor
and liquid ideal gas enthalpies:
Hv’= YiHoV,
(30)
T:
H/=z xiHoyi (31)
1 T,= T/T,
where H‘& is a component ideal gas enthalpy.
The two departure functions, AH” and AH,, are the
quantities for which the simple enthalpy models are
used. Before introducing these models, it will be useful
to consider the roles of the various enthalpy quantities in
the enthalpy balance equation.
The form of the enthalpy balance equation used with
the new method is based on the ‘constant composition’
approach .described by HoUand[‘l], in which the
component mass balance equation is substituted into the
enthalpy balance equation to eliminate either the vapor
or liquid mole fractions. Using primitive Bqs. (2)-(4), and
the defining relations (28)-(31), the following result is
obtained when the vapor mole fractions and the vapor
phase rate are eliminated:
‘I’= (L/F)(AH, - AHL) - T ziH0, - AHv + HF + q = 0.
(32)
This is the form of the error function used inside the R
loop.
An important feature of the form of this function is
that the quantity AH, in the tist term is ordinarily large
in comparison with AH,, since it includes the vaporiza-
tion enthalpy of the liquid. The conventional form, which
is derived by simply combining the enthalpy and total
mass balance equations to eliminate one of the phase
rates, does not have this property. For example, when V
is eliminated the result is:
‘l’=(L/F’)(Hv-HL)-Hv+HF+q. (33)
Since Hv and HL depend on the compositions of
different phases, it is possible for their difference to be
close to zero. This tends to reduce the sensitivity of \y to
changes in the calculated value of L.
The quantities that are required to use I$. (32) are
AH”, AH= and H’&. Consider first the vapor quantities
H’$ and AH”. The component idealgas enthalpy H’$
can be represented adequately by a polynomial in T, and
may be calculated inside the inner loop without adding
much to the computational load. The enthalpy departure
function AH, depends strongly on temperature, pres-
sure, and, in some cases, composition as well. The
115
authors have found that this dependence can be
represented very well over limited ranges of T, P and y
by a model of the form
AH = Cp’ ZWTP.’
V (34)
PC_
where T, and TAP, are mixture pseudo-critical quan-
tities, calculated by the following mixing rules:
i”c = F YiTci (35)
TJP, = 7 yi(T$‘,,). (36)
The form of ELq.(34) was determined from consideration
of various enthalpy departure expressions based on
equations of state and corresponding states correlations.
The reduced temperatures are given by
(37)
T? = T”/T, (38)
where T* is a reference temperature that may be the
same as the one used in the Kb model. The coefficients C
and D are only weakly dependent on T, P and y; hence,
they are analogous to A and B in the K,, model, and are
logical choices of iteration variables to be included with
A, B and II in the outer loop. For given values of C and
4 T,, TAP,, T, and Tt are recalculated as T (or P) and
y change inside the R loop. The simple model of Eq. (34),
then, provides a value of AHv that reflects its depen-
dence on these primitive variables.
The values of C and D are updated in the outer loop
using values of Hv obtained from the actual enthalpy
model and the ideal gas enthalpies H’& at two tempera-
ture levels, T and T’. Two corresponding values of the
departure function, AH” and AH:, are calculated using
the defining relation Eq. (28).
For the liquid, a similar approach is used but a much
simpler model is employed. The dependence of A& on
temperature and composition is represented by the
model
AH,=E+F(T-T*) (39)
where T* is the same reference temperature as that used
for AHv. The coefficients E and F are analogous to C
and D in the AHv model, tid are treated in the same
way, being held constant while the R loop converges.
They become iteration variables of the outer loop, where
they are updated using values of HL obtained from the
actual model at two temperature levels, T and T’. The
component ideal gas enthalpies that are used tb klculate
tie corresponding two values of HLo are the same as
those used to update C and D. The values of AHL at T
and T’ are then calculated using the defining relation Eq.
(2%.
The glaring absence of a composition dependency in
the model for AH,_, analogous to that included in the
AHv model, deserves further comment. It would seem
possible to represent in a realistic way the ‘ideal
contribution to the liquid enthalpy departure by a
function of the form
TfL(1- TX)‘.=
116 J. F. BOSTONand H. I. BRIR

