Documente Academic
Documente Profesional
Documente Cultură
a r t i c l e i n f o a b s t r a c t
Article history: Milling is a required pre-treatment for the use of biomass as a pulverized solid biofuel in some thermochemical
Received 26 March 2014 technologies such as combustion, gasification and bioethanol production, as well as in densification processes.
Received in revised form 4 June 2014 The particle size plays a key role on these energy conversion technologies. Experimental tests for poplar and
Accepted 7 June 2014
corn stover were performed to obtain pulverized material at different physical conditions of the biomass
Available online 28 June 2014
(input particle size and moisture content) and operational parameters (opening sizes of the screen and angular
Keywords:
speed of hammers). Fourteen parameters related to size central trends, dispersion and shape of particle size dis-
Milling tribution (PSD) were calculated and analyzed by a novel data post-processing methodology, combining Artificial
Poplar Neural Networks and statistical analysis. Results show that the characteristic size of the product (geometric mean
Corn stover size) is mainly influenced by the classification of the screen with values from five to eight times lower than their
Particle size distribution openings size. The angular speed of the hammer governs the variability and dispersion of sizes. The higher the
Neural network angular speed, the lower the dispersion on particle size. Physical conditions of the biomass present a negligible
effect on PSD.
© 2014 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.fuproc.2014.06.016
0378-3820/© 2014 Elsevier B.V. All rights reserved.
M. Gil, I. Arauzo / Fuel Processing Technology 127 (2014) 80–87 81
experimental test with all the possible variable combinations, ANN [36, can be read that the input particle size (within the analyzed range)
37] provide a suitable model, able to reproduce complex multivariate has no relevant effect on the particle size of the milled poplar. In addi-
non-linear relationships even from noisy and incomplete data and, at tion, Gil et al. [29] showed that this variable had also no effects on ener-
the same time, they are conceptually simple, computationally efficient, gy consumption, drying of biomass during milling or some handling
and well supported by professional technical computing software properties of the milled poplar. Therefore this variable was not consid-
environments. ered on the test campaign for corn stover in order to decrease the num-
Relevant independent variables such as hammermill operational ber of test from 32 to 16 and reduce the experimental campaign costs.
parameters and physical conditions of biomass (Section 2.4) are the in- On the other hand, three new central point tests for ANOVA analysis
puts of ANN models and the fourteen PSD parameters are the outputs were required. In central point tests, the variables have the medium
(Section 2.2). The individual architecture of neural networks is both value within their mentioned ranges: wH2 O ≈18%, mixture input particle
simple and regular, aiming to be computationally efficient and to size at 50% on mass basis, dtarget = 3.5mm and rev = 2500 rpm. As a
preserve the ability to generalize beyond training data: they are feed- result, 54 experimental tests (32 + 3 for SRF poplar and 16 + 3 for
forward networks, with sigmoid neurons in the hidden layer and one lin- corn stover) were performed.
ear neuron in the output layer. They are trained using back-propagation Obviously, the predictive tool based on Artificial Neural Networks
(Levenberg–Marquardt) algorithms, which include an automated Bayes- and the statistical results have been demonstrated within the variable
ian regularization method [38], allowing for fast training and good gen- ranges (enumerated previously) and any extrapolation outside these
eralization. Feed-forward back-propagation algorithms consist of two ranges must require future works.
phases: training and testing [39]. In this study, two ANN were designed
and trained from experimental data for each biomass resource, one hav-
ing with five sigmoid neurons in the hidden layer for dgm, d50 and lRR as 3. Results
outputs, and another one having with ten sigmoid neurons in the hidden
layer for d10, Sgm, HSgm, LSgm, MRS, Iu, Cu, Cg, Skg, Kg and mRR as outputs. Size is the most widespread studied descriptor of pulverized
ANN models also allow illustrating the influences of inputs on output particles by means of its size-probability distribution (PSD). Size
variables by means of 3D predictive surface from ANN and super- and cumulative size distributions can be used for knowing the
imposed points from experimental tests (see Fig. 2 shows dgm as a func- mass percentage or the cumulative undersize mass percentage for
tion of the hi and dtarget). However, a systematic analysis is required to each size (Fig. 1), showing a log-normal behavior for comminuted
determine the effects of each input variable and their interactions on biomass.
