optimal design of

multiproduct chemical
industries

Name: Ahmed Gamal Ibrahim Saber

Supervisor
 DR: Haitham Abd-Elsamad

Content
1. Optimal Design of Multiproduct Batch Plants
under Imprecise Demand

1.1 Introduction
1.2 Design Approaches under Uncertainty
1.3 A Fuzzy Decision-Making Approach for
Multiproduct Batch Plant Design
1.4 Typical Results
1.5 Conclusions
2. Optimal design of a multi-product reactive
distillation system for silences production .

2.1 introduction
2.2 Design of the silences process
2.3 Initial design and search variables for each
product

2.4 Concluding remarks

1- Optimal Design of Multiproduct Batch Plants
under Imprecise Demand

1.1 introduction

There has been an increased interest in the design of batch plant in

recent years due to the growth of specialty chemicals, food products
and pharmaceutical industries. In the conventional optimal design of a
multiproduct batch chemical plant, a designer speci- fies the
production requirements for each product and total production time.
The number, required volume and size of parallel equipment units in
each stage are to be determined in order to minimize the investment
cost . However, it is often the case that no precise product demand
predictions are available at the design stage .
The market demand for such products is usually changeable, and at
the stage of conceptual design of a batch plant, it is almost impossible
to obtain the precise infor- mation on the future product demand over
the lifetime of the plant. Nevertheless, decisions must be taken on the
plant capacity. This capacity should be able to balance the product
demand satisfaction and extra plant capacity in order to reduce the
loss on the excessive investment cost or than on market share due to
the varying product demands .
The key point in the optimal design of batch plants under
imprecision concerns modeling of demand variations. The most
common approaches treated in the dedi- cated literature represent
the demand uncertainty with a probabilistic frame by means of
Gaussian distributions. Yet, this assumption does not seem to be a
reliable repre- sentation of the reality, since in practice the parameters
are interdependent, leading to very hard computations of conditional
probabilities, and do not follow symmetric distribution rules. In this
work, fuzzy concepts and arithmetic constitute an alternative to
describe the imprecise nature on product demands.
In this study, we will introduce a fuzzy decision-making approach to
solve the fuzzy goal optimization problem. For this purpose, we
extended a multiobjective genetic algorithm, developed in previous
works. For instance, the optimal design of a multiproduct batch
chemical plant is not only to minimize the investment, but also to
minimize the operation cost, to minimize the total production time
and to maximize the revenue, simultaneously. The paper is organized
as follows. Section 2 presents a brief overview design approaches
under uncertainty. Section 3 is devoted to process description,
problem formulation and presents an overview of fuzzy set the- ory
involved in the fuzzy framework within a multiobjective genetic
algorithm. The presentation is then illustrated by some typical results
in Section 4. Finally, the con- clusions on this work are draw.

1.2 Design Approaches under Uncertainty

The literature offers three basic approaches to the problem of design

under uncer- tainty . They can be classified within the wait-and-see
approach, the probabilistic model and the two-stage formulation.
In the wait-and-see formulation, a separate optimal design is found
for each reali- zation of the set of uncertain parameters. The cost of
the plant is then calculated as the expected value of the separate
designs. The practical difficulties with this model are that it is generally
difficult to identify the design which yields a value of the plant cost
which is equal to the expected value, and even if such a design can be
identified there is usually no direct way establishing to what extent
this design will accommodate other values of the uncertain
parameters.
In the probabilistic model, also often called the chance constrained
model, a prob- ability of constraint satisfaction must be specified by
the designer. As show by Char- nes and Cooper , if normal distributions
are assumed, if the uncertain parameters linearly occur in the
constraints and if the constraints can be considered independ- ently,
 then the probabilistic constraints can be reduced to a deterministic
form and thus the model converts to an ordinary deterministic
optimization problem.
In the two-stage formulation, also called the “here-and-now” model,
the design variables are selected “here and now” so as to
accommodate any future uncertain parameter realizations or perhaps
those which fall within some specified confidence limits. The
equipment sizes are determined at the first step or design stage and
the effect of the uncertain parameters on systems performance is
established in the second or operating stage. The second stage is, of
course, the most important part of model since this is the stage at
which the flexibility of the design is checked, possibly by including
considerations of variations of the operating variables to
accommodate the
uncertain parameter realizations. Since this part of the model is also
the most compu- tationally demanding, researchers have sought to
reduce the computational burden by proposing various alternatives
fully solving the second-stage problem . Two- stage stochastic
programming approaches have also been applied for design under
uncertainty
1.3 A Fuzzy Decision-Making Approach for Multiproduct Batch
Plant Design

 Problem Statement
The common approach used by previous research is to describe an
optimal design problem of a multiproduct batch chemical plant as a
single-objective optimization problem. In the single-objective optimal
design problem, the number and size of par- allel equipment units in
each stage as well as the location and size of intermediate storage are
to be determined in order to minimize the investment. Such an
approach formulates the optimal design problem as a single-objective
mixed-integer nonlinear programming (MINLP) problem ([5], [9], [11],
[12], [13], [18], [19]). Indeed, the minimizing investment is not a
unique objective for the optimal design problem. In real application,
designers not only consider to minimize the investment, but also to
minimize the operation cost and total production time, and to
maximize the revenue, simultaneously.

