Materials and Manufacturing Processes

ISSN: 1042-6914 (Print) 1532-2475 (Online) Journal homepage: https://www.tandfonline.com/loi/lmmp20

Simulation and Optimization of Wiped-Film Poly-

Ethylene Terephthalate (PET) Reactor Using
Multiobjective Differential Evolution (MODE)

B. V. Babu , J. H. Syed Mubeen & Pallavi G. Chakole

To cite this article: B. V. Babu , J. H. Syed Mubeen & Pallavi G. Chakole (2007) Simulation and
Optimization of Wiped-Film Poly-Ethylene Terephthalate (PET) Reactor Using Multiobjective
Differential Evolution (MODE), Materials and Manufacturing Processes, 22:5, 541-552, DOI:
10.1080/10426910701319266

To link to this article: https://doi.org/10.1080/10426910701319266

Published online: 30 May 2007.

Submit your article to this journal

Article views: 139

View related articles

Citing articles: 1 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=lmmp20
Materials and Manufacturing Processes, 22: 541–552, 2007
Copyright © Taylor & Francis Group, LLC
ISSN: 1042-6914 print/1532-2475 online
DOI: 10.1080/10426910701319266

Simulation and Optimization of Wiped-Film Poly-Ethylene

Terephthalate (PET) Reactor Using Multiobjective
Differential Evolution (MODE)
B. V. Babu, J. H. Syed Mubeen, and Pallavi G. Chakole

Department of Chemical Engineering, Birla Institute of Technology and Science, Pilani, Rajasthan, India

The multiobjective differential evolution (MODE), which is an extension of the Differential Evolution (DE), is applied to solve the multiobjective
optimization problem (MOOP) of wet film Poly-Ethylene Terephthalate (PET) reactor considering minimization of acid end group and vinyl end
group as the main objectives. The objective function is modified to solve five different possible cases. The results show that a Pareto set (a set of
equally good solutions) is obtained for the cases when two of the parameters (residence time of the polymeric reaction mass,
, and the speed of
the wiped-film agitator, N  are considered as decision variables, unlike a unique solution obtained using Nondominated Sorting Genetic Algorithm
(NSGA). The Pareto optimal front provides wide-ranging optimal sets of operating conditions. And an appropriate set of operating conditions can
be selected based on the requirements of the user.

Keywords Differential evolution; Modeling; Multiobjective Differential Evolution (MODE); Optimization; Pareto optimal front; Polymerization;
Simulation.

1. Introduction suited and gaining popularity for MOO problems. A detailed

The majority of the research work and application in
the field of optimization considers a single objective,
although most of the natural world problems involve
a group conflicting objectives simultaneously. When an
optimization problem involves more than one objective
function, the task of finding one or more optimum
solutions is known as multiobjective optimization (MOO)
[1]. Different solutions may produce trade-offs (conflicting
scenarios) among different objectives.
As multiple objectives are simultaneously optimized in
multiobjective optimization, it may not be possible to have
a solution that is best with respect to all the objectives.
Hence, MOO results in a set of equally good solutions lying
on a front, which is known as Pareto optimal front [1]. A
Pareto optimal set of solutions is such that when we go from
any one point to another in the set, at least one objective
function improves and at least one other worsens [2]. The
choice of selecting a suitable solution from Pareto optimal
solutions depends upon the decision maker.
There are several conventional methods available in
the literature for solving MOO problems [1]. These
methods follow a preference-based approach, in which a
relative preference vector is used to scalarize multiple
objectives. Since these conventional optimization methods
use a point-by-point approach starting from an initial
guess, these methods result in a single optimal solution.
However, Evolutionary Algorithms (EAs) can find multiple
optimal solutions in one single simulation run due to their
population-based search approach. Thus, EAs are ideally
Received August 4, 2006; Accepted February 20, 2007
Address correspondence to B. V. Babu, Dean-EHD and Professor of
Chemical Engineering, Birla Institute of Technology and Science, Pilani
333 031, Rajasthan, India; E-mail: bvbabu@bits-pilani.ac.in
541
account of MOO using evolutionary algorithms and some
of the applications using genetic algorithms are reported in
literature [1, 3, 4].
MOO finds its application in many polymerization
processes due to their composite and complex nature. The
quality of the polymer produced is generally described
in terms of several experimentally measured properties
(e.g., stiffness, tenacity, strength). Mathematical models
of the reactors are available in literature which can
predict the molecular parameters [5]. These include the
average molecular weight, poly-dispersity index, degree
of short-chain branching, concentration and distribution of
functional groups, etc., which can be related to the properties
of polymers. The operating variables, which include
the design variables of a polymerization reactor system,
influence these ‘molecular’ measures of the product in
conflicting ways and hence MOO is of considerable interest
for polymerization systems. Poly-Ethylene Terephthalate
(PET) is one of the key thermoplastic polymers and is
extensively used in the manufacture of fibers, films, bottles,
etc. Optimal design and operation of the wiped PET
reactor are essential, as it is the critical reactor in PET
manufacturing process.
In most of the studies on PET reactor optimization
problem [6, 7], weight factor is introduced for each objective
function. MOO is reduced to single objective optimization
by summing up weighted objective functions. This approach
is not only inefficient but also has the drawback of
missing some of the optimal solutions in the cases where
non-convexity of the objective function gives rise to a
duality gap [1]. MOO of polymerization reactors is studied
by few researchers only [6, 8–11]. Pontryagin minimum
principle or a pattern search technique coupled with the
-constraint method is used to obtain Pareto optimal sets
of nondominant solutions for MOO of a copolymerization
542 B. V. BABU ET AL.