each of the fourteen output parameters of PSD. For such propose, the To characterize the effects of input variables on the particle size,
standard response surface methodology (RSM) by central composite fourteen descriptor parameters of the size distribution were analyzed.
design (CCD) was selected to perform the statistical analysis with a con- Multivariate statistical analysis allows for clustering them (in most
fidence level of 95%. CCD consists of 2n full factorial runs with 2n axial cases R2 N 99%) in three groups in order to describe different character-
runs and n center runs, in which the variable levels must be precisely istics of the size distribution of milled biomass.
established and identified for statistical analysis by (− 1, 0, + 1) low,
medium and high levels, respectively. For this reason, experimental
data could not be analyzed directly by CCD due to the input moisture 3.1. Clustering of PSD parameters
content variability of the biomass. However, predictive model from
the ANN can generate the output results, at the desired levels of input The first group consists of parameters related to average sizes and
moisture content, in order to be further processed by CCD analysis. percentiles (dgm, d50 and d10) of the distribution. These parameters are
named as ‘size parameters’. Multivariate analysis shows very strong
2.4. Test planning correlation R2 99% between dgm and d50, and slightly lower with (d10)
effective size (R2 N 90%).
Experimental test campaign was carried out with a horizontal axis The second group is composed by parameters related to dispersion
hammer mill of 11 kW and 3000 rpm nominal angular speed. Input bio- or variety of particle sizes, named ‘dispersion parameters’. The most
mass rate was regulated by means of the linear velocity of the feeding widespread used is the geometric standard deviation (Sgm) (ISO
conveyor belt as a function of the current electric consumption of the 9276–2:2001), representing the dispersion between each particle size
mill, avoiding mill clogging (more details of the hammer mill and exper- and the geometric mean size. Sgm can be graphically divided into high
imental facility can be found in Gil et al. [40,13]). Tests were carried out graphic Sgm(HSgm = d84 / d50) and low graphic Sgm(LSgm = d50/d16), cor-
on stationary continuous regime in which around 15–20 kg of biomass responding to high-medium and medium-low fraction of the distribu-
were milled. tion, respectively. In addition, other dispersion parameters are used
In relation to the hammer mill, two operational variables may have mostly regarding to ratios between the medium-high and low fractions
important effects on the milled product quality: 1) the opening sizes of the distribution: mass relative span (MRs = (d90 − d10) / d50), unifor-
of the screen surrounding the mill chamber, as a key factor on particle mity index (Iu = d95 / d5), uniformity coefficient (Cu = d60 / d10), and co-
classification and 2) the hammer angular speed and therefore the im- efficient of gradation (Cg = d230 / (d60 − d10)). Multivariate analysis
pact energy between particle and hammer, as a key factor on particle shows strong correlations between them (R2 N 98%).
fracture. The other main factor group with repercussion on milling pro- The third group, denominated ‘shape parameters’ is related to
cess is the physical condition of biomass, mainly, its moisture content the distribution shape in comparison with the log-normal distribution
and particles size. regarding to the lack of symmetry (geometric skewness), as well as
Following, the four factors are enumerated: peakness or flatness of the distribution (geometric kurtosis).