 Assumptions
The model used in this study is derived and adapted from that
proposed by Modi and Karimi [11]. Actually, this model has its early
roots in a work of Grossmann and Sargent [5], who developed a simple
formulation for multiproduct batch plant design. This same
formulation was adopted and extended by several authors such as
Modi and Karimi, who, in addition to the typical batch items,
considered semi-continuous stages and storage tanks (with fixed
location in the process) to the initial formalism. In a majority of these
mono-objective formulations, the typical criterion consists of the
minimization of the investment cost corresponding to all items.
Furthermore, the solution is subject to a major constraint, forcing the
synthesis time of all products to respect a time horizon H. The decision
variables, which define the plant configuration, are the (continuous)
size and (discrete) number of the items for each processing stage.
The model used in this paper, although based in the above-
mentioned formulation, was modified according to the proposal of
Aguilar-Lasserre et al.
The model is based on the following assumptions:
 The devices used in a same production line cannot be used twice
by one same batch.
 The production is achieved through a series of single product
campaigns.
 The units of the same batch or semi-continuous stage have the
same type and size.
 All intermediate tank sizes are finite.
 If a storage tank exists between two stages, the operation mode is
“Finite In- termediate Storage”. If not, the “Zero-Wait” policy is
adopted.
 There is no limitation for utility.
 The cleaning time of the batch items is included into the
processing time.
 The item sizes are continuous bounded variables.

 Model Formulation
The model considers the synthesis of I products treated in J batch
stages and K semi- continuous stages. Each batch stage consists of mj
out-of-phase parallel items with same size Vj. Each semi-continuous
stage consists of nk out-of-phase parallel items with same processing
rate Rk (i.e. treatment capacity, measured in volume unit per time
unit). The item sizes (continuous variables) and equipment numbers
per stage (discrete variables) are bounded. The S-1 storage tanks, with
size Vs*, divide the whole process into S sub-processes.

 Objective Function
Multi-objective optimization provides a framework for understanding
the relation- ships between the various objective functions and allows
an engineer to make deci- sions on how to trade-off amongst the
objectives to achieve performance considered “the best”. If the
objective functions are incommensurate, or competing, then mini-
mization of one objective function requires a compromise in another
objective func- tion. Such an optimal design problem becomes a
N

MOOP as following (maximization of the revenue and minimization of

the investment cost, the operation cost and the total production
time):

Max
(V p )   C (1)
Pi Q i
i1
j  
Min (Cost )  J ( m a V ) K(nbR k) S(cV s) (2)
  k k  s
j jjj k k 1 s s 1

1 N M Q i
Min ( D P )    C E CO (3)
Q
Bi
i 1 j j i
1 is
I Qi L
Min ( H )   (4)
i1 B T
is
is

 Constraint Formulation
The problem statement involves four forms of different constraints as reported in literature [11]:
(i) Dimension constraints: every units has to restrict to its allowable range.
V min
 V j  V max  j  1 ,.., J 
(5)

n min  N j  n max  j  1 ,.., J  (6)

(ii) Time constraint: the summation of available production time for

L
all products is inferior to the total production time.
I
Q
 i
T is  H (7)
i1 B is
 (iii) Limiting cycle time for product i.
Tis L
 i 1 ,.., I 
  j  1 ,.., J 
(8)
n ij T
ij

(iv) Volume constraints: the volume Vj has to be able to process all the products i.
S ij B i  V j
 i 1 ,.., I ;  s  1 ,.., S  (9)

 Study Antecedent: Fuzzy Logics

The emergence of electronic commerce and business to business
applications has, in a recent period, considerably changed the
dynamics of the supplier-customer relationship. Indeed, customers
can change more rapidly their orders to the suppliers and many
enterprises have to organize their production even if the demand is
not completely known at short term. On the other hand, the
increasing need for integration and optimization in supply chains leads
to a greater sensitivity to perturbations resulting of this uncertainty.
These two elements clearly show the interest of taking into account as
soon as possible the uncertainty on the demand and to propagate it
along the production management mechanisms.
In the context of engineering design, an imprecise variable is a
variable that may potentially assume any value within a possible range
because the designer does not know a priori the final value that will
emerge from the design process. The fuzzy set theory was introduced
by Zadeh [21] to deal with problems in which a source of vagueness is
involved. It is well-recognized that fuzzy set theory offers a relevant
framework to model imprecision.
Huang and Wang [7] have introduced a fuzzy decision-making
approach to design a multi-objective optimal design problem of a
multiproduct batch chemical plant. A monotonic increasing or
decreasing membership function is used to define the degree of
satisfaction for each objective function.
The proposed approach involves arithmetic operations on fuzzy
numbers and quan- tifies the imprecision of the demand by means of
fuzzy sets (trapezoidal). In this case, the flat line over the interval
(q2,q3) represents the precise demands with an interval of
confidence at level =1, while the intervals (q1,q2) and (q3,q4)
represent the “more or ~
less possible values” of the demand. The result of the total production
time H i , the
revenue and the operation cost are treated and analyzed through fuzzy
numbers. The demand and the production time are fuzzy quantities as
shown in figure 1.