reactor. Sareen and Gupta [12] obtained optimal pressure
histories and optimal values of the jacket fluid temperature
at different points on the Pareto set for an industrial nylon-
6 semi batch reactor. An adapted version of the Genetic
Algorithm (GA) is used to carry out MOO studies on nylon-
6 and PMMA by Chakravarthy et al. [13], Mitra et al. [14],
Garg and Gupta [15], Garg et al. [16], and Gupta and Gupta
[17]. Pareto sets of nondominated solutions are generated
in all cases. Several researchers extended their study in the
filed of polymer development [18–25].
Bhaskar et al. [5] carried out the MOO of reactors
producing fiber-grade PET. They modified the model of
Laubriet et al. [26] and simulated an industrial wiped-film
reactor by ‘tuning’ the model parameters using three sets of
industrial data. MOO results showed that they obtained a
unique optimum solution (no Pareto set of several equally
good, nondominating points was obtained). Also, they found
that several minima in the decision-variable space were
present, where the values of the two objective functions
were almost identical for all practical purposes.
This result, though interesting, was somewhat unexpected
and so they attempted to explore this problem further to
have more insight into the problem [27]. They found that
Nondominated Sorting Genetic Algorithm (NSGA) fails to
give correct solutions in a single application, particularly
when several decision variables are used. One needs to study
several simpler optimization problems with fewer decision
variables, and also apply NSGA several times in order to
be able to obtain the correct solutions. However, there are
a few problems where NSGA could give Pareto front for
problems with fewer decision variables.
This literature survey spurred interest to consider a
different algorithm technique to be applied for MOO of PET
reactor. Differential Evolution (DE), which is an improved
version of GA, is found to give better results than simple
GA for single objective optimization problems [28–34].
DE has also been successfully applied to different variants
of Multiobjective Differential Evolution (MODE) to solve
MOO problems [35–41]. To the best of our knowledge,
this is the first attempt to study MOO of an industrial
PET reactor using the MOO. The same objectives taken
by Bhaskar et al. [27] were considered for comparison
purposes.
The paper is organized as follows: concepts about DE
and MODE are explained in Section 2. Optimization of
PET reactor is presented in Section 3 with discussion in
Section 4, and conclusions are drawn in Section 5.
to be adopted for a problem are to be determined by trial
and error. The crucial idea behind DE is a scheme for
generating trial parameter vectors. Basically, DE adds the
weighted difference between two population vectors to a
third vector. The key parameters of control in DE are: NP
or the population size, CR or the crossover constant, and F
or the weight applied to random differential (scaling factor).
Price and Storn [44] have given some simple rules for
choosing key parameters of DE for any given application.
Normally, NP should be about 5 to 10 times the dimension
(number of parameters in a vector) of the problem. As for F ,
it lies in the range 0.4 to 1.0. Initially F = 05 can be tried,
then F and/or NP is increased if the population converges
prematurely. As for CR, it lies in the range 0.1 to 1.0.
Babu et al. [45] proposed a new concept called nested DE
to automate the choice of DE key parameters. In addition,
some new strategies have been proposed and successfully
applied to optimization of the extraction process [46].
DE has been successfully applied in various fields.
Some of the successful applications of DE include: digital
filter design [47], batch fermentation process [48], global
optimization of mixed integer non linear programming
(MINLP) problems [29], optimization of biomass pyrolysis
[49], etc. Many engineering applications using various
evolutionary algorithms have been reported in literature
[1, 50–54].
2.1. MODE
Figure 1 describes the working principle of MODE
(MODE flow chart). The pseudo-code for MODE is

2. De
DE [42] is an improved version of GA [43] for faster
optimization. Unlike simple GA that uses binary coding
for representing problem parameters, DE is a simple yet
powerful population-based, direct search algorithm for
globally optimizing functions with real valued parameters.
Among the DE’s advantages are its simple structure, ease
of use, speed, and robustness. Price and Storn [42] gave the
working principle of DE and also suggested ten different
strategies [44]. A strategy that works out to be the best
for a given problem may not work well when applied for
a different problem. Also, the strategy and key parameters Figure 1.—Working principle of MODE.
SIMULATION AND OPTIMIZATION OF WIPED-FILM PET REACTOR
available in our earlier literature [41]. In each generation
of MODE, the dominated solutions are removed from the
list and only the nondominated solutions are allowed to
undergo DE operations. The scaling factor F is generated
from a random generator between 0 and 1. The offsprings
are placed into the population if they dominate the main
parent.
The algorithm works as follows. An initial population is
generated at random. All dominated solutions are removed
from the population. The remaining nondominated solutions
are retained for recombination. Three parent vectors are
selected at random. A child vector is generated from the
three parent vectors and is placed into the population if it
dominates the first parent vector; otherwise, a new selection
process takes place.
The stopping criteria may be of two kinds: 1) there is no
new solution added to the nondominated front for a specified
number of generations, or 2) assign an upper bound on
the number of generations. The stopping criteria may be a
combination of the two as well. However, in this study, the
second criteria is adopted.
3. Optimization of pet reactor
PET is manufactured in three continuous reactors using
Ethylene Glycol (EG) and Purified Terphthalic Acid (PTA).
Usually EG is taken in excess. In the first stage, esterification
is carried out in continuous stirred tank reactor (CSTR) or
plug flow reactor (PFR) at 270–280
C and at atmospheric
pressure. A poly-condensation catalyst, antimony trioxide,
is injected in small concentrations (0.03–0.05 wt.%) into the
oligomer stream leaving this reactor.
In the second stage, pre-polymerization is carried out to
get a Degree of Polymerization (DP) of 30–40. Agitated
vessel is used and maintained at a pressure of 15–30 mmHg
and 270–280
C. In the third stage (wiped-film reactor),
final condensation occurs at a pressure of 1–2 mmHg and
a temperature of 280–295
C. Optimization of third stage is
more important as it controls the final properties of product.
3.1. Modeling of Reaction Scheme
The major functional end groups present in the
polymerizing melt are the hydroxyl group, the carboxylic
acid group, the DEG group, and the unsaturated vinyl group.
An extensive redistribution of these functional end groups
occurs during the course of polymerization. A quantitative
description of the reaction kinetics becomes quite complex.
In developing a PET polymerization model, the first task is
to decide the reactions to be included in the kinetics scheme.
For example, a detailed molecular species model gives a
complete product composition distribution and molecular
weight distributions of various polymeric species; however,
the computational requirement is often extensive and such a
detailed model may not be suitable for the design of reactor
control systems [55–57]. To simplify the overall kinetic
scheme, functional group modeling approach is considered.
By using this approach results of model simulations can
also be validated easily through experimentation.
For the modeling of a continuous finishing melt poly-
condensation reactor, a functional group model that is
similar to the one used by Ravindranath and Mashelkar
543
[55, 58] is employed. The final set of reactions and
corresponding rate expressions are given in Appendix A.
Here, the rate constants are for the reactions between the
two reacting functional end groups, and the equal reactivity
hypothesis is assumed for these end groups. The main poly-
condensation reaction (A1) indicates that ethylene glycol
must be removed from the polymer melt in order to promote
the forward reaction and to produce the polymers of high
molecular weight. Note that reactions A1, A3, A5, and
A8 lead to the formation of ester linkages and reactions
A2, A4, A6, and A7 do not change the polymer chain
length. Reaction (A9) is the main degradation reaction,
which becomes important at high temperature.
3.2. Reactor Modeling
A horizontal vessel equipped with a screw-type or
rotating disk agitator is used as a reactor for continuous
finishing poly-condensation. Such agitators create polymer
films on the screw or disk surfaces continuously. These
films, after exposure to a bulk vapor phase, are mixed
with a bulk polymer melt [26]. Low molecular weight
PET pre-polymer from pre-polymerization reactor is fed to
the wiped-film reactor. Various volatiles such as ethylene
glycol, di-ethylene glycol (DEG), water, and acetaldehyde
are removed from the bulk melt phase by applying high
vacuum (e.g., 0.1–1.0 mmHg) and high temperature (e.g.,
270–300
C).
Bhaskar et al. [59] proposed the model for continuous
finishing stage melt poly condensation pf PET. The
assumptions considered by Bhaskar et al. [59] for proposing
a two-phase model are: 1) flow pattern of melt phase is of
plug flow type and that of vapor phase is of well mixed, 2)
no distinction between the film and bulk part in melt phase,
3) no reaction in vapor phase, and 4) mass transfer resistance
resides in only melt phase. The interfacial concentration of
the volatiles is determined by the vapor–liquid equilibrium
relation (i.e., the Flory–Huggins theory). The concept of the
two-phase reactor model is illustrated diagrammatically in
Fig. 2. The complete reactor model equations are given in
Bhaskar et al. [59].
3.3. Simulation of Wiped-Film PET Reactor
In the previous section, a short summary of the two-
phase model used [5] is given. An adaptation of the two-
phase reactor model of Laubriet et al. [26], as described
Figure 2.—Two-phase model for continuous melt polymerization.
544 B. V. BABU ET AL.

