You are on page 1of 6

70 Ind. Eng. Chem. Res.

2006, 45, 70-75

Application of a Neural Network for the Prediction of Crystallization Kinetics


Meng Yang and Hongyuan Wei*
School of Chemical Engineering and Technology Tianjin UniVersity,
Tianjin, 300072, People’s Republic of China

A method for more-accurate prediction of crystallization kinetics is greatly needed in the field of industrial
crystallization. Traditional empirical correlations cannot give reliable predictions, because of the highly nonlinear
behavior of crystallization kinetics, although they have been used for a long time. In this paper, the development
of a neural network model is presented. The model was trained with limited data obtained from an anti-
solvent crystallization system (ciprofloxacin hydrochloride, H2O, and ethanol). The predictions from the network
then were validated against newly measured data. The results confirm that this approach gives much more-
accurate predictions of the kinetics, in terms of crystal growth and agglomeration as examples. The mean
relative error of the predicted growth rates from this model, versus the measured data, is generally <10%
and, in some cases, is as good as 5%. This is a significant improvement on the relative error of 20% or more
that is typically achieved by traditional correlations.

1. Introduction very well. Such networks can overcome the problems associated
with linear and perception networks. The network constructed
The artificial neural network (ANN) has been developing for
in this work adopts such a multilayer structure, to approach a
many years and has an increasingly significant role across a
function with definite discrete points. The training strategy of
wide range of fields. ANN has been applied to chemical
the network is shown in Figure 1.
engineering over the last 16 years. Since 1988, when Hosking
and Himmelblau1 first applied it to error diagnosis for a chemical Determining a suitable learning rate for a nonlinear network
process, it has been widely used for various applications in is important. A large learning rate may lead to instability.
chemical engineering.2 Conversely, if the learning rate is too small, an unrealistically
Crystallization kinetics (e.g., nucleation rate or crystal growth) long training time may be needed. Unlike linear networks, there
is essential for the analysis, design, and operation of industrial is no easy, prescriptive way of selecting a suitable learning rate
crystallization processes. Traditionally, kinetics is described by for nonlinear multilayer networks.
correlations that are obtained from experimental data using The network constructed in this work generalizes the Wid-
assumed functional forms. The mechanisms of nucleation and row-Hoff study rule; it trains the weightings of nonlinear
crystal growth are extremely complex, with high nonlinearity differentiable functions and is trained by input vector and target
and parameter interactions, which are not easily reflected by vector pairs. The ANN model can finally approximate a function
simple empirical correlations, so that the use of those correla- that relates the input vectors to specific target vectors.
tions normally leads to large errors in predictions. However, A multiple-layer neural network consists of input vectors,
the neural network approach is better able to reproduce the hidden layers, and output layers. In the current work, the network
overall system behavior, with its nonlinearities and parameter is constructed with two layers, which is believed to give
interaction effects. It can not only capture the interactions sufficiently good approximations for predicting kinetics.
between each influential element but also provide the mapping The detailed structure of the network is illustrated in Figure
procedure from input to output, so it is expected to be able to 2, in which P is the input vector and the parameters b1 and b2
predict the complex interplays of the influential factors in represent the hidden-layer and output-layer bias vectors, re-
crystallization processes. spectively. N1 represents the hidden-layer parameter vectors,
This paper describes the development of a neural network which can be defined as
model to predict crystal nucleation, growth, and agglomeration
rates. An anti-solvent crystallization system that is composed N1 ) W1P + b1 (1)
of ciprofloxacin hydrochloride, H2O, and ethanol was used for
the purposes of training and validating the model. where W1 is the hidden-layer weighting matrix.
The hidden-layer propagation function, f1(x), is expressed as
2. Construction of the Network
Several papers have suggested that using a multiple-layer 2
f1(x) ) -1 (2)
neural network is a more effective way to approximate nonlinear 1 + e-2x
relationships between parameters.4-8 In fact, 80%-90% of ANN
applications use a multiple-layer neuron structure or derivatives Its purpose is to send information in the hidden layer to the
thereof. output layer. The output parameter, N2, can be calculated using
Multilayered networks can perform well on any linear or the following equation:
nonlinear problems and can approximate any reasonable function
N2 ) W2P + b2 (3)
* To whom correspondence should be addressed. Tel.:
+86 22 27405754. Fax: +86 22 27400287. E-mail address:
david.wei@tju.edu.cn. where W2 is the output-layer weighting matrix.
10.1021/ie0487944 CCC: $33.50 © 2006 American Chemical Society
Published on Web 11/30/2005
Ind. Eng. Chem. Res., Vol. 45, No. 1, 2006 71