Once parameter clustering was performed, neural networks are
• Biomass moisture content (h): wH2 O ¼ ð5 to 10Þ% and (25 to 30)%
used to obtain the results of the 14 descriptor parameters at any pos-
• Biomass input particle size (is): previously chipped with (50 or 20)
sible combination of the input variables. Obtained results from ANN
mm mesh size.
are statistically analyzed and graphically visualized on the 3D sur-
• Process target diameter (dtarget): (5, 2) mm
face response to find the influences of each input variable on the
• Angular speed on hammer mill rotor (rev): (2000, 3000) rpm
particle size distribution. Experimental data are represented over
A total of 32 tests were performed for SRF poplar as result of 24 var- these 3D surfaces to show the good agreement of the neural
iable combinations and a repeat for each test. In the Results section, it networks.
M. Gil, I. Arauzo / Fuel Processing Technology 127 (2014) 80–87 83
Fig. 2. Predictive surfaces of neural network and experimental data points for geometric
mean diameter for poplar and corn stover at different hi, dtarget and rev.
Other tested variables (hi, IS and rev) presented a much lower signif-
icance than dtarget; hi has a slight relevance for poplar and rev for corn
Fig. 1. Examples of cumulative particle size distribution for poplar and corn stover at three stover. Generally, higher rev and lower moisture content decreases
levels of dtarget (5, 3.5 and 2) mm and different conditions of hi, IS and rev. slightly the mean sizes. Bitra et al. [9] also observed this effect of the
rotational speed on milled corn stover.
was found at 3000 rpm for corn stover. predictive 3D-NN surfaces (not presented) show higher coarse reduc-
• At high moisture contents of the biomass, corn stover presented tion for poplar than corn stover. In fine fraction, LSgm(d50/d16) also de-
higher size dispersion than poplar. crease in the same sense, thus it can be associated with two reasons: a
stronger decrease of medium particles than fine or a possible fine in-
As it can be seen in Fig. 3, a dispersion parameter as Sgm showed ab- crease. To find the cause, it can be assumed the same behavior for d16
solutely different behavior than size parameters (Section 3.2). ANOVA than d10 (Section 3.2.2): higher rev implies higher d10 and thus lower
analyses concluded that Sgm is mainly influenced by the angular speed fine for poplar and opposite trend for corn stover. As a result, the LSgm
and input moisture content and very slight by the dtarget, conversely to reduction for poplar is due to median size decrease (d50) as well as in-
analyze for size parameters. In this case, lower rev and higher hi caused crease of d16 (less fine) causing a particle concentration increase on
less distribution dispersion and therefore more homogeneous sizes medium-low region. However for corn stover, the decrease of d50 and
around dgm. In Fig. 3, it can be observed the significant separation be- d16 allow establishing a global decrease of particle sizes that is stronger
tween surfaces at different rev (2000 rpm on top) and the scarce reper- in mediums than in fine regions.
cussion of dtarget.
Within distribution dispersion, it is possible to analyze the disper- 3.3.2. Mass relative span, uniformity index, uniformity coefficient and
sion related to coarse or fine particles. High graphic geometric deviation coefficient of gradation
(HSgm = d84/d50) defines the dispersion related to the high-medium Mass relative span ((MRS = (d90/d84)/d50), uniformity index
distribution region (coarse), as well as low graphic geometric deviation (Iu = d95 / d5), uniformity coefficient (Cu = d60/d10), coefficient of gra-
(LSgm = d50/d16) for the medium-low region (fine). Predictive neural dation (Cg = d230/(d60 − d10)) are dispersion parameters that evaluate
network surfaces (not shown) show similar behavior than Sgm and percentiles from high-medium against low regions of the distribution.
ANOVA analysis established the same influences and repercussions of Predictive neural network surfaces of these parameters shows al-
the input variables as well as multivariate analysis achieved also strong ways the same trends and evolutions between them (e.g. MRS, Fig. 6),
correlations (R2 N 99%) between them. However, HSgm presents a higher as well as the same influences of input variables from ANOVA analysis
average dispersion than LSgm, 31% for poplar and 36% for corn stover, and strongest lineal correlations with R2 N 99% from multivariate analy-
therefore fine concentration is higher than coarse ones, showing a fine sis. The influences of input variables present high similarities with Sgm.
skew distribution (Section 3.4) and therefore dgm b d50 in all cases. For example, in the case of the variable with the highest influence on
Regarding to the first variable on dispersion size influences (rev), an dispersion size (rpm, rev) it is also concluded that higher angular
increase on angular speed involves more homogeneity size effect on speed implies lower dispersion and a higher size homogenization on
poplar than on corn stover. However, it is unknown in which regions poplar than on corn stover. In other words, generally the ratios between
are exactly more affected: coarse, medium or fine. By means of high-medium and low regions decreases when rev increases.