Table 1.—Feed conditions used for simulation.

Feed concentrations (kmol/m3 ) of pure liquid components

Eg f = 40 × 10−1 EDEG f = 017

Ea f = 257 × 10−3 EGf = 65 × 10−3
Zf = 112 W f = 46 × 10−4
Ev f = 117 × 10−3 DEGf = 40 × 10−4

by Bhaskar et al. [59] is used. The model equations can be

written in the form:
dx
= f x u xz = 0 = x0 (1)
dz
where x is the vector of state variables, defined by

x = Eg  Ea  Z Ev  EDEG  EG W  DEG (2)

and u is the vector of control or decision variables. In

Eq. (1), z represents the (dimensionless) axial position in
the wiped-film reactor.
The initial conditions required to solve the given ODEs
are given in Table 1. Values of the parameters/properties
for the reference case are given in Table 2. MATLAB Figure 3.—a) Prediction of industrial values of DPout with model values.
subroutine ODE 23s is used to solve these stiff equations, b) DP profile of polymer along the length, z.
based on our earlier successful applications [60, 61]. The
model provides values of DP and the concentrations of the
hydroxyl end groups (Eg ), acid end groups (Ea ), di-ester end here are same as those reported by Bhaskar et al. [27],
groups (Z, vinyl end groups (Ev ), DEG end groups (EDEG , so that comparison of MODE results could be done
EG, W , and DEG, as a function of the axial position in the with those obtained using NSGA. The acid end group
reactor. Prediction of industrial values of DP of the product makes the polymer susceptible to hydrolysis during the
with model values using the best-fit value of
= 26647 downstream operations and leads to breakage of the
shown in Fig. 3(a) and DP profile corresponding to set filaments during spinning, where the humidity is very high.
1 is given in Fig. 3(b). The concentration profiles for all The vinyl end groups have been shown to be responsible
the components along the length of the reactor are shown for the unfavorable coloration of PET [5]. Hence the
in Fig. 4. The simulation results are compared with the minimization of these two end groups are considered as
industrial data and the results are given in Table 3. the main objective. The reduction of Ea simultaneously
increases the rate of polymerization of the acid end
3.4. Objective Function Formulation group catalyzed poly-condensation reaction and helps to
In this study, minimization of two objectives, namely, maximize the throughput. Important end-point constraint is
acid end group concentration and vinyl end group to produce polymer with a desired value of the degree of
concentration are considered. The objectives considered polymerization DP. DEG end groups improve the dyeability
of fibre but it also affect the crystallinity unfavorably.
Therefore, it is preferred to have a certain allowable range
Table 2.—Values of the parameters/
for the concentration of DEG end groups in the fiber-grade
properties for the reference case. polyesters (an inequality constraint). In addition, a further
inequality constraint on the maximum allowable limit for
Parameters Values the acid end group concentration is imposed to ensure that it
k1 aref 2.6875
is not only minimized but also lies below an upper limit. The

2.6647 objective function formulation is discussed in the following
ao 1.0378 subsection and optimization is done using MODE.
bo 2.1838 Five decision variables are used for optimization in this
EG (kg/m3 ) 1108 study. These are the reactor pressure P, temperature
W (kg/m3 ) 1000
DEG (kg/m3 ) 1118
(isothermal, T , catalyst concentration (Sb2 O3 ), residence
P (mmHg) 2.00∗ , 1.5∗∗∗ time of the polymeric reaction mass inside the reactor
T (K) 566 , 567∗∗
∗ (
), and the speed of the wiped-film agitator N . All of
∗
these variables can easily be changed in any industrial,
Set 1.
∗∗
Set 2. wiped-film reactor for PET manufacture, including the
∗∗∗
Set 3. one being studied, and are therefore used to obtain the
SIMULATION AND OPTIMIZATION OF WIPED-FILM PET REACTOR 545