where TcL is a liquid pseudocritical temperature and Table 5. Type I adiabatic flash, the new algorithm
r,L = lYT,. The temperature and composition depen-
dence of this function are similar to those of a number of (1) Assume II, A, B, C, D, E, F
corresponding states correlations. If the behavior of the
(2) Assume R
actual liquid enthalpy model were realistic, its deviation
from that of the function given above should be rela- (3) Calculate p, Kb, T, x, J, L using Eqs. (22)-(27)
tively small, even for systems that have a strongly
composition dependent excess enthalpy. Unfortunately, (4) Calculate T,, TJP,, AH,, AHL using Eqs. (34)-(39)
it is sometimes found that the actual liquid enthalpy
model is in gross disagreement with the typical behavior, (5) Calculate Hv’
especially with regard to composition dependence. In
such cases, using a simple model with a typical composi- (6) Calculate Y using Eq. (32)
tion dependence can have a detrimental effect on con-
: (7) Assume new R and return to (3) until (YI< E
vergence characteristics. A model sulhciently compli-
cated to be universally suitable would defeat the purpose (8) Calculate V using Eq. (3)
of minimizing the inner loop computational load. Thus,
the simple model of Eq. (39), having no explicit (9) Calculate K, HL, Hv using actual models
composition dependence, was adopted. Clearly, then, the
coefficients E and F must absorb all the composition (10) Calculate i, a, 6, % (and 4, fi, P first iteration only)
dependence of AHL.
It should be noted that there is an implicit liquid 7 (11)Assume new values of II, A, C, E (and E, D, F first
enthalpy composition dependence in the inner loop iteration only) and return to (2) until calculated
values match assumedvalues
through the ideal gas enthalpy of the liquid phase. While
this is not apparent in the final form of Rq. (32), it in fact
arises in the term T ZJH“vi. responding vector of calculated values, then the problem
may be stated as the determination of U such that
When the actual vapor enthalpy model is a linear
combination of the form G(U)=U(U)--u=o. (41)

HV = 2 YiHvi
To facilitate the discussion of the iterative solution of
i this system, the following notation is introduced. Let k
be an iteration index, and SUk denote a direction of
where the component enthalpies Hvi are simple displacement away from Uk. Then, in general:
functions of T, such as a low degree polynomial, then no
excessive computational burden is imposed on the R Uk+’ = Uk + psUk (42)
loop by simply replacing the H$‘s withthe Hv’s. In this
case, A& is set to zero, and C and D are ignored in the where p is a scalar distance factor. The corresponding
outer loop. error function vector is given by:
The liquid enthalpy simple model case is analogous to
that for the vapor enthalpy. The value of F is set to zero, Gk+lIG(uk+l)2ek+I_uk+l.
(43)
and AH= becomes:
Note that, since each element of G is the difference
AHL=&Hq-H~o+E between the assumed value and the corresponding cal-
i culated value of the same variable, the problem is
amenable to treatment by direct iteration. Here, the
where the component enthalpies HLi are simple correction vector is simply
functions of T, and E is either zero or the composition
dependent excess enthalpy. p8Uk = G’ (44)
A stepwise outline of the new algorithm is given in
Table 5. For the sake of simplicity, Table 5 includes only and the assumed values for the next iteration are given
the complicated enthalpy model cases. by:

~OuTIIllLoop
CoNvmGmcE Uk+l=Uk+Gk=,, (45)
The problem of generating a new set of assumed
values of the iteration variables in step (11) of the new The advantage of direct iteration over other methods is
algorithm is considered next. Let the complete variable its simplicity. It almost always converges when applied
set be denoted as: to this problem, but’its rate of convergence is slow in
comparison with that of Newton type methods.
If the problem were solved by the Newton-Raphson
(N-R) method, the direction vector would be given by:

SU’ = EjkGk (46)

