new approach in modeling of metallocene-catalyzed olefin polymerization using artificial neural...
TRANSCRIPT
Full Paper
New Approach in Modeling ofMetallocene-Catalyzed Olefin PolymerizationUsing Artificial Neural Networks
Mostafa Ahmadi, Mehdi Nekoomanesh,* Hassan Arabi
A new approach for the estimation of kinetic rate constants in olefin polymerization usingmetallocene catalysts is presented. The polymerization rate has been modeled using themethod of moments. An ANN has been used and trained to behave like the mathematicalmodel developed before, so that it gets polymerization rate at different reaction times andpredicts reaction rate constants. The network wastrained using modeling results in desired oper-ational window. The polymerization rates werenormalized to make the network work indepen-dent of operational conditions. Themodel has alsobeen applied to real polymerization rate data andthe predictions were satisfactory. This model isspecially useful in comparing different newmetallocene catalysts.
Introduction
Development of an accurate kinetic model plays a vital role
in the prediction of reactor behavior as well as in
maximizing the yield of desired products. It means that
all the information about the mechanism of reactions and
their unknown parameters, such as kinetic rate constants,
have to be known in order to optimize the process.
Determination of the kinetic rate constants has been the
main subject of several researches.[1–22] Earlier, it was
conventional to obtain the rate constants from time
average reaction rates and final molecular weights,[1–5]
while in recent models, instantaneous reaction rates are
M. Ahmadi, M. Nekoomanesh, H. ArabiDepartment of Polymerization Engineering, Iran Polymer andPetrochemical Institute, P.O. Box 14965/115, Tehran, IranFax: þ98 2 144580 0213; E-mail: [email protected]
Macromol. Theory Simul. 2009, 18, 195–200
� 2009 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
being used.[6–22] In these methods, the algebraic function,
developed for desired output[7–9] (like polymerization rate)
or the nonlinear differential equations from mass con-
servation law,[10] is coupled with an optimization algo-
rithm in order to find the best rate constants. Generation of
certain rate constants needs almost numerous experi-
mental runs, which is considered as a serious limitation of
these methods.
The use of artificial neural networks (ANNs) as powerful
tools for modeling, optimization, and control of chemical
processes has become increasingly popular, especially
where the process is complex or the exact values of
parameters are unknown.[23–31] In most cases, ANNs are
used directly, for example where the network is trained to
get the reactor’s input variables (like temperature and
pressure) and reproduce its dependent outputs (like
polymerization yield or molecular weight).[23–26] In the
inverse method, they are used to find the boundaries of
DOI: 10.1002/mats.200800088 195
M. Ahmadi, M. Nekoomanesh, H. Arabi
196
reactor’s input variables required for desired outputs.[27–29]
ANNs answer much faster than a mathematical model,
especially where the mathematical model is complex. This
enhanced speed can be very useful for optimization
problems. Therefore, in some cases ANNs are used to
compensate the shortcomings of conventional mechan-
istic models such as the long time required for calcula-
tions[23,24] or the discrepancy between calculated and
measured molecular weight distribution in Ziegler-Natta
catalyzed ethylene polymerization.[25] However, these are
the most practical uses of ANNs in polymerization
modeling, but training of a reliable network needs a vast
number of experimental samples, uniformly distributed in
the entire design space that is not always available.
Hornik, Stinchcombe, and White proved that standard
multilayer feed forward networks are capable of approx-
imating any measurable function to any desired degree of
accuracy.[30] Consequently, the lack of success in many
applications arises from inadequate learning, forgetting
some effective inputs, or unsuitable network structure for
that particular task. Therefore, the necessity of being
skilled in use of ANNs is another problem in addition to the
need of numerous experimental data.
The mentioned problems associated with the optimiza-
tion of mathematical models and conventional applica-
tions of ANNs make one to look for an alternative way to
predict the polymerization outputs with a few experi-
mental results. This paper introduces a new approach
based on ANNs to study the kinetics of olefin polymeriza-
tion using metallocene catalysts. The main advantage of
this procedure is that it can be used to predict polymeriza-
tion behavior using a few experimental results, where the
neccesity of having a lot of experimental samples is the
most important shortcoming of current mathematical and
conventional ANN-based models.