Q=(x) H =(x)
1

0
q1 q2 q3 q40 H1 H2 H3 H4

Fig. 1. Two trapezoidal fuzzy numbers, ~

= (q ,q ,q ,q ) and H=(H ,H ,H ,H )
Q 123 4 1234
 The criterion minimizes the delays and advances of the production
time necessary ~
for the synthesis of all the products. We must compare the time repre
horizon H ~ -
sented by a fuzzy expression (rectangle) and the Hi
production time (trapezoidal).
For the comparison of fuzzy numbers, Liou and Wang’s method [10]
was adopted.
The production time necessary to satisfy each product demand
must be less than a given time horizon, but due to the nature of the
fuzzy numbers, three different cases for determination of the
criterion may occur. The different cases are reported in figure 2.
The temporal criterion selected is called “common surface”,
representing the inter- section between the sum of the production
time (trapezoid) and the horizon of time to respect (rectangle). The
calculation of the criterion depends on each case: for exam- ple, case1
illustrate the solutions which arrive just in time.

 Overview of Multiobjetive Genetic Algorithm Approach

General principles on evolutionary computation will not be recalled
here, since the aim is only to insist on the particular details of the used
Genetic Algorithm (GA). Actually, this one is a standard multiobjective
algorithm, similar to the specific one developed in [1]. The proposed
GA procedure implies the following steps:

The encoding of the solutions was carried out

1) Encoding of solution.
dividing the chromosome, i.e. a complete set of coded variables, into
two parts. The first one deals with the items volumes, which are
continuous in the initial formulation. Nevertheless, they were
discretized here with a 50 unit range, while their upper and lower
bounds

H Hs H

Hs

Hs1 H1 Hs4 H4
 Hs1<H1
Hs4 H4

Figure 2a) Case 1 Figure 2b) Case 2: The advances

Hs
H

Hs1 H1 Hs4>H4

Figure 2c) Case 3: The delays

Fig. 2. Three
cases for the minimization of a criterion that penalizes the
delays and advances of the time horizon necessary for the synthesis of
all the products
were preserved. The second part of the chromosome handles the
number of equipment items per stage: the value of these discrete
variables is coded directly in the chromo- some.
2) Initial population creation. The procedure of creating the initial
population corre- sponds to random sampling of each decision variable
within its specific range of variation. This strategy guarantees a
population various enough to explore large zones of the search space.
3) Fitness Evaluation. The optimization criterion considered for fitness
evaluation involves the revenue and three other performance criteria,
i.e. the production de- lay/advance, the investment and the operation
cost.
4) Selection Procedure. The multi-objective aspects are taken into
account in the selection procedure. The better individuals, that are the
surviving individuals, are chosen with the Goldberg’s roulette. The
same number of surviving individuals is chosen for each criterion (four
roulettes). Each individual is represented by a slice of the roulette
wheel, proportional to its fitness value. Since the criteria are
represented by fuzzy numbers, they were defuzzified (the defuzzified
value was calculated as the centre of gravity) in the roulette wheel.
5) Crossover and mutation. Concerning genetic operators, the crossover
obeys to a classical one cut-point method, while the mutation
technique is in agreement with the encoding procedure: inversion of a
bit value on the continuous zone (0 becomes 1 and conversely), and
decrease of one unit for the discrete loci (if possible, of course).
6) Registration of all non-dominated individuals in Pareto set . Pareto’s sort
proce- dure is carried out at the end of the algorithm over all the
evaluated solutions; at the end of the procedure, the whole set of the
non dominated Pareto’s optimal solutions, are obtained.