where m is the negative inverse of the Jacobian matrix of

G(U). Although the convergence characteristics of the
N-R method are favorable, a serious drawback is that
If U is a vector of assumed values and 6(U) a cor- the Jacobian matrix must be re-evaluated after every
 A radicallydifferentformulationand solution of the single-stageflash problem
iteration. If numerical differentiation were employed for
this, hJ+7 passes through steps (2)-(10) of Table 5
would be required just to evaluate the Jacobian matrix
one time for an iV component system. This would clearly
be prohibitive even for a system having only a few
components. Analytical diierentiation is also impracti-
cal, in view of the complexity of composition dependent
enthalpy departure and equilibrium ratio models.
The quasi-Newton (Q-N) methods offer an attractive
alternative to the N-R method. They possess conver-
gence characteristics similar to those of N-R, hut do not
require explicit reevaluation of the Jacobian matrix after
each iteration. Rather, starting with some initial estimate
of the Jacobian matrix, they directly update it or its
inverse afpr each iteration, using information consisting
only of U(U) and U from the current iteration. A
comprehensive review of Q-N methods was given by
Dennis & More[81. In the present work, the Q-N
method proposed by Broyden[6], with the identity matrix
as the initial B matrix, was selected for the solution of
Eq. (41).
The above considerations concerning N-R vs Q-N
methods apply to an arbitrary system of nonlinear equa-
tions. The full significance of the use of the Broyden
method for this algorithm is a result of the special
characteristics of G(U). The iteration variables U were
chosen to be as independent as possible of the primitive
variables and of each other. If these goals were met
completely the Jacobian matrix of G(U) would be the
negative identity matrix and convergence would be
achieved in a single iteration. For real problems the
Jacobian matrix will not differ greatly from the negative
identity matrix. In other words the Jacobian matrix will
be diagonally dominant with diagonal elements the order
of magnitude of unity. This result has two implications.
First, the identity matrix is a good, not arbitrary, choice
for the initial H matrix, which means that not even one
explicit evalmiion (or inversion) of the Jacobian matrix
is required. Second, the off-diagonal elements of the
Broyden update should be small and not greatly affect
the new estimates of U. The net result of these two
considerations is that convergence is usually achieved in
fewer iterations than would be required for evaluation of
the initial Jacobian matrix by llnite difference.
The strongest interactions (largest off-diagonal ele-
ments) occur in the case of .highly nonideal systems. This
is because no attempt was made to eliminate composition
dependence among the volatility parameters, II. While
the absence of composition dependence in the simple
model for K may seem to be a violation of the principle
of the iteration variable independence, it actually
represents a trade-off between the effects that the simple
models ought to account for and the effects that the
Broyden convergence procedure ought to account for.
Since composition dependence is seldom so great as to
negate diagonal dominance, and since the representation
of composition effects by a simple model would require
N additional K evaluations per iteration, composition
dependence is most efficiently accounted for by the
convergence procedure.
By now it should be apparent that the desirable con-
vergence properties of the new algorithm result from a
matching of the characteristics of iteration variables and
the Broyden method. Another benefit obtains when the
tlash calculation is restarted from a previously converged
solution for diflerent values of z and Q. This situation
occurs during the convergence of flow-sheet simulation
117
problems. The iteration variables, to the extent they are
independent of primitive variables, do not change, while
the correlation among the Ui'sis accounted for in the
most part by the H matrix saved from the previous
convergence. As a result convergence for the new
conditions is achieved in a few, frequently one, itera-
tions.
The convergence characteristics described above have
been veritied by observation over approximately a one
year period at Union Carbide.
Some details of the implementation are considered
next. In the outer loop of the new algorithm, all of the
iteration variables may be updated using the final T, x
and y resulting from the converged R loop. However, the
correctness of the final solution requires only that a and
one each of the pairs (A, B), (C, D), (E FL be updated,
and that the calculated values of these variables match
the assumed values. This is not required of the remaining
three variables, whose values may be tixed.
The significance of this observation may be explained
by considering the variable B. When B is updated, two
evaluations of K are required, whereas if only II and A
are updated, a single evaluation is suflicient. Since B is
much less temperature and composition dependent than
u and A, little is lost by not updating J3, and much is
gained in reduction of the computational load associated
with equilibrium ratio evaluation.
Similar remarks may be made with regard to the
updating of D and F, with corresponding implications in
the reduction of the required number of vapor and liquid
enthalpy evaluations.
In our implementation, the following strategy is adop-
ted. Initial values of all of the outer loop iteration vari-
ables are generated as follows. The K’s are first evalu-
ated at an assumed temperature assuming no composi-
tion dependence in either the vapor or the liquid phase.
Using these ideal K’s, Eq. (8) is solved for /3, and x and y
are calculated using Eqs. (6) and (7). The u’s are cal-
culated from the K’s using the defining relationship Eq.
(13), and A, B, C, D, E and F are calculated using the
K’s and vapor and liquid enthalpies evaluated at two
temperature levels, as discussed previously. Since all of
the iteration variables are largely independent of
temperature, their values are affected little by the choice
of the assumed temperature. The values may, however,
be affected significantly by the ideal K assumption. Using
these initial values, the R loop is converged and the
iteration variables updated using the actual, rather than
ideal, K model. At this point the values of B, 4 and F
are tixed and the iteration variable set redefined as:
rui
u=
11
“c
E
as indicated in steps (10) and (11) of Table 5.
Implementation of the Broyden method requires a
procedure for determining the (scalar) distance factor p
in 4. (42). We do not invoke norm reduction but instead
use p = 1 each iteration as suggested by Broydent91.
While the previous discussion has been related to the
Type I llash, the other five ilash types listed in Table 1
may he accomodated readily by simple mod%cations to
the algorithm. The required changes in the basic.al-
118 J. F. BOSTONand H. I. BRETT