HSgm = (d84/d50), a stronger decrease of coarse fraction can be conclud- Numerically, mass relative span (MRS, Fig. 4) was greater than 1.0,
ed since d50 decreases likewise (Section 3.2.1). Experimental data and which indicated a wide distribution, with values around 2 for both bio-
mass. The higher the MRS the higher the heterogeneity in particle sizes.
Coefficient of gradation was between 1 and 1.3, showing a well graded
particle size [10,9]. In Fig. 1, it can be observed that distributions present
a continuum evolution without irregularities or jumps. Cu showed
values from 3.1 to 6 for poplar and from 3.75 to 6.3 for corn stover.
Cu b 4 means a uniform mix of particles, and in the case of Cu = 1, all
particles have the same size.
Fig. 4. Predictive surfaces of neural network and experimental data points for mass Fig. 5. Predictive surfaces of neural network and experimental data points for geometric
relative span for poplar and corn stover at different hi, dtarget and rev. kurtosis for poplar and corn stover at different hi, dtarget and rev.
behavior and trends like dispersion parameters but noticeable differ- Therefore, both RR parameters appear as useful parameters to describe
ences on the effects of input variables. In fact, dtarget presents again the the central trend and dispersion of the PSD and may be used as input
highest influences, following by rev or hi. In addition, multivariate anal- variables for closed loop control of milling processes.
ysis showed also moderate correlation between Kg and distribution pa- In order to easily check and visualize the evolution of PSD when
rameters (R2 ≈ 0.73), lower one than agreements obtained between (mainly) opening size and angular speed are varied, Fig. 8 shows the
dispersion parameters themselves (R2 ≈ 0.99). RR fitting of experimental PSD with extreme values of lRR and mRR
obtained during the experimental campaign.
3.5. Rosin–Rammler function
4. Discussion
Rosin and Rammler [35] (definition in Section 2.2) obtained a distri-
bution equation to characterize the coal fineness. Subsequently, it was The developed characterization of PSD by means of a semi-empirical
observed that it was well suited to skewed distribution of particle sizes tool allows the prediction of the guidelines to control the milling
and was used to fit hammer-milled switchgrass, wheat straw and corn process in order to adjust the particle size to the requirement of each
stover [9], as well as knife-milled corn stover [10] and switchgrass [11]. final application.