 Treatment of an Illustrative Example

We consider an example to illustrate the approach fuzzy-AG based on
arithmetic operations on fuzzy numbers and quantifying the
imprecision of the demand. The example was initially presented by
Ponsich and al. [15]: the plant, divided into two sub-processes,
consists of six batch stages to manufacture three products.
The GA parameters are the following ones: Population size 400
individuals, num- ber of generations 500 iterations, crossover
probability 40%, mutation probability 30% and the stop criterion
considered in this study concerns a maximum number of generations
to reach.
For the considered example, table 1 shows the values for processing
times, size factor for the units, cost data, and the production
requirement for each product quanti- fying the imprecision of the
demand by means of fuzzy numbers representing the “more or less
possible values”.
For the construction of the trapezoid which represents the request
for each product, the original values of the demand were used as a
reference. To determine the support and the core, one calculated a
percentage of opening taking as reference the demand of the original
data is computed.
 Table 1. Data used in example

Processing time i,j (h) Size factors (1/kg)

B1 B2 B3 B4 B5 B6 B1 B2 B3 B4 B5 B6

Minimum size =250 l A 1.15 3.9 9.8 5.28 1.2 3.5 8.2 6.9 9.7 2.9 6.5 10.6
8 6 7 8 2 5 7
Maximum size = 10 000 l B 5.95 7 1.0 8.0 6.57
7.5 7.0 8 5.7 5.5 8.0 9 3.2 6.1
C 3.96 2 1 5.13 8 8 3 7 7 3.14
0.6 10.
j 0.4 5.0 6.0 0.3 6 4.3 2.3 9.1 3 5.7 5.9
7 1 7 4 9 8
0.2
0.2 0.3 0.3
9 3 5

Unit price for product i Coefficients ci,j Q1=(428260, 432630, 441370, 445740)
($/Kg)
Q2=(317520, 320760, 327240, 330480)

Q3=(252840, 255420, 260580, 263160)

H = (5800, 5800, 6200, 6200) The

horizon H (rectangle) (h)

CP CO B1 B2 B3 B4 B5 B6 Cost of mixer=$250V0.6
A 0.70 0.08 0.2 0.36 0.24 0.4 0.5 0.4 Cost of reactor=$250V0.6
B 0.74 0.1 0.15 0.5 0.35 0.7 0.42 0.38 Cost of extractor=$250V0.6
C 0.80 0.07 0.34 0.64 0.5 0.85 0.3 0.22 Cost of centrifuge=$250V0.6
(Volume V in liter)
Operating cost factors
B1 B2 B3 B4 B5 B6
CE 20 30 15 35 37 18

1.3 Typical Results

The result obtained of the multi-objective optimization problem
involves 44 non- dominated solutions: among them, 42 represent the
first case, where the production time Hi (trapezoidal) is within of time
horizon H (rectangle) The solutions 43 and 44 represent respectively
cases 2 and 3 (the criterion that penalizes the delays and ad- vances of
the time horizon necessary for the synthesis of all the products). This
study leads to three different scenarios as a fuzzy decision-making
approach to preliminary design.
Table 2 shows the solution that minimizes the investment cost for
case 1. Tables 3 and 4 present the results that minimize the
penalization of delays and advances (cases 2 and 3), respectively.
These tables display typical results obtained after ten runs to guaranty
the stochastic nature of the algorithm. Since a multi-objective analysis
is considered, a statistical analysis of the results is difficult to carry out
and was not performed as it was the case in our previous studies in
mono-objective optimization.
From these results, it can be concluded that:
(i) The fuzzy optimal design of batch plant is the solution of case 2
(the advance) because its operation cost is lower than the solution
of case 1. The production
~
time of the case of advance H i = 5788.7, 5847.8, 5965.9, 6025 is
lower than
~
case 1 H i = 5800.1, 5859.3, 5977.7, 6036.9.
(ii)The fuzzy design of batch plant for case 1 is a sure solution
because the pro-
~
duction time H i = 5800.1, 5859.3, 5977.7, 6036.9 is within the
imposed~ time
horizon H = 5800, 5800, 6200, 6200.
 (iii) The fuzzy design of batch plant for the case 3 is a more
hazardous solution, ~
because the H i = 5958.5, 6019.3, 6140.9, 6201.7 is
production time ~
greater
than the time horizon H = 5800, 5800, 6200, 6200].

Table 2. Fuzzy optimal design of batch plant for case 1

Product Bis kg TLi h Optimal objective function value Storage Tanks

A 589.6 3.9 I = 721078.1 [$] Vs = 2000.9 [l]

~
B 951.2 5.8 V = [737018.8 744539.4 759580.6 767101.2] [$]
p
C 873.7 5.0
~
D = [270064.4 272820.2 278331.7 281087.5] [$]
p

∑ ~H= [5800.1 5859.3 5977.7 6036.9] [h]

i

Table 3. Fuzzy optimal design of batch plant for the criterion that penalizes the advances

Product Bis kg TLi h Optimal objective function value Storage Tanks

A 832.3 4.4 I = 693084 [$] Vs = 2121.9 [l]

~
B 1008.7 6.6 V = [737018.8 744539.4 759580.6 767101.2] [$]
p
C 771.9 5.8
~
D = [251027.7 253589.2 258712.9 261273.0] [$]
p

∑ ~H= [5788.7 5847.82 5965.9 6025] [h]

i

Table 4. Fuzzy optimal design of batch plant for the criterion that penalizes the delays