Figure 1. Training strategy of the constructed artificial neural network


(ANN) model.
Figure 3. Experimental setup. Legend: 1, crystallizer; 2, agitator; 3,
thermometer; and 4, thermostat.

weighed. A sample of the mother liquor was also analyzed, so


that the solution concentration and suspension density were
obtained. Crystal size distributions of samples were obtained
using a laser diffraction granulometer (Malvern Mastersizer).
Figure 2. Structure of the two-layer network. The experimental apparatuses are shown in Figure 3.

The definition of the output-layer propagation function, f2(x),


4. Experimental Data Analysis
is
The crystal growth rate (G) and nucleation rate (B0) are
f2(x) ) x (4)
calculated using the following equations with equilibrium data9
and y is the output vector. and experimental data:10

[ ]
In this application, supersaturation (S), temperature (t), volume
fraction (a), solid suspension density (Mt), and agitation rate dC(t) d ln V(t)
(Nstr) are defined as the input vectors (suspension density is - + (C(t) + Mt(t))
dt dt

[ ]
defined as the mass of solid particles per unit volume of slurry), G) (6)
whereas growth rate (G), nucleation rate (B0), and agglomeration n wti%
coefficient (Ka) comprise the output vectors. The use of the Mt(t) ∑
i)1 (L - Li)
tansig function (f1(x)) as the hidden propagation function, and i+1
the purelin function (f2(x)) as the output propagation function,
enable this network to approximate any function reasonably well. B0 ) n(L1,t) × G (7)
All weights and biases in the network are obtained through a
training procedure that is based on the measured data. After a The agglomeration coefficient (Ka) is the constant (assuming
predefined level of precision is achieved, the training procedure size independence) for describing the agglomeration rate (ragg),
can be terminated and the multiple-layer network is then ready as shown in eq 8 and can be obtained by discretizing population
to be used for prediction. balance:
To evaluate the accuracy of the network constructed, the
following mean relative error, Q, is used:

Q)
1 N

∑1 |
the prediction value - the experimental value
|
ragg ) Ka (∑
1 i-1
2 j)1
Ni-jNj - Nim0 ) (8)

N the experimental value


(5) 5. The Traditional Correlation Approach for Kinetics
Predictions
3. Experiments
The experiments were performed in a 6-L baffled crystallizer The empirical correlation for predictions of crystallization
agitated by a propeller at different stirring speeds, ranging from kinetics in the traditional format can be obtained through the
250 rpm to 750 rpm. The diameter of the crystallizer is 18 cm, regression of experimental data, as shown in eqs 8 and 9. The
and the height is 25 cm. The working volume in each run was detailed descriptions for such method can be found in ref 11.
5 L. The temperature of the crystallizer was controlled by The format for linear growth rate (number-length coordinate)
recycling water from a thermal basin. In a typical run, a in the traditional correlation is
preheated and filtered saturated aqueous solution of high-purity
E
ciprofloxacin hydrochloride was placed into the crystallizer. The
temperature of the crystallizer was maintained 5 °C higher than
G ) C exp - ( )
N d(S - 1)b
RT str
(9)
that of the solution, to ensure that no crystals exist in the solution
prior to crystallization. After 0.5 h, the temperature of the The constants C, E, d, and b can be obtained by a regression
crystallizer was reduced to the saturation temperature of the procedure from experimental data. For this studied system, C
ciprofloxacin hydrochloride solution. A certain amount of seeds ) 4.32 × 10-10, E ) 12761.3, d ) 0.887, and b ) 0.451. Figure
was fed into the crystallizer before the batch experiment 4 shows the predicted values (the solid line) coming from the
commenced. Supersaturation was obtained by slow cooling. A correlation developed by fitting the measured values against
small amount of sample was removed and filtered at intervals measured data (depicted by solid dots). The mean relative error
of every 15 min during the process. Particles were dried at a (Q) can be up to 19.5%, which reflects the accuracy of the
temperature of 95 °C under vacuum conditions and then prediction.
72 Ind. Eng. Chem. Res., Vol. 45, No. 1, 2006