Distribution parameter (mRR) and size parameter (lRR) characterize Energy conversion technologies such as combustion, gasification or
Rosin–Rammler distribution. The first one is the slope of the distribu- lignocellulose hydrolysis in which the residence time and the kinetic
tion. Lower mRR means lower slope and therefore a wider distribution of thermo-chemical conversion govern, require an optimum size to
and therefore higher diversity of the particles sizes. The second one achieve the total particle conversion. The ideal PSD would be a step
corresponds mathematically to lRR = d63, therefore it is a distribution function with the 100% mass percentage for this size. However, particle
percentile. According to the mathematical definition, both parameters fracture by impact produce multiple particles with different morphol-
can be enclosed to the two main described groups of the analyzed ogies and the step function is unachievable in practice. In order to obtain
distribution parameters: dispersion and size parameters, respectively. a product that approaches requirements, the opening size of the mill
Multivariate analysis confirm the strong correlation (R2 N 99%) screen (dtarget) and the angular speed of the rotor (rev) can be con-
between lRR and size parameters (Section 3.2) and between mRR and trolled. The former (dtarget) determines the mean sizes (Section 3.2.1);
dispersion parameters (Section 3.3). ANOVA analysis shows the same in Fig. 8, PSD shows x-axis displacement of PSD as a result of variation
influences and predictive 3D-NN surfaces (Figs. 6 and 7), following the on dtarget. The latter (rev) mainly affects the dispersion above the
same evolutions and trends as function of the four input variables. In mean size (Section 3.3); the higher the rpm the lower the size disper-
the case of mRR, predictive surfaces present an inverse behavior since sion, and therefore, the cumulative particle size distribution can be
higher slopes (mRR) of cumulative distribution involves lower disper- slightly closer to the ideal step function around the characteristics size
sion on particle size, inversely to lower values of dispersion parameters. at 3000 rpm. As a conclusion, the opening size of the screen should be
86 M. Gil, I. Arauzo / Fuel Processing Technology 127 (2014) 80–87
Fig. 6. Predictive surfaces of neural network and experimental data points for lRR for poplar
Fig. 7. Predictive surfaces of neural network and experimental data points for mRR for pop-
and corn stover at different hi, dtarget and rev.
lar and corn stover at different hi, dtarget and rev.
from five to eight times higher than the required size (Section 3.2.1) and between 5 to 8 times lower. Therefore, dtarget must be only considered as
the hammermill should be operated at higher rev to decrease the size a reference size imposed in the process, but not as a characteristic size of
dispersion and obtain a PSD focused around the mean size. milled product.
Other technologies such as pelletization of ground materials may Dispersion parameters confirm that comminuted biomass presents
require higher dispersion on particle size since smaller particles refill wide and well graded distribution with a considerable variety and
the holes between the bigger ones and make the cohesion for pelletizing dispersion on particle size, uniform mix of particles, predominance of
easier. If the handling labors are also taken into account, particles with medium-fine particles (fine skew) and very platykurtic distribution. In
similar sizes contribute to improve the relative movement between par-
ticles due to the lack of smaller ones between them that would increase
the frictional effects.
As a conclusion, the operational parameters, dtarget for central trends
and rev for dispersion size, can be adjusted in order to approach the
requirements established by the final energy conversion technology.
5. Conclusions
milling process, the angular speed (rev) of rotor governs the variability [13] M. Gil, I. Arauzo, E. Teruel, Influence of input biomass conditions and operational
parameters on comminution of short-rotation forestry poplar and corn stover
and dispersion of the particle size. An increase of rev produces higher using neural networks, Energy & Fuels 27 (2013) 2649–2659.
homogenization on particle size for poplar than that for corn stover [14] S. Van Loo, J. Koppejan, The Handbook of Biomass Combustion and Co-firing,
due to higher particle concentration around central sizes. Earthscan Publications Ltd, London, 2008.
[15] J. Pallarés, A. Gil, C. Cortés, C. Herce, Numerical study of co-firing coal and Cynara
Rosin–Rammler function perfectly fits the cummulative particle size Cardunculus in a 350 MWe utility boiler, Fuel Processing Technology 90 (2009)
distribution of comminuted biomass (R2 N 99%) and, in addition, both 1207–1213.
Rosin–Rammler parameters (lRR and mRR) represent the main character- [16] J. Steer, R. Marsh, A. Griffiths, A. Malmgren, G. Riley, Biomass co-firing trials on a
down-fired utility boiler, Energy Conversion and Management 66 (2013) 285–294.
istics of their distributions (central sizes and dispersion). [17] H. Lu, E. Ip, J. Scott, P. Foster, M. Vickers, L.L. Baxter, Effects of particle shape and size
As a conclusion, a computational tool based on neural networks on devolatilization of biomass particle, Fuel 89 (2010) 1156–1168.
was carried out to predict the screen opening size and the rotor angular [18] D.R. Sudhakar, A.K. Kolar, Experimental investigation of the effect of initial fuel par-
ticle shape, size and bed temperature on devolatilization of single wood particle in a
speed required to approach the particle size to the requirements
hot fluidized bed, Journal of Analytical and Applied Pyrolysis 92 (2011) 239–249.
established by the final users. Opening size and angular speed [19] J. Reuther, G. Karsner, S. Jack, Plane flame furnace combustion studies of pulverized
govern the central trends and the dispersion of PSD, respectively. wood, in: 1Fundamentals of Thermochemical Biomass Conversion Conference.