Subject to
DPout = DPd (4)
−3
Ea out ≤ 1038 × 10 3
kmol/m (5)
01660 ≤ EDEG out ≤ 017 kmol/m 3
(6)
dx
= f x u xz = 0 = x0 (7)
dz
005 ≤ P ≤ 0266 kPa 04 ≤ P ≤ 20 mm Hg (8)
564 ≤ T ≤ 570 K (9)
003 ≤ Sb2 O3  ≤ 0045 wt.% (10)
∗
09 ≤
≤ 106 (11)
∗
093 ≤ N ≤ 105 (12)

The variables,
∗ and N ∗ , represent dimensionless values,

/
ref and N /Nref , where
ref and Nref are values being
used currently in the industrial reactor being studied. These
two values are kept confidential in the literature and hence
the value of
is found by a trial-and-error method. The
value that satisfies with the industrial value is taken to be

ref and it is found to be 99 minutes. It may be noted
that Nref is not directly required but it is the range of N ∗
=N /Nref , which is required for simulation of the reactor
model. The range of N ∗ is taken to be 0.93–1.05 as reported
in the literature [59].
Meaningful bounds have been chosen on the five decision
variables, u, based on industrial practice. These are given in
Eqs. 8–12. The catalyst concentration as a decision variable
for the optimization of stage 3 is as it affects the operation
more significantly used even though the catalyst is added
in the form of the glycolate at the inlet of stage 2 [5].
Catalyst concentration is equilibrium-controlled [58] and is
Figure 4.—Concentration profiles along the length of reactor: a) Hydroxyl influenced primarily by the pressure and temperature used
end group ([Eg ]); b) Acid end group ([Ea ]); c) Ester Linkages ([Z]); d) Vinyl in that reactor. But as the data is available only for catalyst
end group ([Ev ]); e) Diethylene glycol end group ([EDEG ]); f) Ethylene glycol concentration of 0.04%, the catalyst concentration is kept
([EG]); g) Water ([W ]) h) Free DEG ([DEG]). constant and not used as a decision variable in the study. The
use of the residence time,
∗ , as a decision variable for the
film reactor while keeping the feed composition unchanged,
best optimal operating conditions. The MOO problem is can also be justified. It is assumed that a change in the
described mathematically by: residence time of the film reactor could be achieved by
changing the flow rate of the reaction mass in this reactor,
IP T  Sb2 O3 
∗  N ∗  ≡ I1  I2 T rather than by changing the holdup volume of the melt in this
Minimize
T reactor. Since the first- as well as the second-stage reactors
= Ea out  Ev out has considerable flexibility in industry, and since both these
stages are equilibrium-controlled, a change in the flow rate
(3) does not necessarily lead to any changes in the composition
of the feed to the film reactor. The choice of the four
decision variables and the use of constant feed conditions
Table 3.—Comparison of the simulation results with the reported data. are thus justified and give a great deal of flexibility for
Set Product Industrial Model predicted value
optimization, which can easily be implemented in industry.
no. property value (
= 26647) The end-point constraint on DP Eq. (4) is incorporated
in the form of a penalty function with a large value of the
1 DP 82.00 82.00 weightage factor, w1 (106 in this study), as
EDEG  (kmol/m3 ) 0.17 0.167
Ea  (kmol/m3 ) 1038 × 10−3 1028 × 10−3  2
2 DP 82.60 82.53 DPout
3 DP 82.70 82.65 Min I1∗ = Ea out + w1 1− + Pe (13)
DPd
546 B. V. BABU ET AL.

Table 4.—Computational variables used in this study.
Parameter Values
Number of population 100
Number of generations 600
Scaling factor (F  Using random generator
Crossover Constant (CR) 0.9
An additional large “penalty” value, Pe (104 in this study),
is added to the objective function, I1∗ , if either of the two
inequality constraints in Eqs. (5) and (6) are violated. If
these constraints are satisfied, then the value of Pe in
Eq. (13) is taken as zero. This ensures that bad points in
the MODE used for optimization are not reproduced in
the successive generations. Also, a very large value of Pe
is added if the concentrations of three of the vaporizing
species (EG, W , and DEG) in the vapor phase are higher
than their respective vapor-liquid interfacial equilibrium
concentrations (EG∗ , W ∗ , and DEG∗ ) at any axial location,
in order to ensure the vaporization rather than condensation
of these species at all axial locations. Minimization of I1∗
leads to a decrease in the acid end group concentration while
simultaneously giving reference to solutions satisfying the
several requirements just discussed.
The second objective function, I2∗ , is also modified to
incorporate the penalties for the several constraints as
discussed for I1∗ , and is given by
 2
DPout
Min I2∗ = Ev out + w2 1− + Pe
DPd
case for the industrial reactor simulated earlier. Such a study
would help to explore whether changes in the operating
variables could improve the performance of the industrial
reactor. The MOO problem described in Eqs. (3)–(12) is
solved here.
Figure 5(a) shows the Pareto optimal set obtained
after 600 generations when the two objective functions
considered are Eaout and Evout [Eq. (3)]. Each point on the
Pareto set corresponds to a set of decision variables given
in Figs. 5(b)–(e), where each of these decision variables is
plotted against one of the objectives, Eaout . From Fig. 5(a)
it is evident that the proposed algorithm is able to give a
Pareto optimal front and not a unique solution as reported
by Bhaskar et al. [27] using NSGA. NSGA does not take
into account the dealing with crowding and hence diversity
of population is not accounted for as the generations
increase, and hence it might have converged to a single
solution. However, MODE has inherent capability (due
to its better recombination, i.e., combined mutation and
crossover operation) of giving a better spread among the
Pareto points. The Pareto optimal front thus provides the
user a set of solutions to choose from. Figures 5(b) and (c)
reveals that the optimum value for the decision variable T
is nearly constant, and for decision variable P it is more
scattered. The decision variables are also plotted against
objective function Evout to gain further understanding of
(14)

Minimization of the second objective function leads to a

minimum value of the vinyl end group concentration in
the product, thus reducing the product degradation due to
thermal effects and improving the color of the polymer.
The MOOP discussed above is solved using the proposed
algorithm MODE, the computational variables of which are
given in Table 4. The optimization results are discussed in
the next section.

4. Results and discussion

The solution of the MOO problem described in Eq. (3)
is obtained using MODE, as proposed by our group using
values of the computational parameters as given in Table 4.
A software package is developed in MATLAB 6.1. The
computer code was tested for a few simple two-objective
function problems taken from literature [1, 51] as well
as on some other problems, such as the optimization of
industrial styrene reactor and other complex test problems
[39–41, 62] to ensure the correctness of the program. One
of the optimal points obtained from the code was used
in our earlier simulation code. The same output variables
are obtained from this as from the optimization code, thus
confirming that the developed software package is error
free.