gorithm outlined inTable 5 for the Type I flash will now (4) Assumenew R and return to (3)until (i- L(< t.
be discussed for each of the other types. (5) CalculateKb, T using Eqs. (23)and (24).
Flash Type II requires a ZL model in terms of P
rather than T:
’ The bubble point and dew point special cases require
In (KbP) = A + BP In (P/P*) (W no R loop iterations since B = 0 implies R = 0 and /3 = 1
implies R = 1.
where P* is a reference pressure. The primary pressure In Type VI there is no temperature or pressure varia-
dependence derives directly from the fact that, for any tion and hence no Kb model is needed. The discussion of
individual equilibrium ratio K,, the product KIP is nearly Types IV and V applies with the exceptions that Kb is
independent of P. set to unity and as a result the outer loop iteration
The secondary pressure dependence [BP In (P/P*)] is variables become simply the equilibrium ratios, K, while
necessary for systems in which this simple pressure R reduces to /3.
dependence does not hold, e.g. systems in which large
amounts of supercritical components are present in the
liquid phase. With this KS model, flash Type II can be Type VI
handled simply by solving Eq. (48) for P instead of Eq.
(2) AssumeR
(15) for T in step (3) of Table 5. No other changes are (3) Calculatep. & using Eqs. (22) and (23)
required. The vapor enthalpy departure model, Eq. (34), (4) Assume new R and return to (3)until I&, - l( <c
already includes pressure dependence while no pressure
dependence is necessq for liquid enthalpy.
Flash Type III reqmres that the R loop be split into It should be noted that the R loop in this case is
two consecutive loops, each of which depends on a equivalent to the solution of I%. (8) for B given K.
single variable. The K,, model for this flash type is a
function of both T and P and is a combination of Eqs.
(15) and (48):
The performance of the new algorithm will be illus-
trated using several example problems based on four
ln(KbP)=A+& T-~*
’ L + BP ln(P/P*). (49) different systems. The first two systems were chosen to
( >
demonstrate the convergence characteristics in extreme
Steps (2)-(8) of Table 5 should be replaced with the cases of wide- and narrow-boiling systems. The third and
following modified procedure: fourth illustrate the performance for moderately and
highly nonideal ternary systems.
(2a) CalculateL using J?q.(3) Example system 1 is a wide-boiling binary system
(2b) AssumeR consisting of ethane and n-nonadecane. The difference
(3) Calculatep, i using Eqs. (Z?),(27) between the dew point and bubble point of a mixture
(4) Assumenew R and return to (3)until Ii - L(< P, containing equal parts of the two components is 508°F. A
(5) Calculate4,I, Jusing 4s.(2%(2%(26) generalized, reduced form of the BWR equation of state
(6) Assume T
(7a) CalculateP using Eq. (49) due to Starling and Han[lOl was used as the equilibrium
(7b) Sameas steps (4)-(a) of Table 5 ‘ratio and vapor and liquid enthalpy model. The feed
(8) Assumenew T and return to (7al until pj c l2. contained 5Omole per cent of each component in all
cases.
Steps (9)-( 11) of Table 5 remain unchanged. In Example l(a) the feed was a saturated vapor-liquid
In Types IV, and V the enthalpy balance is employed mixture at 6OWRand 1 atm created by performing a Type
only to calculate q. Hence, steps (4)-(S) of Table 5 may VI flash and mixing the products. Sufficient heat was
be omitted, only K need be calculated in step (9), and II, then added to give a temperature increase in a Type I
A and B alone constitute the variable set of the outer flash of 228.3”F. The folloking results from the R loop
loop. After the outer loop is converged, q may then be for the first iteration of the outside loop show that the
calculated using Eq. (9), after HL and HV are evaluated. temperature dominates in driving the enthalpy balance
It should also be noted that x and y need not be cal- error to zero, with L remaining nearly constant. This
culated until after the R loop is converged. The minor agrees with the expected behavior for a wide-boiling
differences in step (3) and in the error function used to system.
drive the R loop to convergence are indicated in the
following outlines:

Type IV R L/F VW

(2a)Calculate L using l!?q.(3) 0.4940 0.5060 600.0

(2b) AssumeR 0.5940 0.5040 631.9
(3) Calculatep, t using Eqs. (22)and (27) 0.8985 0.5oa7 891.6
(4) AssumenewRandretumto(3)untilli-YcE 0.9746 0.5001 1323.4
(5) CalculateK,,,P using Eqs. (23)and (48). 0.9175 0.5005 939.3
0.924M 0.5005 948.0

(2a) calculate L using Eq. (3)

(2b) AssumeR The outside loop converged in a total of six iterations,
(3) calculate p, i using Eqs. (22)and (27) with the following iteration history:
 A radically
differentformulationaad solutionof the single-stageflashproblem 119

The feed stream in Example 2(a) was a saturated

Iteration vapor-liquid mixture at 701.2”R and 1.75atm, having a
number UF Wg) U1 u2 liquid fraction of approximately 0.76. Sufficient heat was
then added in a Type I flash to decrease the liquid
1 0.5005 947.99 0.1287 -6.290
2 0.3956 806.34 0.1630 -7.965 fraction to about 0.21. The results from the R loop of the
3 0.4885 828.86 0.1580 - 7.723 iirst outside loop iteration are tabulated below. They
4 0.4852 828.57 0.1580 -7.725 show that, as expected, the temperature remains nearly
5 0.4831 828.34 0.1581 -7.726 constant as R changes, while L changes almost linearly
6 0.4827 828.31 0.1581 -7.727 with R. Convergence of the outside loop was achieved in
only two iterations, with no signilicant change in L or T
after the lirst iteration. The liquid phase activity
Examples l(b) and l(c) are Type V flashes with rela- coefficients in the final solution ranged from 0.590 to
tively poor initial temperature estimates, and l(d) is the 3.15 x lo”, reflecting the strong nonideality of hydro-
Type VI flash used to generate the feed for Example carbon-water systems. It is worth noting that a con-
l(a). The following tables summarize the iteration ventional flash algorithm would fail on this problem unless
histories for these examples: T were taken as the iteration variable of the inner loop,
and p as that of the outer loop, as indicated in Table 2
Example l(b): Specified V/F=O, initial T =6OW’R for the ‘narrow-boiling’ case.

Iteration history:
R UF T("N
0.5 0.5002 701.23
Iteration 0.6 0.4002 701.23
number T("R) k 0.7932 o.m70 701.25

I 503.18 -12.55
2 480.45 - 12.03 Examples 2(b) and 2(c) are Type V bubble point and
3 481.90 -l2.07 dew point flashes, respectively, with relatively poor in-
4 481.74 -12.07 itial temperature estimates. Summaries of the iteration
histories are tabulated below. Components 1, 4 and 7
were hydrocarbons, and 13 was water.
Example l(c): Specified V/F = 1, initial. T = WR Example 2(b): Specified V/F = 0, initial T = 585.71“R
Iteration history: Iteration history:

Iteration Iteration
number TCR) UI number Wg) MI u7 U13
1 986.69 5.885 1 697.% 7.193 14.17 -9.840
2 989.44 5.857 676.30 7.456 14.34 - 10.44
3 989.43 5.857 : 671.06 7.520 14.34 - 10.44
4 669.38 7.540 14.39 - 10.65
5 669.24 7.542 14.40 - 10.65
Example l(d): Specitied T = 6OtPR
Example 2(c): Specilied V/F = 1, initial T = 585.71”R
Iteration history:
Iteration history:
Iteration
number UF UI u2
Iteration
1 0.5060 2.451 -8.510 number WR) UI ~4 Ul3

2 0.5471 2.558 -8.516

3 0.5419 2.545 -8.515 1 673.26 4.299 7.780 -6.768
4 0.5424 2.547 -8.515 721.70 5.759 8.967 - 6.499
: 718.73 6.391 9.679 -6.600
4 717.83 6.092 9.343 -6.590
5 717.98 6.193 9.455 -6.5%
Example system 2 is a muhicomponent system consis- 9.451 -6.597
6 717.92 6.189
ting of water and 12 assorted hydrocarbons. While in a 7 717.90 6.188 9.451 - 6.597
strict sense it is not a narrow-boiling system, it behaves
like an extremely narrow-boilii system because the sum
of the mole fractions of the hydrocarbons in the feed is
only 3.6 x lo-‘. The NRTL equation of Renon [ 1l] was Example 2(c) is particularly interesting because of the
used as the liquid phase activity coefficient model, the extremely strong liquid composition dependence of the
PrausnitzChueh[l21 modification of the Redlich-Kwong equiWrhlm ratios, and the resultillg large changes in
equation was used for the vapor fugacities and enthalpy, liquid composition as the outer loop converges. It shows
and the liquid enthalpy model was a polynomial in T. that the Broyden quasi-Newton algorithm is capable of
120 I. F. BOSTONand H. I. BRITT