Figure 4. Correlation between predicted and measured values of the growth


rate, G.

Figure 6. Correlation between predicted and measured values of the


agglomeration coefficient, Ka.

Figure 5. Correlation between predicted and measured values of the


nucleation rate, B0.

The nucleation rate, B0, can be regressed as the following


function:
Figure 7. Correlation between predicted and measured values of G.
B ) (5.29E + 17)(1 - a)
0 0.103
Mt0.932
Gv0.326
(10)
6. The Neural Network Approach for Kinetics
Predictions
The predicted values against measured data are shown in Figure
5. The mean relative error Q is 20%. The ANN model is sensitive to the number of neurons in its
The agglomeration coefficient, Ka, can be estimated by a hidden layers. Too few neurons may lead to underfitting.
semiempirical correlation, which is the function of fluid However, too many would contribute to overfitting, in which
conditions such as temperature, viscosity, and shear rate and is the fitting curve exhibits wild oscillations between well-fitted
assumed to have a function similar to that of the growth rate: training points. Hence, it is very important to select a suitable
number of neurons. After many trials, the model can give highly
7945.5
Ka ) (3.64E - 20) exp - ( RT )
Nstr0.448(S - 1)0.230
(11)
satisfactory results with 10 neurons in the hidden layer and 1
neuron in the output layer.
The weightings and biases in the model are as follows:
The predicted values against measured data are shown in Figure W1 ) [wij], where i is the number of input parameters and j
6. The mean relative error Q is 17%. is the number of neurons in the hidden layer
The experiments were also performed in two crystallizers with W2 ) [wij], where i is the number of output parameters and
different sizes (0.5 and 1.5 L), and data were regressed for the j is the number of neurons in the hidden layer
linear growth rate G, as shown in eq 12: b1 ) [bj], where j is the number of neurons in the hidden
layer
46329.5
G ) (5.579E - 3) exp - ( RT )
Nstr2.043(S - 1)0.349
(12)
b2 ) [bj], where j is the number of output parameters
The network was first trained with limited data obtained from
experiments, and predictions from the trained network were then
The predicted values against measured data are shown in Figure validated against newly measured data.
7. Very obviously, the error is even larger (Q > 40%). Network for Linear Growth Rate Prediction under the
To summarize the aforementioned comparisons, the traditional Number-Length Coordinate. The model constructed in this
approach for kinetic predications is not satisfactory and can lead case was trained with 31 sets of experimental data. Figure 8
to large errors. shows that the resulting correlation between the predicted values
Ind. Eng. Chem. Res., Vol. 45, No. 1, 2006 73

Figure 8. Training the multilayer ANN network: correlation between the


experimental and predicted values of G.

Figure 10. Training the multilayer ANN network: correlation between


the experimental and predicted values of B0.

Figure 9. Multilayer ANN network prediction: comparison between the


experimental and predicted values of G.