4.1. Case 1: With T , P,
∗ , and N ∗ as Decision Variables

The proposed code (MODE) for the MOO was first run Figure 5.—Optimization results (comparison of Pareto fronts), with decision
for a DPd of 82.0, the value corresponding to the reference variables compared with Eaout (Case 1).
SIMULATION AND OPTIMIZATION OF WIPED-FILM PET REACTOR 547

As temperature does not cause any conflict among the

objectives, the MOO problem is modified as described in
Eqs. (15)–(20).

IP
∗  N ∗  ≡ I1  I2 T
Minimize
T
= Ea out  Ev out (15)
Subject to Eqs. (4)–(7) and
T = C1 (16)
Sb2 O3  = 004 wt.% (17)
005 ≤ P ≤ 0266 kPa 04 ≤ P ≤ 20 mm Hg (18)
∗
09 ≤
≤ 106 (19)
∗
093 ≤ N ≤ 105 (20)

In this case both temperature and catalyst concentration

Figure 6.—Optimization results (comparison of Pareto fronts), with decision
are taken to be constant values and not included in the
variables compared with Evout (Case 1).
objective function formulation as decision variables. This
was done to develop further insight into the MOO problem.
The bound on rest of the decision variables are taken to
be same. Here we have considered two constant values for
the problem. As seen in Figs. 6(a), (b), the temperature temperature T . One is the optimal value that we got in the
and pressure profiles are similar to those seen in Figs. 5(b), previous run (564.02 K) and the other one is the reference
(c), thus they don’t conflict with the objective functions value (566 K).
considered. Figures 5(d), (e) and 6(c), (d) shows that
∗ and As expected each constant value of T gave different
N ∗ have a strong opposing effect on Eaout and Evout . The Pareto front. Figure 7(a) shows the Pareto optimal set
value of both
∗ and N ∗ , when increased, decreases Eaout obtained after 600 generations for temperature value of
but increases Evout . Table 5 compares objective functions 564.02 K (Optimal value obtained in Case1). The algorithm
(Eaout and Evout , decision variables (T , P,
∗ and N ∗ ) should reproduce the same Pareto front, as the same optimal
for three selected points, A, B, and C [shown in Fig. 5(a)]. value of 564.02 K (obtained in Case 1) is used as a constant
Table 5 as well as Fig. 5(a) reveals that as we move from input value for this case. From Fig. 7(a) it is evident that the
point A to point C, Eaout increases while Evout decreases Pareto optimal front matches exactly with those obtained
primarily due to decrease of
∗ and N ∗ . High
∗ and N ∗ using temperature as the decision variable [Figs. 5(c) and
minimizes acid end group produced Eaout , while low 6(b)]. This confirms that MODE is able to reproduce the

∗ and N ∗ minimizes vinyl end group Evout  produced. same Pareto front, which was not the case with NSGA, as
The optimal results also show that the decision variable, T reported by Bhaskar et al. [27].
[Figs. 5(c) and 6(b)], hits the lower bound and is almost Figure 7(b) shows the Pareto optimal set obtained after
constant at 564.02 K. As both the reactions (A2 and A9) 600 generations for temperature value of 566 K (reference
are degradation reactions, increase in temperature increases value). The result clearly suggests that the Pareto front is
production of Eaout and Evout , thus the temperature is found not as diverse and it is always above that obtained using T
to be at its lower bound. value of 564.02 K. Thus temperature value of 564.02 K is
best suited for industrial use rather than 566 K. Figures 7(c)–
4.2. Case 2: With P,
∗ , and N ∗ as the Decision Variables (e) shows the effect of other decision variables with respect
to Eaout . Again it is found that
∗ and N ∗ are responsible
An interesting point suggested in Figs. 5 and 6 is for producing the Pareto-optimal front.
that the value of the decision variable temperature T 
remains constant and it is at its lower bound (564.02 K). 4.3. Case 3: With
∗ and N ∗ as the Decision Variables
In order to gain further understanding into the problem
and to check the robustness of the code developed, the MOO
Table 5.—Comparison of operating conditions and objective function (Case is further modified. As
∗ and N ∗ are the main decision
1) values corresponding to points A, B, and C in Fig. 5(a) with results reported
variables that are responsible for producing Pareto optimal
by Bhaskar et al. (2001).
front, we tried to simplify the formulation with only these
Parameter Point A Point B Point C Bhaskar et al. (2001) two as the decision variables. For pressure reference value
(2.0 mmHg) is taken, and for the temperature the optimal
Eaout (mol/m3 ) 0.969 0.984 1.004 1.0 value (564.02 K from case 1) is chosen. Now the MOO
Evout (mol/m3 ) 0.776 0.771 0.765 0.755
T (K) 564.02 564.02 564.02 564.01
problem can be defined as:
P (mm Hg) 1.51 1.64 1.187 1.39

∗ 
/
ref  0.986 0.963 0.948 0.967 I
∗  N ∗  ≡ I1  I2 T
N ∗ N /Nref  1.034 0.998 0.946 1.0 Minimize = Ea out  Ev out T (21)
548
Figure 7.—Optimization results for Case 2.
Subject to Eqs. (4)–(7), (6), (17) and
P = 20 mm Hg (22)
09 ≤
∗ ≤ 106 (23)
093 ≤ N ∗ ≤ 105 (24)
Since both the variables
∗ and N ∗ , which are responsible
for producing Pareto front, are considered, the algorithm
should give Pareto set of solutions. Figure 8(a) shows the
Pareto front obtained after 600 generations. As reference
value of pressure (2 mmHg) is considered, the Pareto front
is slightly above than, that reported in case 1. This also
underlies the fact that the reference value of pressure is not
the optimal value for operation of PET reactor. The trend
followed by the two decision variables considered, are same
as seen in the previous cases (cases 1 and 2) [Figs. 8(b), (c)].
4.4. Case 4: Only with P as the Decision Variable
The importance of
∗ and N ∗ in producing a Pareto set
of solutions is highlighted in this case. Here all the values
except pressure are taken to be constant. The bounds on the
pressure are kept same, while optimal value of temperature
is taken (564.02 K), and reference values are taken for
∗
and N ∗ . The reason for considering only pressure as the
decision variable is to study how it affects the operation of
PET reactor when other variables are considered constant.
B. V. BABU ET AL.
Figure 8.—Optimization results for Case 3.
From cases 1 and 2 it is evident that pressure plays an
important role in the optimization as its effect on the
objective function is scattered and not constant one. The
multi-objective optimization problem is thus formulated as:

T
Minimize IP ≡ I1  I2 T = Ea out  Ev out (25)
Subject to Eqs. (4)–(7), (17) and
005 ≤ P ≤ 0266 kPa 04 ≤ P ≤ 20 mm Hg (26)
∗

=1 (27)
N ∗ = 1 (28)
The results are found to be interesting. Figure 9(a) shows
that when
∗ and N ∗ are considered constant, there is no
Pareto front, but a unique solution is obtained. Also the
values of the pressure are found to be almost same, around
0.6 mm Hg and not scattered [Fig. 9(b)]. These results show
that the effect of pressure is dependent on the speed of
rotation of agitator. If the speed is kept constant at reference
value, single optimum value of pressure is obtained and
not scattered one. Thus the effect of pressure on producing
Pareto front depends on the agitator speed, and when
∗ and
N ∗ are constant there is no Pareto front produced.
SIMULATION AND OPTIMIZATION OF WIPED-FILM PET REACTOR 549

Figure 10.—Optimization results for Case 5.

Figure 9.—Optimization results for Case 4.

Table 6.—Summary of the optimization results for PET reactor corresponding

to point A in Figs. 5(a), 7(a), 8(a), 9(a), 10(a).
4.5. Case 5: With T and P as the Decision Variable
The importance of reference time and reference speed Parameter Industrial Case 1 Case 2 Case 3 Case 4 Case 5
values in producing a Pareto set of solutions is further Eaout (mol/m ) 3
1.038 0.969 0.968 0.973 0.977 1.0
stressed by considering only T and P as decision variables. Evout (mol/m3 ) 0.78 0.776 0.776 0.776 0.775 0.778
The modified objective function is given below: T (K) 566 564.02 564.02∗ 564.02∗ 564.02∗ 565.07
P (mm Hg) 2.0 1.51 1.044 2.0∗∗ 0.576 1.08

∗ 
/
ref  1 0.986 0.993 0.984 1.0∗∗ 1.0∗∗
IT  P ≡ I1  I2 T N ∗ N /Nref  1 1.034 1.03 1.044 1.0∗∗ 1.0∗∗

Minimize
T DP 82.0 82.0 82.0 82.0 82.0 82.0

= Ea out  Ev out
Subject to Eqs. (4)–(7), and
005 ≤ P ≤ 0266 kPa 04 ≤ P ≤ 20 mm Hg
564 ≤ T ≤ 570 K

∗ = 1
∗
N = 1
The results, as expected, produced a single optimal
solution. There are no trade-off solutions seen, the unique
solution obtained is shown in Fig. 10(a). The solution
obtained in this case is slightly different from the one
obtained in case 4, also a constant temperature value of
565 K is obtained and not 564.02 K. This discrepancy in
the value of temperature is because reference values are
considered for
∗ and N ∗ , which are different from the
optimal values obtained in case 1. Figure 10(b) shows the
effect of pressure on the objective function Eaout . The
values are again found to be almost constant as in case 4,
but a higher value (1.08 mm Hg) is obtained. Table 6 gives
a summary of the optimization results for PET Reactor for
all the cases considered. For all the cases, the values of
Eaout and Evout are less than those reported by the industrial
data [59].
(29) ∗
Optimal value from case 1.
∗∗
Reference value (Bhaskar et al., 2001).
(30)
5. Conclusions
(31)
In this paper, DE approach is presented for solving
(32) the MOO problems. The proposed algorithm (MODE)
(33) is used to find the optimal operating conditions for the
manufacture of PET. Several MOO problems for PET
reactors were formulated, and then solved by MODE.
The results are found to be extremely encouraging as
Pareto set was obtained and not as a single solution as
reported by Bhaskar et al. [27]. The Pareto set and optimal
operating conditions for the first three combinations are
same. Following important points are also noted in this
study: Pareto front obtained was due to the conflicting
effect of decision variables
∗ and N ∗ (cases 1, 2, and 3),
temperature is found to be at its lower bound at 564.02 K
and pressure is found to be scattered. The trends followed
by decision variables are same and could be explained
qualitatively, which proves the reliability of the results
obtained. The results also show that the value of the
objectives could be improved further.
Appendix A
The final set of reactions are the following.
550 B. V. BABU ET AL.