Rosin–Rammler parameters (lRR and mRR) appear as useful key parame- [20] R.J. Gravelsins, Studies of Grinding of Wood and Bark-Wood Mixtures with the
Szego Mill, Ph.D. thesi University of Toronto, 1998.
ters indicators of these PSD characteristics and may be used as inputs for [21] T.G. Bridgeman, L.I. Darvell, J.M. Jones, P.T. Williams, R. Fahmi, Influence of particle size
control systems. on the analytical and chemical properties of two energy crops, Fuel 86 (2007) 60–72.
[22] F.V. Tinaut, A. Melgar, J.F. Pérez, A. Horrillo, Effect of biomass particle size and air
superficial velocity on the gasification process in a downdraft fixed bed gasifier. An ex-
Acknowledgments perimental and modelling study, Fuel Processing Technology 89 (2008) 1076–1089.
[23] J.J. Hernández, G. Aranda-almansa, A. Bula, Gasification of biomass wastes in an
This work was partially supported by Spanish Ministry of Education, entrained flow gasifier: effect of the particle size and the residence time, Fuel
Processing Technology 91 (2010) 681–692.
Culture and Sports (project ENE2008-03358/ALT), funded by the Minis- [24] A. Hendriks, G. Zeeman, Pretreatments to enhance the digestibility of lignocellulosic
try of Education, Culture and Sports, Spain. The authors are grateful to biomass, Bioresource Technology 100 (2009) 10–18.
Mr. Óscar Puyó for his extensive support in the laboratory work and to [25] V.B. Agbor, N. Cicek, R. Sparling, A. Berlin, D.B. Levin, Biomass pretreatment:
fundamentals toward application, Biotechnology Advances 29 (2011) 675–685.
the agricultural cooperative of Sariñena for the corn stover supply. [26] V. Chang, B. Burr, M. Holtzapple, Lime pretreatment of switchgrass, Applied
Biochemistry and Biotechnology 63–65 (1997) 3–19.
References [27] K.I. Ileleji, B. Zhou, The angle of repose of bulk corn stover particles, Powder Technol-
ogy 187 (2008) 110–118.
[1] E. Commission, Europe 2020, 2010. [28] M. Gil, I. Arauzo, E. Teruel, C. Bartolome, Milling and handling Cynara Cardunculus L.
[2] S. Paulrud, J.E. Mattsson, C. Nilsson, Particle and handling characteristics of wood for use as solid biofuel: experimental tests, Biomass and Bioenergy 41 (2012)
fuel powder: effects of different mills, Fuel Processing Technology 76 (2002) 23–39. 145–156.
[3] C. Igathinathane, L. Pordesimo, E. Columbus, W. Batchelor, S. Methuku, Shape iden- [29] M. Gil, D. Schott, I. Arauzo, E. Teruel, Handling behavior of two milled biomass: SRF
tification and particles size distribution from basic shape parameters using ImageJ, poplar and corn stover, Fuel Processing Technology 112 (2013) 76–85.
Computers and Electronics in Agriculture 63 (2008) 168–182. [30] CEN/TS 14778–1, Solid biofuels. Sampling. Methods for sampling, European
[4] C. Igathinathane, L. Pordesimo, E. Columbus, W. Batchelor, S. Sokhansanj, Sieveless Committe for Standardization, Brussels, 2005.
particle size distribution analysis of particulate materials through computer vision, [31] CEN/TS 14774–1, Solid biofuels. Methods for the determination of moisture content.