Modeling
Polymerization Kinetics
At the first step, the simplest kinetic model was developed
for the prediction of the reaction rate in olefin polymer-
ization using metallocene catalysts. Despite the catalyst
being used, the basic scheme generally accepted for all
olefin polymerizations includes chain initiation Equation
(1), chain propagation Equation (2), and chain deactivation
Equation (3):
Macrom
� 2009
C� þ M �!kiP1 (1)
Pn þ M �!kpPnþ1 (2)
ol. Theory Simul. 2009, 18, 195–200
WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
kd
Pn �! Dn (3)Site activation was assumed instantaneous, even in the
absence of premixing time. Population balance for
polymerization components is written below:
d½C��dt
¼ �ðki½M� þ kdÞ½C�� (4)
d½P �
1dt¼ ki½M�½C�� � ðkp½M� þ kdÞ½P1� (5)
d½P �
ndt¼ kp½M�ð½Pn�1� � ½Pn�Þ � kd½Pn� (6)
d½D �
ndt¼ kd½Pn� (7)
Using the moments of chain length distribution one can
calculate polymerization rate and yield according to the
following equations:
rate ¼ ½M� ki½C�� þ kpl0
� �(8)
yield ¼ M ðl þ m Þ (9)
0 1 1where ln and mn are the nth order moments for the live and
dead chains, respectively and M0 is the monomer molar
mass. The following equations can be derived for the
calculation of the moments of live and dead chains:
dl0
dt¼ ki½M�½C�� � kdl0 (10)
dl
1dt¼ kiMC� � kdl1 þ kpMl0 (11)
dl
2dt¼ ki½M�½C�� � kdl2 þ kp½M�ð2l1 þ l0Þ (12)
dm
idt¼ kdli (13)
Training the Network
A common type of ANN with multiple layers and
supervised learning rule called feed forward was used.
Standard back propagation with Levenberg-Marquardt
algorithm was used to train the network. A set of
polymerization rates reported at 32 time points calculated
with the kinetic model proposed before was used as the
DOI: 10.1002/mats.200800088
New Approach in Modeling of Metallocene-Catalyzed Olefin Polymerization . . .
Figure 1. (a) The optimal network structure and (b) the corre-sponding training errors.
network’s input and the corresponding rate constants (ki,
kp, and kd) were used as the network’s outputs. The
considered time points were more at the beginning of
polymerization to feel the effect of initiation rate constant.
Eight levels were considered for ki between 0.001 and
10 L �mol�1 � s�1, seventeen levels for kp between 1 and
20 L �mol�1 � s�1, and seven levels for kd between 10�6
and 0.0028 s�1 to produce a total number of 952 sets of
polymerization rates. In this scenario, the network will
modify the weight of interconnections between neurons
to reproduce the given rate constants. Limits of the
rate constants were selected according to the published
data[7–10] and the information about the kinetics of our
target catalyst. It should be mentioned that the steps of
initiation and deactivation rate constant were selected
logarithmic to emphasize the effect of low values as well as
high values.
The most common problem with ANN-based models is
the lack of generalization capability which means that
they would memorize only the training set if not properly
trained. This problem can be overcome by carefully
selecting the range of data and using early stopping
method in training. Thus, early stopping method was used
in the training step to ensure the generalization. In this
method, the calculated polymerization rate sets were
divided into three subsets. The first one is used for training
(training set). The error on the second subset (validation
set) is monitored during training and the training stops if
the validation error increases. The error of the third subset
(test set) is also monitored to check whether the network
works well in the test area. In the ideal situation, the errors
of three subsets decrease concurrently during the training
course and any increase in each subset leads to incorrect
predictions. In addition, ANNs do not always predict well
near the borders of their training range. Extending the
training range so that the region of interest falls within
95% of the total training range would minimize the border
problem.
All neuron’s transfer functions were tan-sigmoid except
in the output layer where log-sigmoid function was used.
All the input and output data were coded to vary between
0.1 and 0.9 according to the following equation:[31]
Macrom
� 2009
Pcoded ¼ 0:8P � Pmin
Pmax � Pmin
� �þ 0:1 (14)
There is no rule for finding the best ANN architecture
and it depends on the specific characteristics of each
unique problem. The best ANN architecture should be
found using trial and error. This is done by testing several
topologies and comparing the three errors of training
course. Different number of hidden layers and neurons (in
each layer) were tested to get the least error. In our case,
ol. Theory Simul. 2009, 18, 195–200
WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
the optimal network contains two hidden layer with 20
neurons in the first and 10 neurons in the second layer.
Figure 1 shows the optimal network structure and the
corresponding training errors.
Results and Discussion
In order to generalize the network to work independent
from operational conditions, the calculated polymeriza-
tion rates were divided by the monomer and catalyst
concentrations [since the rate is proportional to these
parameters according to Equation (8)] before coding
Equation (14). Once the network has been trained, it can
be fed with different polymerization rate curves (experi-
mental or calculated ones) and calculate the kinetic
parameters, endlessly. ANN predictions can be used in
the mechanistic model to reproduce the input data. In this
way we have modeled an unknown experimental data
according to our desired reaction mechanism. Since there
is no comprehensive kinetic scheme for all catalysts, the
www.mts-journal.de 197
M. Ahmadi, M. Nekoomanesh, H. Arabi
Figure 3. The experimental [32] (points) and predicted (lines)polymerization rates, polymerization conditions: pressure¼1 bar, [Zr]¼ 1.4� 10�5 mol � L�1, [Al]/[Zr]¼ 5 000, in 200 mL ofdecane.