Table 1. Validation: Experimental and Predicted Values of the


Growth Rate, G
Growth Rate, G (m/s)
experiment value predicted value
3.07 × 10-10 3.04 × 10-10
4.64 × 10-10 4.52 × 10-10 Figure 11. Training the multilayer ANN network: correlation between
4.97 × 10-10 5.64 × 10-10 the experimental and predicted values of Ka.
5.20 × 10-10 5.82 × 10-10
7.47 × 10-10 6.39 × 10-10 Table 2. Validation: Experimental and Predicted Values of the
9.54 × 10-10 1.00 × 10-9 Nucleation Rate, B0
Nucleation Rate, B0 (number/(m3 s))
(the solid line) and measured data (the open circles) is extremely experimental value predicted value
good. The mean relative error Q is only ∼1%.
2.17 × 1011 2.06 × 1011
Another six sets of experimental data were selected to validate 2.16 × 1011 2.16 × 1011
the model predictions away from training points. The experi- 2.30 × 1011 2.38 × 1011
mental value and the predicted values are shown as Table 1 3.27 × 1011 2.84 × 1011
and Figure 9. All predictions lie within 15% of the measured 3.00 × 1011 2.57 × 1011
2.789 × 1011 2.63 × 1011
values, with a mean relative error of Q ) 8.01%.
Network for Simultaneous Prediction of Nucleation Rate values and the predicted values for B0 and Ka are shown in
and Agglomeration Coefficient under the Number-Volume Tables 2 and 3, respectively, and, correspondingly, Figures 12
Coordinate. The model constructed in this case was trained and 13. The predicted values of B0 and Ka were within 15%
with 34 sets of experimental data. and 6%, respectively, of measured data, with corresponding
Figures 10 and 11 show the correlation between the predicted mean relative errors of 7.05% and 3.66%.
and measured values of B0 and Ka, respectively. The corre- Network for Growth Rate Prediction with Different
sponding mean relative errors are 9.42% and 5.8%. Crystallizers, under the Number-Length Coordinate. Two
Another six sets of experimental data were selected to validate sizes of crystallizers were used for the experiments; these had
the model predictions between training points. The experimental working volumes of 0.5 and 1.5 L.
74 Ind. Eng. Chem. Res., Vol. 45, No. 1, 2006

Figure 12. Multilayer ANN network prediction: comparison between the


experimental and predicted values of B0.

Figure 14. Training the multilayer ANN network: correlation between


the experimental and predicted values of growth rate.

Figure 13. Multilayer ANN network prediction: comparison between the


experimental and predicted values of Ka.

Table 3. Validation: Experimental and Predicted Values of


Agglomeration Coefficient, Ka
Agglomeration Coefficient, Ka (µm3/s) Figure 15. Multilayer ANN network prediction: comparison between the
experimental value predicted value experimental and predicted values of G.
3.38 × 10-20 3.22 × 10-20 Table 4. Validation: Experimental and Predicted Values of the
3.55 × 10-20 3.43 × 10-20 Growth Rate, G
4.50 × 10-20 4.24 × 10-20
3.27 × 10-20 3.26 × 10-20 Growth Rate, G (µm/s)
7.09 × 10-20 6.84 × 10-20 experimental value predicted value
7.37 × 10-20 7.08 × 10-20
2.91 × 10-5 2.93 × 10-5
2.59 × 10-5 2.61 × 10-5
The model constructed for this study was trained with 34 4.32 × 10-5 4.00 × 10-5
sets of experimental data. Figure 14 shows the correlation 4.62 × 10-5 4.60 × 10-5
between the predicted and measured values of G. The mean 5.49 × 10-5 5.46 × 10-5
relative error is Q ) 2%. 5.15 × 10-5 5.18 × 10-5
Another six sets of experimental data were selected to validate forms.3 Therefore, making use of these empirical correlations
the model predictions between training points. The experimental can lead to large errors.
and predicted values are shown in Table 4 and Figure 15. The mechanism of crystallization kinetics is very complex
Predicted values are within 7.5% of the experimental data, with with strong nonlinearity, which cannot be reflected accurately
a mean relative error of Q ) 1.73%. by a simple regression function. However, the neural network
approach is better able to reproduce overall system behaviors,
7. Conclusions with its nonlinearities and parameter interaction effects. It can
not only capture the interactions between each influential
Crystallization kinetics (such as crystal nucleation, growth, element but also provide the mapping procedure from input to
and agglomeration rates) is essential for the analysis, design, output. Its ability to predict the complex interplays of influential
and operation of industrial crystallization processes. The factors in crystallization processes result in more-accurate
traditional correlation approach for predicting kinetics is predictions of the crystallization kinetics, as demonstrated in
arbitrary and subjective in its selection of assumed functional this paper.
Ind. Eng. Chem. Res., Vol. 45, No. 1, 2006 75