Reaction (A1): Ester Interchange Reaction (Main Poly- References

condensation) 1. Deb, K. Multi-objective Optimization Using Evolutionary
Algorithms; Wiley: New York, 2001.
2. Yee, A.K.Y.; Ray, A.K.; Rangaiah, G.P. Multi-objective
optimization of industrial styrene reactor. Computers and
Chemical Engineering 2003, 27, 111–130.
Reaction (A2): Acetaldehyde Formation (a) 3. Rajesh, J.; Gupta, S.K.; Rangaiah, G.P.; Ray, A.K.
Multiobjective optimization of steam reformer using genetic
algorithm. Industrial and Engineering Chemistry Research 2000,
39, 706–717.
4. Srinivas, N.; Deb, K. Multiobjective function optimization
Reaction (A3): Acetaldehyde Formation (b) using non-dominated sorting genetic algorithms. Evolutionary
Computation 1995, 2, 221–248.
5. Bhaskar, V.; Gupta, S.K.; Ray, A.K. Modeling of an industrial
wiped film poly ethylene terephthalate reactor. Polymer Reaction
Engineering 2001, 9 (2), 71–99.
6. Farber, J.N. In Handbook of Polymer Science and Technology
1(429); Cheremisinoff, N. P.; Ed; Marcel Dekker: New York,
1989.
Reaction (A4): Diethylene Glycol Formation (a) 7. Tieu, D.; Cluett, W.R.; Penlidis, A. Optimization of
polymerization reactor operation: review and case studies with
the end-point collocation method. Polymer Reaction Engineering
1994, 2, 275.
8. Tsoukas, A.; Tirrell, M.V.; Stephanopoulos, G. Multiobjective
Reaction (A5): Diethylene Glycol Formation (b) dynamic optimization of semibatch copolymerization reactors.
Chemical Engineering and Science 1982, 37, 1785.
9. Farber, J.N. Steady state multiobjective optimization of
continuous copolymerization reactors. Polymer Engineering and
Science 1986, 26, 499.
10. Butala, D.; Choi, K.Y.; Fan, M.K.H. Multiobjective dynamic
optimization of a semibatch free-radical copolymerization
process with interactive CAD tools. Computers and Chemical
Reaction (A6): Diethylene Glycol Formation (c)
Engineering 1988, 12, 1115.
11. Choi, K.Y.; Butala, D.N. An experimental-study of
multiobjective dynamic optimization of a semibatch
copolymerization process. Polymer Engineering and Science
1991, 31, 353.
12. Sareen, R.; Gupta, S.K. Multiobjective optimization of an
industrial semibatch nylon 6 reactor. Journal of Applied Polymer
Reaction (A7): Water Formation (a) Science 1995, 58, 2357.
13. Chakravarthy, S.S.S.; Saraf, D.N.; Gupta, S.K. Use genetic
algorithms in the optimization of free radical polymerizations
exhibiting the Trommsdorff effect. Journal of Applied Polymer
Science 1997, 63, 529.
14. Mitra, K.; Deb, K.; Gupta, S.K. Multiobjective dynamic
Reaction (A8): Water Formation (b)
optimization of an industrial nylon 6 semibatch reactor using
genetic algorithm. Journal of Applied Polymer Science 1998,
69, 69.
15. Garg, S.; Gupta, S.K. Multiobjective optimization of a free
radical bulk polymerization reactor using genetic algorithm.
Macromolecular Theory of Simulation 1999, 8, 46.
16. Garg, S.; Gupta, S.K.; Saraf, D.N. On-line optimization of free
Reaction (A9): Degradation of Diester Group radical bulk polymerization reactors in the presence of equipment
failure. Journal of Applied Polymer Science 1999, 71, 2101.
17. Gupta, R.R.; Gupta, S.K. Multiobjective optimization of an
industrial nylon-6 semibatch reactor system using genetic
algorithm. Journal of Applied Polymer Science 1999, 73,
729.
18. Raha, S.; Majumdar, S.; Mitra, K. Effect of caustic addition
in epoxy polymerization process: a single and multiobjective
SIMULATION AND OPTIMIZATION OF WIPED-FILM PET REACTOR
evolutionary approach. Macromolecular Theory and simulations
2004, 13, 152–161.
19. Deb, K.; Mitra, K.; Dewri, R.; Majumdar, S. Towards a
better understanding of epoxy polymerization process using
multiobjective evolutionary computation. Chemical Engineering
Science 2004, 59 (20), 4261–4277.
20. Mitra, K.; Majumdar, S.; Raha, S. Multi-objective optimization
of a epoxy polymerization process using the elitist genetic
algorithm. Industrial and Engineering Chemistry Research 2004,
43 (19), 6055–6063.
21. Mitra, K.; Majumdar, S.; Raha, S. Multi-objective dynamic
optimization of epoxy polymerization process. Computer and
Chemical Engineering 2004, 28 (12), 2583–2594.
22. Majumdar, S.; Mitra, K.; Raha, S. Optimized species growth in
epoxy polymerization with real coded NSGA II. Polymer 2005,
46, 11858–11869.
23. Majumdar, S.; Mitra, K.; Sardar, G. Kinetic analysis and
optimization for the catalytic for the catalytic esterification step
of PPT polymerization. Macromolecular Theory and Simulations
2005, 14, 49–59.
24. Mitra, K.; Majumdar, S. Selection of catalyst for the esterification
step of the PPT polymerization process. Macromolecular Theory
and Simulations 2006, 15, 497–506.
25. Kasat, R.B.; Ray, A.K.; Gupta, S.K. Applications of genetic
algorithm in polymer science and engineering. Materials and
Manufacturing Processes 2003, 18 (3), 523–532.
26. Laubriet, C.; LeCorre, B.; Choi, K.Y. Two-phase model for
continuous final stage melt poly-condensation of poly ethylene
terephthalate: 1. steady-state analysis. Industrial and Engineering
Chemistry Research 1991, 30, 2.
27. Bhaskar, V.; Gupta, S.K.; Ray, A.K. Multi-objective
optimization of an industrial wiped film poly(ethylene
terephthalate) reactor: some further insights. Computers and
Chemical Engineering 2001, 25, 391–407.
28. Babu, B.V.; Sastry, K.K.N. Estimation of heat transfer
parameters in a trickle bed reactor using differential
evolution and orthogonal collocation. Computers and Chemical
Engineering 1999, 23, 327–339.
29. Babu, B.V.; Angira, R. A differential evolution approach for
global optimization of MINLP problems. Proceedings of 4th
Asia-Pacific Conference on Simulated Evolution and Learning
(SEAL’02), Singapore, November 18–22, 2002; Vol. 2, 880–884.
30. Babu, B.V.; Angira, R. Optimal Design of an Auto-
thermal Ammonia Synthesis Reactor. Computers and Chemical
Engineering 2005, 29 (5), 1041–1045.
31. Babu, B.V.; Chakole, P.G.; Mubeen, J.H.S. Differential evolution
strategy for optimal design of gas transmission network. Journal
of Multidisciplinary Modeling in Materials and Structures 2005,