198
best kinetic scheme and best range of rate constants
should be chosen according to the experimental results.
Four sets of rate constants were generated randomly
between 0.1 and 0.9 and transferred to the actual values to
produce network inputs using the mechanistic model,
randomly. The calculated polymerization rates were fed to
the trained network and the output rate constants were
used in the mechanistic model to reproduce the poly-
merization rates as well as polymerization yields. Figure 2
shows the calculated and predicted coded polymerization
rates and yields. It is shown that predictions are very close
to the actual values, which means that the network can
predict the outputs between the training steps. In addition,
the combined model can predict the polymerization yield
even when the network was trained using polymerization
rate data.
The network was also applied to the real case of
propylene polymerization using ansa-zirconocene catalyst
system of Me2Si(Ind)2ZrCl2/MAO in decane.[32] Details of
polymerization conditions can be found elsewhere.
Figure 3 shows the experimental rates and the model
Figure 2. The polymerization rates and polymerization yieldscalculated with the mechanistic model (lines) and predictionsof the ANN model (points).
Macromol. Theory Simul. 2009, 18, 195–200
� 2009 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
predictions. There are some discrepancies between mea-
surements and predictions, especially at shorter reaction
times and lower temperatures. Evidently, this simple
kinetic scheme could not capture all the aspects of the real
complex behavior of metallocene catalyzed olefin poly-
merization. However the model is successful in predicting
the general behavior and this simple example reveals the
validity of the proposed method. These discrepancies can
be attributed to the simplicity of the selected kinetic
scheme and as can be found in Ref.[32] some more complex
kinetic schemes should be considered for predicting this
special decay type rate curve. In addition to the complexity
of the real kinetic behavior it should be noted that the
reaction media is not the same for all runs, because the
polymer becomes soluble as the polymerization tempera-
ture goes up and accordingly the molecular weight
becomes lower. This phenomenon causes the diffusion
limitations become more important in the prediction of
the polymerization kinetics.
Table 1 lists the kinetic rate constants at different
temperatures. When the judgment is based on observa-
tion, there are many acceptable rate constants that can
draw a specific rate curve but using this method, there is
only a unique set of rate constant that can capture the
desired rate curve. Therefore, this method is useful in
comparing kinetics of different catalysts. Two catalysts
can be compared in different operational conditions,
specifically different cocatalyst concentrations, only if
they show similar trends toward cocatalyst concentration
or the kinetic scheme should be modified to include
excessive complexation of cocatalyst molecules with
active centers.
DOI: 10.1002/mats.200800088
New Approach in Modeling of Metallocene-Catalyzed Olefin Polymerization . . .
Table 1. Kinetic rate constants at different temperatures derived by the ANN model and the corresponding Arrhenius parameters.
Run no. Temperature ki kp kd
-C L �molS1 � sS1 L �molS1 � sS1 sS1
1 40 0.056 1.240 5.520T 10S5
2 60 0.041 2.321 4.650T 10S5
3 81 0.381 5.198 1.020T 10S4
4 90 0.454 6.200 1.020T 10S4
5 105 0.126 10.186 3.170T 10S4
6 120 0.297 12.697 6.520T 10S4
7 129 5.472 16.901 1.138T 10S3
Ea/(kJ �molS1) – 17.640 13.431 16.218
log k0 – 12.226 12.081 3.860
Conclusion
This paper uses ANNs in a new way to predict the
polymerization behavior, which does not need many
experimental samples for training. The polymerization
behavior can be simulated using only one experimental
run data. The model works independent of operational
conditions such as monomer and catalyst concentrations
and polymerization temperature, so one can predict the
rate constant Arrhenius parameters if polymerization
rates at different temperatures are available. The advan-
tage of using ANNs is that the training of the network is
done only once, and after that it could be used to calculate
rate constants of different catalysts under different
operational conditions. Adding the resulted molecular
weight information, this method also has the potential of
predicting chain transfer constants.
Received: November 10, 2008; Accepted: December 22, 2008; DOI:10.1002/mats.200800088
Keywords: artificial neural networks; catalysts; chains; olefinpolymerization; polymerization kinetics; polymerization modeling
[1] P. A. Charpentier, Ph.D. thesis, McMaster University, Hamil-ton 1997.