In summary, the application of an artificial neutral network Literature Cited


(ANN) overcomes, to a significant degree, the problems that
(1) Hosking, J. C.; Himmelblau, B. M. Artificial neural network models
result from the traditional use of a regressional analysis of kinetic of knowledge representation in Chemical Engineering. Comput. Chem. Eng.
data to an assumed functional form. It provides us with an 1988, 12, 881-891.
effective and applicable way to obtain dependable kinetics (2) Bhagat, P. An introduction to neural nets. Chem. Eng. Prog. 1990,
parameters. Future work will examine the applicability of this 8, 55-65.
method to other systems and situations. (3) Mullin, J. W. Crystallization; 3rd Edition; Butterworth-Heinemann:
Oxford, U.K., 1993.
(4) Larachi, F.; Iliuta, I.; Chen, M.; Grandjean, B. P. A. Onset of Pulsing
Nomenclature in Trickle Beds: Evaluation of Current Tools and State-of-art Correlation.
Can. J. Chem. Eng. 1999, 77, 751-757.
B0 ) nucleation rate (number/(m3 s))
(5) Larachi, F.; Bensetiti, Z.; Grandjean, B. P. A.; Wild, G. Two-Phase
b1 ) hidden-layer bias Frictional Pressure Drop in Flooded-Bed Reactors: A State-of-the-art
b2 ) outer-layer bias Correlation. Chem. Eng. Technol. 1998, 21 (11), 887-893.
f1(x) ) tansig function (6) Bensetiti, Z.; Larachi, F.; Grandjean, B. P. A.; Wild, G. Liquid
f2(x) ) purelin function saturation in cocurrent upflow fixed-bed reactors: a state-of-the-art cor-
G ) growth rate (m/s) relation. Chem. Eng. Sci. 1997, 52 (21-22), 4239-4247.
(7) Iliuta, I.; Larachi, F.; Grandjean, B. P. A. Pressure Drop and Liquid
Ka ) agglomeration coefficient (µm3/s) Holdup in Trickle Flow Reactors: Improved Ergun Constants and Slip
mi ) ith moment Correlations for the Slit Model. Ind. Chem. Eng. Res. 1998, 37, 4542-
Mt ) suspension density (g/L) 4550.
Ni ) number concentration at size i (number/m3) (8) Iliuta, I.; Larachi, F.; Grandjean, B. P. A.; Wild, G. Gas-liquid
Nstr ) stirring rate (rpm) interfacial mass transfer in trickle-bed reactors: state-of-the-art correlations.
Chem. Eng. Sci. 1999, 54, 5633-5645.
P ) input vector
(9) Liu, Y.; Wang, J.; Yin, Q. Study on the thermodynamics of
S ) supersaturation crystallization of ciprofloxacin hydrochloride monohydrate. Tianjin Daxue
t ) temperature (°C) Xuebao, Ziran Kexue Yu Gongcheng Jishuban 2002, 35 (3), 313-316.
W1 ) inner-layer weighting (10) Liu, Y.; Wang, J. K.; Wei, H. Y. Determination of crystallization
W2 ) outer-layer weighting kinetics in solution. J. Cryst. Growth 2004, 271, 238-244.
(11) Liu, Y. Study on the synthesis and crystallization process of
ciprofloxacin hydrochloride monohydrate, Ph.D. Thesis, Tianjin University,
Acknowledgment Tianjin, PRC, 2003.
The authors wish to acknowledge the National Natural
Science Foundation of China for financial support, and they ReceiVed for reView December 14, 2004
also thank Dr. Liu Yong (Tianjin University) and Dr. Simon ReVised manuscript receiVed October 25, 2005
Accepted November 3, 2005
Leefe (Chembridge Technologies, Limited, U.K.) for their
valuable assistance with this work. IE0487944

You might also like