Prod ct Bis TLi h Optimal objective function value Storage Tanks

u kg
A 943.5 6.2 I = 698906.2 [$] Vs = 2150.6 [l]
C B 964.9 5.7 ~ = [737018.8 744539.4 759580.6 767101.2] [$]
1226.6 6.3 ~
= [232624.1 234997.8 239745.27 242118.9] [$]
Dp

∑ ~H= [5958.5 6019.3 6140.9 6201.7] [h]

i
1.4 Conclusions
For the most common form of design of a multiproduct batch chemical plant, the
designers specify the production requirement of each product and the total
production time. However, no precise product demand predictions are generally
available. For this reason, an alternative treatment of the imprecision by using fuzzy
concepts is introduced in this paper.
In real application, designers not only require to minimize investment cost, but also
to minimize the operation cost and the delays and advances of the time horizon
necessary for the synthesis of all the products, and to maximize the revenue,
simultaneously.
In this study, we have introduced a fuzzy-AGs approach to solve the problem of
multi-objective optimal design of a multiproduct batch chemical plant. The results
obtained on the treated example have shown that three different scenarios were
obtained as a fuzzy decision-making approach. The analysis tended to be helpful for
decision making.

Nomenclature
CEj= Operation cost in stage j($)
COi= Operation cost of product i to be produced ($/kg)
CPi= Price of product i ($/kg)
Sij= Size factor of product i in stage j (1/kg)
Bis= Size of the batch of product i
Qi= Production requirement of product i (kg)
TLi= Cycle time for product i (h)
j= Cost coefficient for unit j
j= Cost exponent for unit j
ij= Processing time of product i in stage j (h)
 2. Optimal design of a multi-product reactive distillation
system for silences production .

2.1 Introduction

The quest for new sustainable fuels has become an important effort because of
fossil fuels depletion and its negative impact on the environment. As part of that
search, solar energy has received sig- nificant attention recently. For instance, solar
cells or photovoltaic cells have been manufactured in the last decades at increasing
rates. Single crystal silicon is the most important material for modern solar and
semiconductor industries. Ultra-high purity of polysili- con is achieved first by the
preparation of a volatile silicon hydride,
i.e. silane, and its purification, generally using fractional distil- lation. This is
followed by the decomposition of this hydride to hyper-pure elemental silicon by
reductive pyrolysis or chemical vapor deposition. The preparation of the volatile
silicon (Si) com- pound involves external reactants and its decomposition generates
by-products, mainly silicon tetrachloride and other chlorosilanes; such by-products
are typically recycled to improve the process effi- ciency, although market conditions
could promote their sales for applications in the manufacture of optical fibers.
Three processes can be mentioned as the commonly used routes to produce
polysilicon. The most popular process, developed in the late 1950s, is based on the
thermal decomposition of trichlorosilane (TCS) on a heated silicon rod or filament
placed inside a deposi- tion chamber. This process is commonly referred to as the
Siemens process, is energy intensive, and a major part of the energy is dis- seminated
and lost. In 2001 processes based on this technology still accounted for at least 60% of
the worldwide production of polysil- icon (Ceccaroli and Lohne, 2010). The by-product
SiCl4 (STC) is a low-value waste product that cannot be used efficiently for polysil-
icon growth. However, STC can be used for its reconversion to the valuable starting
material SiHCl3 to form a closed-loop production process (Sarma and Rice, 1982).
The second process came from an improvement of the Siemens process developed
by Union Carbide corporation (Luque and Hegedus, 2011). Fig. 1 presents the
flowsheet for such a process for the production of polysilicon from mineral grade
silicon. The trichlorosilane was replaced by monosilane SiH4, with the same principle
of decomposition on a heated silicon rod inside a closed deposition chamber. In this
process silicon STC is converted to TCS via a hydrogenation reaction, after which TCS is
converted to dichlorosilane (DCS) and then to monosilane through redistribu- tion
reactions. This process has found significant market acceptance over the last 30 years.
The other commonly used process, developed in the 1980s–1990s, is the Ethyl
Corporation process (Luque and Hegedus, 2011). It also makes use of monosilane, for
which the heated silicon rod in the closed reaction chamber is replaced by a fluidized
bed of heated silicon particles, which act as seeds for the continuous decomposition
of SiH4 into larger granules of hyperpure silicon. As opposed to the previous two
technologies, this process operates on a continuous basis.
Dichlorosilane and monochlorosilane are intermediate com- pounds in the
production of silane that are typically recycled, but they also have other important
applications. Some of them are described below.