building up information about the composition depen-

dency of the u’s in the updating of the II matrix. As an R 4F TCR)
indication of the magnitude of the compo&ion effect, the
range of values of activity coelIicients in the final solu- 0.0 1.000 639.5
0.1 0.9972 639.5
tion of 2(c) was 0.749-1756. In Example 2(b) the final 0.775 0.9217 640.0
range was 0.535-3.56~ 106, but the composition effect 0.9438 1.7128 641.5
was less important since the liquid composition was 0.8172 0.9009 640.2
constant. 0.9269 0.7653 641.1
The third example system is a moderately nonideal
ternary system containing ethanol (1) iso-octane (2), and
benzene (3). As an indication of the magnitude of the
The following table summarizes the outer loop iteration
history:
liquid mixture nonideality, the infinite dilution activity
coefficients are given at 300 K:
Iteration
Component Infinite dilution number 4F T(a) UI u2 u3

paid activity coefficient

-0.4101 0.1652
1 0.7653 640.57 1.400
2 0.6618 627.85 1.933 -0.4652 0.1568
192 95.1
3 0.6500 627.47 2.177 -0.4881 0.1514
271 5.08
4 0.6472 626.67 2.326 -0.5017 0.1479
1.3 18.3
5 0.6393 626.72 2.362 -0.5057 0.1475
391 3.03
6 0.6387 626.66 2.371 -0.5066 0.1474
293 2.51
7 0.6388 626.67 2.369 - 0.5065 0.1474
392 2.12

tThe component corresponding to the first index given is
infinitely dilute in the component corresponding to the second
index.
These values indicate that the strongest nonideality oc-
curs when ethanol is present in relatively low concen-
tration. Thus, the composition of the feed for this
example was taken to be 0.1 mole fraction ethanol, with
equal parts of the other two components.
The liquid phase activity coefficients were modelled by
the Wilson equationil31, the vapor phase was assumed
to be ideal with respect to both fugacities and enthalpies,
and the Watson pure component vaporization enthalpy
equation was used for ideal liquid enthaipies. Excess
liquid enthalpies were obtained from the Wilson equa-
tion.
Example system 4 is a highly nonideal ternary system
containing chloroform, water and acetone. The
UNIQUAC [ 141equation was used to model liquid phase
activity coefficient, and the vapor phase was assumed to be
ideal. The vapor enthalpy was expressed as a polynomial in
temperature, and the liquid enthalpy was obtained using
the Watson vaporization enthalpy equation for each com-
ponent.
Three Type I flash examples were constructed, each
using a feed created by performing a Type V flash at
1 atm, with a specified V/F of 0.4, mixing the products
and adding varying amounts of heat. The feed composi-
tion was 0.4 mole fraction each of chloroform and water,
and 0.2 mole fraction acetone. In all three examples, the
.initial temperature estimate was 450 K. A few of the final
results are tabulated be1ow.S

Number Final activity coefficients

Example iterations V/F T(K) Chloroform Water Acetone

2; 11
7 0.4000
0.6623 329.5
324.7 3.16
168 2.38
1.02 0.663
5.59
4(c) 10 0.8110 340.2 413 I.00 7.80

A Type I gash was performed, with a feed stream that The following table is an iteration history for Example
was created by performing a Type VI flash at 626.7”R 4(c), the most nonideal of the three:
and 1 atm, which resulted in a liquid fraction of 0.6388,
and mixing the product streams together. The initial
temperature provided to the Type I flash was 491.71, or
Iteration
135”less than the correct value. number 4F T(g) ~1 u2 u3
The tabulation below of the iteration summary for the
tirst Pass through the R loop shows that the temperature I 0.2185 332.29 0.6210 -0.3862 -0.0197
changed very quickly from its rather poor initial value of 2 0.2107 335.82 1.290 -0.6486 -0.2239
491.7”R.This behavior reflects the fact that the system is 3 0.2058 336.69 1.583 -1.112 0.1050
relatively narrow-boiling, and the enthalpy balance error 4 0.1974 338.99 2.594 - 1.856 0.2124
is therefore being driven to zero mainly by the change in 5 0.2393 370.65 4.336 -3.111 0.3672
liquid fraction. 6 0.1880 338.26 3.769 -2.861 0.5072
7 0.1880 339.36 4.184 -3.135 0.5143
8 0.1891 340.38 4.330 - 3.225 0.5082
k!Msequcnt analysis showed that the liquid phases resulting 9 0.1890 340.29 4.295 -3.204 0.5096
from Exampks 4fa) snd 4(b) were unstabk, splitting into two IO 0.1890 340.24 4.294 -3.201 0.5095
coexisting liquid phases.
 A radicallydiierent formulation
and solution of the single-stage flash problem 121