1 (4), 315–328.
32. Babu, B.V.; Angira, R. Modified differential evolution (MDE)
for optimization of non-linear chemical processes. Computers
and Chemical Engineering 2006, 30 (6–7), 989–1002.
33. Angira, R.; Babu, B.V. Optimization of process synthesis and
design problems: a modified differential evolution approach.
Chemical Engineering Science 2006, 61 (14), 4707–4721.
34. Babu, B.V.; Munawar, S.A. Differential evolution strategies
for optimal design of shell-and-tube heat exchangers. Chemical
Engineering Science 2007 (In Press).
35. Abbass, H.A. The self-adaptive pareto differential evolution
algorithm. Proceedings of Congress on Evolutionary
Computation (CEC-2002) 1, IEEE Press, 2002; 831–836.
551
36. Abbass, H.A.; Sarker, R.; Newton, C. PDE: a pareto-frontier
differential evolution approach for multi-objective optimization
problems. Proceedings of IEEE Congress on Evolutionary
Computation, 2001, 971–978.
37. Madavan, N.K. Multi-objective optimization using a pareto
differential evolution approach. Proc. of Congress on
Evolutionary Computation (CEC-2002), IEEE Press, 2002; Vol.
2, 1145–1150.
38. Xue, F.; Sanderson, A.C.; Graves, R.J. Pareto-based
multiobjective differential evolution. Proceedings of the 2003
Congress on Evolutionary Computation (CEC-2003), IEEE
Press, 2003; Vol. 2, 862–869.
39. Babu, B.V.; Jehan, M.M.L. Differential evolution for
multiobjective optimization. Proceedings of 2003 Congress on
Evolutionary Computation (CEC-2003), Canberra, Australia,
December 8–12, 2003; 2696–2703.
40. Babu, B.V.; Chakole, P.G.; Mubeen, J.H.S. Multiobjective
differential evolution (MODE) for optimization of adiabatic
styrene reactor. Chemical Engineering Science 2005, 60 (17),
4822–4837.
41. Babu, B.V.; Mubeen, J.H.S; Chakole, P.G. Multiobjective
optimization using differential evolution. TechGenesis-The
Journal of Information Technology 2005, 2 (2), 4–12.
42. Price, K.; Storn, R. Differential evolution — a simple evolution
strategy for fast optimization. Dr. Dobb’s Journal 1997, 22,
18–24 and 78.
43. Goldberg, D.E. Genetic Algorithms in Search, Optimization, and
Machine Learning; Addison-Wesley: MA, 1989.
44. Price, K.; Storn, R. Home page of Differential Evolution as on
July, 2006. The URL of which is: http://www.ICSI.Berkeley.edu/
∼storn/code.html.
45. Babu, B.V.; Angira, R.; Nilekar, A. Optimal design of
an auto-thermal ammonia synthesis reactor using differential
evolution. Proceedings of The Eighth World Multi-Conference
on Systemics, Cybernetics and Informatics (SCI-2004), Orlando,
Florida, USA, July 18–21, 2005.
46. Babu, B.V.; Angira, R. New strategies of differential
evolution for optimization of extraction process. Proceedings
of International Symposium and 56th Annual Session of IIChE
(CHEMCON-2003), Bhubaneswar, December 19–22, 2003.
47. Storn, R. Differential evolution design of an IIR-
filter with requirements for magnitude and group delay.
Technical report TR-95–026, ICSI, May 1995. Available:
http://www.icsi.berkeley.edu/cgi-bin/pubs/publication.pl? ID =
000966 (Accessed: April 2007).
48. Chiou, J.P.; Wang, F.S. Hybrid method of evolutionary
algorithms for static and dynamic optimization problems with
application to a fed-batch fermentation process. Computers and
Chemical Engineering 1999, 23, 1277–1291.
49. Babu, B.V.; Chaurasia, A.S. Optimization of pyrolysis of
biomass using differential evolution approach. Proceedings
of The Second International Conference on Computational
Intelligence, Robotics, and Autonomous Systems (CIRAS-2003),
Singapore, December 15–18, 2003.
50. Onwubolu, G.C.; Babu B.V. New Optimization Techniques in
Engineering; Springer-Verlag: Heidelberg, Germany, 2004.
51. Babu, B.V. Process Plant Simulation; Oxford University Press
India, 2004.
52. Angira, R.; Babu, B.V. Multi-objective optimization using
modified differential evolution (MDE). International Journal
of Mathematical Sciences: Special Issue on Recent Trends in
552
Computational Mathematics and Its Applications 2006, 5 (2),
371–387.
53. Angira, R.; Babu, B.V. Performance of modified differential
evolution for optimal design of complex and non-linear chemical
processes. Journal of Experimental and Theoretical Artificial
Intelligence, 2006, 18 (4), 501–512.
54. Babu, B.V. Improved differential evolution for single- and multi-
objective optimization: MDE, MODE, NSDE, and MNSDE. In
Advances in Computational Optimization and Its Applications;
Deb, K., Chakroborty, Iyengar, P.N.G.R., Gupta, S.K. Eds.;
Universities Press: Hyderabad, 2007; 24–30.
55. Ravindranath, K.; Mashelkar, R.A. Finishing stages of PET
synthesis: a comprehensive model. American Institute of
Chemical Engineers Journal 1984, 30, 415–423.
56. Kumar, A.; Sharma, S.N.; Gupta, S.K. Optimization of
polycondensation step of PET formation with continuous
removal of condensation products. Polymer Engineering Science
1984, 24, 1205.
57. Lei, G.D.; Choi, K.Y. A melt polymerization of poly(ethylene
terephthalate) in semibatch stirred reactors. Journal of Applied
Polymer Science 1990, 41, 2987–3024.
B. V. BABU ET AL.
58. Ravindranath, K.; Mashelkar, R.A. Modeling of polyethylene
terephthalate reactors: 5. A continuous prepolymerization
process. Polymer Engineering Science 1982, 22,
610–618.
59. Bhaskar, V.; Gupta, S.K.; Ray, A.K. Multi-objective
optimization of an industrial wiped film PET reactor. American
Institute of Chemical Engineering Journal 2000, 46 (5),
1046–1058.
60. Babu, B.V.; Chaurasia, A.S. Dominant design variables in
pyrolysis of biomass particles of different geometries in
thermally thick regime. Chemical Engineering Science 2004, 59,
611–622.
61. Babu, B.V.; Chaurasia, A.S. Heat transfer and kinetics in the
pyrolysis of shrinking biomass particle. Chemical Engineering
Science 2004, 59, 1999–2012.
62. Babu, B.V.; Gujarathi, A.M.; Katla, P.; Laxmi, V.B.
Strategies of Multi-Objective Differential Evolution (MODE)
for Optimization of Adiabatic Styrene Reactor. Proceedings of
International Conference on Emerging Mechanical Technology-
Macro to Nano (EMTMN-2007), February 16–18, 2007, BITS
Pilani; 243–250.