Robinson and Goldsmith (1975) studied silicon epitaxial growth using

dichlorosilane. They found that dichlorosilane is a suitable material for the deposition
of high quality epitaxial layers of sili- con. A process developed by Hemlock
Semiconductor Corporation showed the feasibility of using dichlorosilane instead of
trichlorosi- lane for the production of polycrystalline silicon (McCormick et al., 1989).
They reported a faster deposition rate and low power con- sumption when
dichlorosilane was used instead of trichlorosilane. Dichlorosilane (DCS) is mainly used
for the preparation of semi- conducting epitaxial silicon layers of very high purity, as
required for their application in the microelectronics industry (Marin and Lefort,
1977). Among the chlorosilane products, dichlorosilane has been found to be the best
material for epitaxial growth of inte- grated circuits (Chowdhury et al., 2011). DCS has
also been used for the production of silicon nitride, a material in microelectronic and
optoelectronic industries; the conversion is based on chemical vapor deposition or
plasma-enhanced chemical vapor deposition (Santana et al., 2005). Among the routes
to obtain dichlorosilane, the one based on thrichlorosilane disproportionation is the
most widely used. Li and Huang (1988) investigated the redistribution reaction of
trichlorosilane to dichlorosilane in a fixed bed reactor using different types of ionic-
exchange resin catalysts, and pro- posed a reaction mechanism to explain the
redistribution reaction. The other silane product of interest in this work, monoclorosi-
lane (MCS), was investigated as a precursor for the vapor phase
epitaxy of silicon. Tomasini and Weeks (2011) found that MCS is an equally
effective Si precursor in comparison to SiH4 and DCS. Monochlorosilane can also be
used as reagent to produce chem icals such as trisilylamine (Hoppe et al., 2014),
which is used as precursor for the deposition of high purity silicon oxide films for gap
fill applications in the semiconductor industry. Additional applica- tions of
chlorosilanes have been developed (Bailey and York, 1959; Seyferth et al., 1984).
Given the importance of the three types of silanes mentioned above, this work
focuses on the analysis, design and optimiza- tion of a reactive distillation system that
has the flexibility to shift the production of silane, monochlorosilane and
dichlorosilane as a function of the market demand for each product. The system is to
be operated to produce one product at a time. For the sake of this paper, the problem
is simply defined as the design of a multi- product reactive distillation column.
Structurally, the problem can be viewed as the substitution of the highlighted section
of Fig. 1, which includes two reactors and four distillation columns, into a single piece
of equipment, i.e. a reactive distillation column.
This work is organized in the following way. The design strategy and the equations
for the chemical kinetic reactions for silanes pro- duction are presented. The design of
reactive distillation columns for each silane product is then carried out. From these
results, a base design is developed for the production of any of the three products
with a single unit, which is then subject to an optimization task to find the best
structure and operating conditions for each product. Given the importance of the
temperature of the reactive zone due to catalyst restrictions, an initial dynamic
analysis is then reported on the changes of the hottest stage against feed
disturbances. The last section gives the relevant conclusions from this work.

2.2 Design of the silanes process

A single reactive distillation unit is to be designed for the pro- duction of silane,
monochlorosilane and dichlorosilane with high purities. The reaction step starts
with the disproportionation of trichlorosilane, which produces the three silane
products as part of the following reaction mechanism involving five chemical
species and three reversible reactions,
catSiCl + SiH Cl
2SiHCl3↔ 4 2 2

2SiH2Cl2 cat
↔ SiHCl3 + SiH3Cl
2SiH3Cl cat
↔ SiH2Cl2 + SiH4

The first two reactions are slightly endothermic, while the last one is exothermic.
According to the reaction mechanism, dichlorosilane can be produced if only the
first reaction is car- ried out. Similarly, monochlorosilane is produced if only the first
two reactions take place, while the production of silane requires all three reactions
to occur. Therefore, for the production of the silane product of interest, the
operating conditions must be set accordingly so as to favor the reaction(s) of
 interest
For this mechanism, the following reaction rate equations have been reported by
Huang et al. (2013)

where x0, x1, x2,x3, and x4 correspond to the liquid mole fraction of silicon
tetrachlorine, trichlorosilane, dichlorosilane, monochlorosi- lane and silane
respectively; r1, r2, and r3 are the reaction rates of trichlorosilane, dichlorosilane,
and monochlorosilane dispro- portionation; k is the forward rate constant and K the
chemical equilibrium constant, for which the following relationships apply:

1. Initial design and search variables for each product

The main goal of this work is the design of a multi-product reactive distillation (RD)
system. The first step we follow is the development of individually intensified
processes via reactive dis- tillation systems for the production of each silane product.
The design of reactive distillation systems involves the detection of rec- tifying,
reactive and stripping sections; a conceptual structure of this system is shown in Fig. 2.
Subawalla and Fair (1999) report some guidelines for several design variables such as
the location of the reactive zone, feed location, number of stages, pressure, and reflux
ratio. Initial interest in the reactive distillation process for silane production is reflected
in the patent by Bakay (1976) and the work by Müller et al. (2002). A relevant aspect
for silane produc- tion is the need for refrigeration utilities at the top of the column.
Also, it is important to avoid thermal decomposition of the cata- lyst in the reactive
trays, which restricts the use of high pressure operation. In this work, the design
reported by Huang et al. (2013) for the production of silane with a purity of 99.3% was
taken as a basis. The analysis will serve to establish conditions that promote the
production of each of the three silane products, which will then be subjected an the
optimization procedure. The RD column con- sists of 60 stages, out of which 29 are
reactive; the reactive zone starts at stage 16, with a residence time per stage of 2.5 s.
The feed stage is located at stage 46, just below reactive zone, and the feed
flowrate is 10 kmol/h of TCS. The operating specifications include a reflux ratio of
63 and a feed to distillate ratio of 0.25. The pressure
on the top of the column is 5 atm, with a pressure drop of 0.5 kPa
per stage. The design was validated in ASPEN Plus using a RADFRAC
block.
In order to identify the conditions under which the production of each silane
product would be favored, a sensitivity analysis was conducted through variations on
the design pressure and the dis- tillate to feed ratio (D/F). The results are shown in Fig.
3. From