Computers and Electronics in Agriculture 66 (2009) 147–158. Oven dry method, Part 1: Total moisture. Reference method, European Committe for
[5] Q. Guo, X. Chen, H. Liu, Experimental research on shape and size distribution of Standarization, Brussels, 2004.
biomass particle, Fuel 94 (2012) 551–555. [32] CEN/TS 15149–2, Solid biofuels. Methods for the determination of particle size
[6] M. Gil, E. Teruel, I. Arauzo, Analysis of standard method for milled biomass through distribution, Part 2: Vibrating screen method using sieve apertures of 3.15 mm
image processing, Fuel 116 (2014) 328–340. and below, European Committe for Standarization, Brussels, 2006.
[7] M. Himmel, M. Tucker, J. Baker, C. Rivard, K. Oh, K. Grohmann, Comminution of [33] ISO3310-1, Technical requirements and testing — part 1: test sieves of metal wire
biomass: hammer and knife mills, Biotechnology and bioengineering symposium, cloth, International Organization for Standardization, Geneva, 2000.
vol. 15, Wiley, J, New York, 1986, pp. 39–58. [34] R.L. Folk, W.C. Ward, Brazos River bar: a study in the significance of grain size pa-
[8] S. Mani, L.G. Tabil, S. Sokhansanj, Grinding performance and physical properties of rameters, Journal of Sedimentary Petrology 7 (1957) 3–26.
wheat and barley straws, corn stover and switchgrass, Biomass and Bioenergy 27 [35] P. Rosin, E. Rammler, Laws governing the fineness of powdered coal, Journal of the
(2004) 339–352. Institute of Fuel 7 (1933) 29–36.
[9] V.S.P. Bitra, A.R. Womac, N. Chevanan, P.I. Miu, C. Igathinathane, S. Sokhansanj, D.R. [36] C.M. Bishop, Neural Networks for Pattern Recognition, Oxford University Press,
Smith, Direct mechanical energy measures of hammer mill comminution of switch- New York, 1995.
grass, wheat straw, and corn stover and analysis of their particle size distributions, [37] M. Hagan, H. Demuth, M. Beale, Neural Network Design, PWS Publishing, 1996.
Powder Technology 193 (2009) 32–45. [38] H. Demuth, M. Beale, M. Hagan, Neural Network Toolbox, For Use with Matlab, The
[10] V.S.P. Bitra, A.R. Womac, Y.T. Yang, P.I. Miu, C. Igathinathane, S. Sokhansanj, Mathe- MathWorks, 2001.
matical model parameters for describing the particle size spectra of knife-milled [39] D. Rumelhart, G. Hinton, R. Williams, Learning Internal Representations by Error
corn stover, Biosystems Engineering 104 (2009) 369–383. Propagation, MIT Press Cambridge, USA, 1986.
[11] V.S.P. Bitra, A.R. Womac, Y.T. Yang, C. Igathinathane, P.I. Miu, N. Chevanan, S. [40] [40] M. Gil, I. Ramos, I. Arauzo, J. Román, Characterization of a biomass milling
Sokhansanj, Knife mill operating factors effect on switchgrass particle size distribu- pilot plant, in: G. F. De Santi (Ed.), 17th European Biomass Conference, Hamburg,
tions, Bioresource Technology 100 (2009) 5176–5188. pp. 447–454.
[12] V.S.P. Bitra, A.R. Womac, Y.T. Yang, P.I. Miu, C. Igathinathane, N. Chevanan, S. [41] S.J. Blott, K. Pye, GRADISTAT: a grain size distribution and statistics package for the
Sokhansanj, Characterization of wheat straw particle size distributions as affected analysis of unconsolidated sediments, Earth Surface Processes and Landforms 26
by knife mill operating factors, Biomass and Bioenergy 35 (2011) 3674–3686. (2001) 1237–1248.