The example problems demonstrate the rapid NcwmucL4TlJRB

temperature convergence characteristics of the new al- Kb model coefficients
gorithm. By far most of the temperaturechange from the 2; AH, model coefficients
initial estimate typically occurs during the first two or iF AH‘ model coe&iits
three iterations. The following table illustrates this F total molar flow rate of feed
component molar flow iate in feed
behavior in terms of the temperature errors initially and
after each of the tirst three iterations:
A iteration variable error function
H negative inverse’of Jacobian iteration matrix
HF feed enthalpy
Magnitude
of temperature
error(“F) HL liquid enthalpy
HLOideal gas enthalpy of liquid phase
Initial First Second Third AH, liquid enthalpy departure function
Example estimate iteration iteration iteration Hv vapor enthalpy
Hv” ideal gas enthalpy of vapor phase
l(a) 228.3 119.7 22.0 0.6 A& vapor enthalpy departure function
I(b) 118.3 21.4 1.3 0.2 K vapor-liquid equilibrium ratio
l(c) 89.4 2.7 0.0 0.0 weighted average equilibrium ratio
2(b) 83.5
132.2
28.7
44.6
7.1
3.8
1.8
0.8
2L reference value of Kb
total molarflow rate of liquid phase
Yc’ 135.0 13.9 1.2 0.8 1 component molar Bow rate in liquid phase
216.9 0.7 4.9 3.0 N number of components
2:; 225.5 14.3 16.9 6.4 P absolute system pressure
4(c) 197.6 10.5 8.0 16.0 PC pseudocritical pressure
modifiedcomponentmolarflow rate

In Examples 4(a)-(c) the apparently erratic behavior,

EJ
rate of flow of heat into flash, excluding feed stream
(positive for heat in)
after the large error reduction of the fist iteration, may Q/F
be attributed to the temperature dependence of the ii
iterationvariablefor inner loop
T absolutesystem temperature
activity coefficients as predicted by the UNIQUAC
equation. In such a highly nonideal system, these TC pseudocriticaltemperahue
temperature effects would be expected to dominate the T, pseudoreducedtemperature
T* referencetemperature
temperature convergence behavior while the Broyden V outer loop iterationvariableset
algorithm is building up .information about composition volatilityparameter
effects and the composition is still changing substantially. ; total molarflow rate of vaporphase
In order to illustrate the convergence behavior of the u componentmolarflow rate in vaporphase
volatility parameters, the component having the most w Kb model weightingfactor
widely varying II was selected for each example. The x componentmole fractionin liquidphase
magnitudes of the errors in u for the selected Y componentmole fractionin vaporphase
z componentmole fractionin feed
components after the first, second, third and sixth itera-
tions are tabulated below: Greek symbols
(I relativevolatility
Magnitude
of errorin II B V/F
CD constitutive function
second Third Sixth 0 enthalpy balance function
Fiit
Example Component iteration iteration iteration iteration p scalar distance factor
c convergence tolerance