this analysis, one can identify that with distillate to feed ratios

of 0.25, 0.35, and 0.5 the corresponding distillate compositions of silane,

monochlorosilane and dichlorosilane are favored. When the pressure was varied,
values of 5, 1.5 and 1 atm increased the dis- tillate composition of silane,
monochlorosilane and dichlorosilane, respectively. One can also observe in Fig. 3
how those D/F ratios and pressure values generate the highest amount of the silane
com- pound of interest. Such values were then used as starting points for the
optimization step for the production of the individual products. It should be noted
that the reflux ratio was not analyzed since one degree of freedom is lost due to the
purity specification.
1.1. Individual designs for the production of each silane product

The three individual designs for reactive distillation columns, one for the
production of each silane product, were then carried out. The optimization
procedure was based on the minimization of the total annual cost (TAC), which
includes a total investment cost (TIC), the total energy cost (TEC), and the total
catalyst cost (TCC). The optimization variables are the number of stages, feed stage,
number of stages for the reactive zone and their location, total liquid holdup of
reactive zone, and column pressure. The number of stages in the reactive zone and
liquid holdup are associated to the amount of catalyst needed annually (mcatalyst ).
The optimization model is written as,

MinTAC = TIC + TEC + TCC

s.t. TIC Costshell + Costcondenser + Costreboiler

paybackperiod =

TEC = Putility ∗ Qcondenser + Pvapor ∗ Qreboiler

TCC = Pcatalyst ∗ mcatalyst

The costing equations for the distillation system were taken from Kiss (2013) and
are reported in Table 1. The price of Amberlyst catalyst (Pcatalyst ) was fixed at
$330/kg. Three months of catalyst life and a volume of 50% of the liquid holdup of
reactive zone were
assumed. A product purity specification of 99.5% was used for each silane product.
The top temperature in the column affects the type of refrigeration utility needed for
the condenser. The utilities con- sidered in this work, along with their prices, are
given in Table 2.
The optimization of reactive distillation processes presents major challenges,
particularly when use of gradient methods is con- sidered. The problem involves
simultaneous reaction-separation operations, which drives a highly nonlinear model
with heavy interactions among design variables. To tackle this problem, derivative-
free optimization methods, such as the use of genetic algorithms (GAs), provide a
suitable approach. This technique has shown its effectiveness in several design
applications, for instance in the optimization of heat transfer processes (Ponce-
Ortega et al., 2009) and the design of complex separation systems (Gutiérrez-
Antonio and Briones-Ramírez, 2009; Vázquez-Ojeda et al., 2013).
The optimization approach was implemented in a similar way as in Medina-
Herrera et al. (2014) for the optimal design of extractive distillation systems. The GA
method available in the MATLAB envi- ronment was used, and a link between
MATLAB and ASPEN Plus was created so that each search point was provided from
rigorous sim- ulations as part of the automated procedure. The general approach,
 shown schematically in Fig. 4, was as follows. An initial point for the optimization
variables was provided, along with bounds on the search variables and the GA
parameters. The GA sent the values of the optimization variables to ASPEN Plus, in
which the rigorous simulations were performed using RADFRAC blocks. The simula-
tions proved to be straightforward, with no convergence issues for any of the runs
that were required. The results from the simulation were then sent back to MATLAB,
where the genetic algorithm calculated the objective function (total annual cost).
The steps of the genetic algorithm are described briefly. First, an initial
population (i.e. a set of RD designs) is created randomly. Then a sequence of new
populations is created, with each popula- tion based on the best individuals of the
current generation. There are six steps to create the next population. (1) Each
individual of the current population is evaluated and its corresponding objective
function is computed. (2) An escalation based on objective function values is carried
out so that more suitable or promising ranges of the search variable are obtained.
(3) The best individuals (i.e. the best designs) are selected as parents for the next
generation. (4) Some of the individuals of the current population with the best
objec- tive function (lower total annual costs) are chosen as elite and pass directly
to the next generation (elite count). (5) Children are cre- ated based on random
changes to a single parent (mutation) or a combination of two parents (crossover).
(6) The next generation is formed with the children of the current population. The
algorithm finishes when a given stopping criterion such as a maximum num- ber of
iterations (generations) is reached. The GA parameters used in this work were 50
generations, a population size of 250 indi- viduals, and a parent selection option of
“stochastic uniform”. For children creation, a cross over fraction of 0.8, a mutation
fraction of 0.2 and elite count of 2 were used.
The base case and the optimal designs obtained for silane, monochlorosilane and
dichlorosilane individual products are given in Table 3, where a comparison to the
design reported by Huang et al. (2013) for silane production is included. The
optimization task
Table 1