[2] W. J. Wang, D. Yan, S. Zhu, A. E. Hamielec, Macromolecules1998, 31, 8677.
[3] N. P. Khare, Ph.D. thesis, Virginia Polytechnic Institute andState University 2003.
[4] M. F. Bergstra, G. Weickert, Macromol. Mater. Eng. 2005, 290,610.
Macromol. Theory Simul. 2009, 18, 195–200
� 2009 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
[5] W. Kaminsky, F. Muller, O. Sperber, Macromol. Mater. Eng.2005, 290, 347.
[6] P. G. Belelli, M. L. Ferreira, M. H. Lacunza, D. E. Damiani, A.Brandolin, Poly. Eng. Sci. 2001, 41, 2082.
[7] R. A. Gonzalez-Ruiz, B. Quevedo-Sanchez, R. L. Laurence, M. A.Henson, E. B. Coughlin, AIChE J. 2006, 52, 1824.
[8] B. Quevedo-Sanchez, J. F. Nimmons, E. B. Coughlin, M. A.Henson, Macromolecules 2006, 39, 4306.
[9] Y. V. Kissin, Makromol. Chem. Macromol. Symp. 1993, 66, 83.[10] M. Ahmadi, M. Nekoomanesh, R. Jamjah, G. H. Zohuri, H.
Arabi, Macromol. Theory Simul. 2007, 16, 557.[11] Y. V. Kissin, R. I. Mink, T. E. Noelin, J. Polym. Sci., Part A: Polym.
Chem. 1999, 37, 4255.[12] V. Matos, A. G. Matos Neto, J. C. Pinto, J. App. Poly. Sci. 2001, 79,
2076.[13] V. Matos, A. G. Matos Neto, M. Nele, J. C. Pinto, J. App. Poly. Sci.
2002, 86, 3226.[14] G. C. Han-Adebekun, M. Hamba, W. H. Ray, J. Polym. Sci.,
Part A: Polym. Chem. 1997, 35, 2063.[15] M. Hamba, G. C. Han-Adebekun, W. H. Ray, J. Polym. Sci.,
Part A: Polym. Chem. 1997, 35, 2075.[16] Z. Gene Xu, S. Chakravarti, W. H. Ray, J. App. Poly. Sci. 2001, 80,
81.[17] S. Chakravarti, W. H. Ray, J. App. Poly. Sci. 2001, 80, 1096.[18] S. Chakravarti, W. H. Ray, J. App. Poly. Sci. 2001, 81, 2901.[19] S. Chakravarti, W. H. Ray, S. X. Zhang, J. App. Poly. Sci. 2001, 81,
1451.[20] B. Kou, K. B. McAuley, C. C. J. Hsu, D. W. Bacon, Macromol.
Mater. Eng. 2005, 290, 537.[21] B. Kou, K. B. McAuley, C. C. J. Hsu, D. W. Bacon, K. Z. Yao, Ind.
Eng. Chem. Res. 2005, 44, 2428.[22] K. Z. Yao, B. M. Shaw, B. Kou, K. B. McAuley, D. W. Bacon, Poly.
Reac. Eng. 2003, 11, 563.[23] G. Arzamendi, A. d’Anjou, M. Grana, J. R. Leiza, J. M. Asua,
Macromol. Theory Simul. 2005, 14, 125.[24] A. d’Anjou, F. J. Torrealdea, J. R. Leiza, J. M. Asua, G. Arzamendi,
Macromol. Theory Simul. 2003, 12, 42.[25] M. Hinchliffe, G. Montague, M. Willis, A. Burke, AIChE J. 2003,
49, 3127.
www.mts-journal.de 199
M. Ahmadi, M. Nekoomanesh, H. Arabi
200
[26] C. Kuroda, J. Kim, Neurocomputing 2002, 43, 77.[27] F. A. N. Fernandes, L. M. F. Lona, Pol. Reac. Eng. 2002, 10, 181.[28] F. A. N. Fernandes, L. M. F. Lona, A. Penlidis, Chem. Eng. Sci.
2004, 59, 3159.[29] T. Hanai, T. Ohki, H. Honda, T. Kobayashi, Comp. Chem. Eng.
2003, 27, 1011.
Macromol. Theory Simul. 2009, 18, 195–200
� 2009 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim
[30] K. Hornik, M. Stinchcombe, H. White, Neural Netw. 1989, 2,359.
[31] F. A. N. Fernandes, L. M. F. Lona, Brazil. J. Chem. Eng. 2005, 22,401.
[32] T. S. Wester, H. Johnsen, P. Kittilsen, E. Rytter, Macromol.Chem. Phys. 1998, 199, 1989.
DOI: 10.1002/mats.200800088