l(a) 2 1.44 0.24 0.00 Subscripts

l(b) 2 0.48 0.04 0.00 i, j component indices
l(c) 1 0.03 0.00 0.00 r reference component
l(d) l 0.10 0.01 0.00
0.81 0.21 0.21 Superscripts
7;C ‘: 1.89 0.43 0.20 k inter&onindex
1 0.97 0.44 0.19 A calculated quantity
&) t 0.36 0.19 0.12 0:oO
2.85 2.48 2.30 0.03
?!C I 3.67 3.00 2.71 0.53 1. J. F. Boston & S. L. Sullivan, Jr., Can. 1. Gem. Engng 52,52
(1974).
2. H. H. Rachford, Jr. & J. D. Rice, J. Petrol. Te&tol: 4(10)
(19.52).
3. J. R. Friday & B. D. Smith, A.Z.Ch,E.Jl 10,698 (1964).
The table shows that the II errors tended tu decrease
4. H. C. Kliesch, Ph.D. Dissertation, Tulane University, New
rapidly over the first three iterations for all except the Orleans, Louisiana (1967).
two most highly nonideal examples, 4(b) and (c). In the 5. C. M. Crowe B M. Nishio, A.J.CAEfi 1X3), 528 (1975).
latter two examples, however, the progress over the next 6. C. G. Broydcn,MathsComp. 19,577 (l%S).
three iterations was remarkable, as shown by the tabu- 7. C. D. Holland, Multicomponent ZXstiUatiott, Prentice-Hall,
lated errors after the sixth iteration. Englewood Cls. NJ (1963).
8. J. E. Dennis, Jr. Bt J. J. More, SZAb4Reu.19(l), 46 (1977).
AcknowMgemmts-The authorswish to thankDr.J. S. Newell, 9. C. 0. Byyden, Comput. Jl 12,94 (1970).
of Union Carbide Corporation and Mr. R. L. Four&r, MS. IO. K. E. Starl& & M. S. Han, Hydrocarli. Z+ucess 51(5), 129
candidate at the University of Toledo, for providing assistance (1972)and Sl(6). 107(1972).
on the example problems. 11. H. Renon C J. M. Prausnitz, A.Z.Ch.i?.II14. 135(1968).
122 J. F. BWON and H. I. BRETT

12. P. L. Chueh & J. M. Prausnitx, KEC Fundamentals 6, 492

(1%7).
13. G. M. Wilson, J. Am. Chem. Sot. 86,127 (1964).
14. D. S. Abrams &J. hi. Prausnitx, A.J.CkE.R 21, 116(1975).
wi =i

i’i
I
(IlO)

where

rig&. (Ill)
Lkri’oationof Kb weightingcoejicient expressions
(I) Flash Types IV and V. Consider a Type V Rash in which the For a Type IV flash a sin&r derivation with P replacing T
K’s are independent of composition. The temperature must everywhere leads to precisely the same expression for wi.
satisfy the condition (II) Flash Types I, II and III. Consider a Type I Bash m which
the K’s are independent of composition. For relatively narrow-
a(T)=0 (11) boiling systems, the distribution of values of the Kb weighting
coefficients is relatively unimportant. For a relatively wide-boil-
where ing system, the temperature is controlled largely by the enthalpy
balance
(12) ‘W,B)=O (III)

where
Now introduce the variables Ui and Ks through their defining
relations
~(T,B)~BH~+(~-B)HL-HF-~ (112)
Ui= In (K/K,) (13)
HV = F #vi (113)
and
HL = x XiH,_, (114)
i
In Kb s z wiIn Kf (14)
I and where I$, and fib are partial molar vapor and liquid enthal-
pies, respecttvely. The liquid and vapor mole fractions may be
where the w’s are weighting coefficients constrained by:
expressed in terms of the feed mole fractions, /3 and the K’s as
follows:
T wi=l.
Kiri
yi=l-,!W/3Ki (ID)
Next define the function d as:
xi=+ (1%)
(15)
and the partial molar enthalpies are related by:
in which the u’s are taken to be independent of temperature. Let
T* be a base temperature where &,-&=KTZ !$ . cm)
( >P.X.?
K,,(T*)e”’ = Ki(T*) (16)
Let &T,B) be defined the same as $(T,B) except that Ki is
so that replaced by Kb e”’ everywhere in expressions (ID), (II6) and
(In), and the u’s are taken to be independent of temperature.
&T*)=$(T*). (IS) Following a line of reasoning similar to that under (I) above,
the condition
The ideal choice of the. weighting coefficients would be that set
for which the value of
@,,I,.= @,lr. (II@
AT=T-T*
is adopted for determining the weighting coefficients. DiSeren-
required to satisfy Eq. (ll) is the same as that which would result tiating, substituting into (IIS) and simplifying yields:
in

4,(T) = 0. (17) ( ~‘;I”-‘l:~~~)~(r.=~~d_l;f~~~(~)p(r..

The simplest analytical approximation to this ideal condition is (119)

its local equivalent: Hence it is clear that

(18) wi =‘i (1110)

7 ti
which is now adopted for determuting the weighting coelkients.
Differentiating (l2) and (U) with respect to T, substitutit~into where
(IS), and simplifying leads to:
yid In K&VT
(IIll)
‘I- l-Bt@Ki’

Sin&r derivations for Types II and III lead to expressions for Wi

From (14)it is then clear that precisely the same as (1110).