Equipment sizing and economics equations.

took 8, 5 and 5 h for silane, monochlorosilane and dichlorosilane designs, respectively. One can observe
that the design obtained in this work for silane requires a higher number of stages than the one by Huang
et al. (2013), mostly due to the number of stages for the reactive zone (67 vs. 30). However, the design
obtained here pro- vides savings in total annual cost of 45%, driven by energy savings of 56% and investment
savings of 24%. The investment savings were obtained because of a smaller column diameter and smaller
areas for the condenser and the reboiler. The feed is placed in stage 17, within the reactive zone, which differs
from the one reported by Huang et al. (2013) in which the feed stream is located in stage 46, just below the
reactive zone. It should be noticed that the silane optimized design requires 1.7 times more catalyst than the
design of Huang et al. (2013) Nonetheless, the energy and investment savings observed for the design
obtained here provide a more economical design.
When comparing the designs for the three products, one can see that the silane reactive distillation
column requires the highest
Table 3
Reactive distillation design for each silane product.

Design of Huang et al. (2013) Optimized Silane Optimized Monochlorosilane Optimized Dichlorosilane

Number of stages 60 91 72 71
Feed stage 46 17 32 40
Feed flowrate of SiHCl3 (kmol/h) 10 10 10 10
Reactive Zone (Stage to Stage) 16–45 6–73 14–70 30–65
Residence time (s) 2.5 – – –
3
Liquid Holdup (m ) ≈0.06 0.16 0.045 0.107
Purity of product in D1 (mol%) 99.3 (silane) 99.5 99.5 99.5
Purity of SiCl4 in B1 (mol%) 100 99.83 100 100
Distillate to feed ratio 0.25 0.25 0.3335 0.5
Reflux ratio 63 27 44 11.3
Operating pressure (atm) 5 4.84 2.72 1.00
Pressure drop per stage (kPa) 0.5 0.5 0.5 0.5
Column diameter (m) 0.58 0.4 0.75 0.59
◦
Maximum temperature in the reactive zone ( C ) 95.8 97 97 65.1
◦
Top Temperature ( C ) −78 −78.9 5.1 5.6
Type of cooling utility R1 R1 R3 R3
Qc (kW) 535 228 786 377
QR (kW) 568 269 814 388

Total Investment Cost (TIC) (105$/year) 3.6183 2.7385 3.4698 2.3332

Total Catalyst Cost (TCC) (104$/year) 1.3068 3.4848 0.9772 2.3378
Total Energy Cost (TEC) (105$/year) 9.6375 4.2289 5.5612 2.6576
Total Annual Cost (TAC) (105$/year) 13.386 7.3159 9.286 5.2245
TAC savings (%) – 45.35 – –
Energy savings (%) – 56.12 – –
Investment savings (%) – 24.32 – –
Catalyst savings (%) – −166.67 – –

number of stages (91) with respect to those needed for monochloro and
dichlorosilane of 72 and 71 trays. Dichlorosilane requires the lowest number of
stages for the reactive zone (35) but the highest for the separation (36). The feed
stage is located inside the reac- tive zone for all designs, a different feature with
respect to the initial design taken from Huang et al. (2013) In agreement with the
constraint imposed for the reactive zone, the maximum tempera- ture is below the
thermal decomposition of the Amberlyst catalyst of 100 ◦C; the highest temperature
in the reactive zone for silane and monochlorosilane column was the same, 97 ◦C,
while that for dichlorosilane was quite lower, 65.1 ◦C. Although the energy con-
sumption required by the silane design is the lowest, the energy cost is the highest
because of the low temperature of the distil-
late product (-78.9 ◦C), compared to temperatures of −5.1 ◦C and
5.7 ◦C required at the top of the column for monochlorosilane and
dichlorosilane. As a result, the production of silane requires the most expensive type
of refrigerant. In terms of the total annual cost, monochlorosilane would require the
most expensive process, while the one for dichlorosilane would be the most
economical.
The composition and temperature profiles for each design are presented in Fig. 5
(the reactive zones are highlighted). For silane production the stages of the reactive
zone account for 74% of the total stages, while the rectifying section has only 6
stages. On the other hand, for dichlorosilane column, the reactive zone
comprises49% of its design, while the rectifying zone accounts for 42%. These
observations serve as a basis to develop an initial design for the
optimization of a multi-product reactive distillation system.