Accepted Manuscript

Title: Logic hybrid simulation-optimization algorithm for

distillation design.

Author: J.A. Caballero

PII: S0098-1354(14)00097-0
DOI: http://dx.doi.org/doi:10.1016/j.compchemeng.2014.03.016
Reference: CACE 4929

To appear in: Computers and Chemical Engineering

Received date: 15-1-2014

Revised date: 18-3-2014
Accepted date: 20-3-2014

Please cite this article as: Caballero, J. A.,Logic hybrid simulation-optimization

algorithm for distillation design., Computers and Chemical Engineering (2014),
http://dx.doi.org/10.1016/j.compchemeng.2014.03.016

This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
 Logic hybrid simulation-optimization algorithm for
distillation design.
J. A. Caballero

t
Department of Chemical Engineering, University of Alicante. Apartado de correos 99, 03080,

ip
Alicante Spain. email: caballer@ua.es.

cr
us
Highlights

!! A novel Generalized Disjunctive Model is presented for rigorous distillation design.

an
!! A logic based algorithm without MINLP reformulation is used for solving the problem

!! The model is solved using a hybrid simulation-optimization approach.

M
!! We take advantage of the best of algebraic models and process simulators.
d
p te

Abstract
ce

In this paper, we propose a novel algorithm for the rigorous design of distillation columns that
integrates a process simulator in a generalized disjunctive programming formulation. The
optimal distillation column, or column sequence, is obtained by selecting, for each column
Ac

section, among a set of column sections with different number of theoretical trays. The selection
of thermodynamic models, properties estimation etc., are all in the simulation environment. All
the numerical issues related to the convergence of distillation columns (or column sections) are
also maintained in the simulation environment. The model is formulated as a Generalized
Disjunctive Programming (GDP) problem and solved using the logic based outer approximation
algorithm without MINLP reformulation. Some examples involving from a single column to
thermally coupled sequence or extractive distillation shows the performance of the new
algorithm.

Keywords: Distillation; Generalized Disjunctive Programming; Simulation; Optimization.

1. Introduction

Page 1 of 40
The general separation problem was defined more than 40 years ago by Rudd & Watson (1968)
as the transformation of several source mixtures into several product mixtures. More than forty
years later we can say that this general problem has not been completely solved. We will focus,
in this work, in the more restricted problem of separating a single source mixture into several
products using only distillation columns.

Distillation is likely the most important separation and purification operation in chemical process

t
industries. Typically more than half of the process heat distributed to a plant is dedicated to

ip
supply heat in the reboilers of distillation columns (Kunesh et al., 1995). However, the energy is
provided to the bottom of the column and approximately the same amount of energy removed in

cr
the top, although at lower temperature, which yields an inefficient process, but still one of the
most effective for homogeneous mixtures separations. To get an idea of the importance of

us
distillation, Humphrey (1995) estimated that distillation handles more than 90% of all the
separations and purifications. Soave & Feliu (2002), using data by Mix et al (1978) estimated
that distillation accounts about 3% of the total United States energy consumption. This is

an
18
equivalent to 2.87·10 J per year (91 GW or 54 million tons of crude oil). The capital investment
for these distillation systems was estimated to be around 8 billion US$.

The optimization of distillation columns involves the selection of the number of trays, the feed
M
location and the operating conditions to minimize a performance function, usually the total
annualized cost that involves investment and operating costs. Discrete decisions are related to
the calculation of the number of trays and feed and products location, and continuous decisions
d

are related to the operation conditions. Due to the discrete-continuous nature of the problem
and to the complex equations involved, it is common use shortcut or aggregated models
te

together with some rules of thumb that under some assumptions have proved to produce good
results, at least in the first stages of design where a rigorous design is neither necessary nor
p

convenient due to the large computational effort needed. Some of the most successful shortcut
ce

methods are:

Fenske – Underwood- Gilliland (FUG). (Fenske, 1932; Gilliland, 1940; Underwood, 1948). The
FUG method assumes a constant molar overflow and constant relative volatilities in all the trays
Ac

of the distillation column. This method considers two extreme ideal situations. a) The distillation
column operates at total reflux (no feed is entering or exiting from the column), which allows
calculating the minimum number of trays for a given separation of two key components, and b)
when the column operates at pinch conditions, (infinite number of trays), which allows
calculating the minimum reflux. The optimal situation is in some point in between these two
extreme cases. Group methods (GM) (Edmister, 1943; Kamath et al., 2010; Kremser, 1930).
GMs use approximate calculations to relate the outlet stream properties to the inlet stream
specifications and number of equilibrium trays. They provide only an overall treatment of the
stages in the cascade without considering detailed changes in the temperature and composition
of individual stages. However, they are much easier to solve because they involve fewer
variables and constraints. Aggregated models (AG) (Bagajewicz & Manousiouthakis, 1992; J. A.

Page 2 of 40
Caballero & Grossmann, 1999). AG models are similar to group methods, they are based on
mass balances and equilibrium feasibility, expressed in terms of flows, inlet concentrations, and
recoveries. The Boundary Value Method (BVM) (Barbosa & Doherty, 1988; Fidkowski et al.,
1993; Fidkowski et al., 1991; Julka & Doherty, 1990; Levy & Doherty, 1986; Levy et al., 1985).
BVM can be used to determine the minimum reflux ratio and feasible design parameters for a
column separating a ternary mixture. It allows to obtain the number of trays, composition

t
profiles etc. The Rectification Body Method (RBM) (Bausa et al., 1998; von Watzdorf et al.,

ip
1999). RBM is used for the determination of minimum energy requirements for a specified
separation. The method approximates rectification bodies by straight lines. The intersection of

cr
the rectification bodies of two sections indicates its feasibility. Driving Force Method (DFM)
(Gani & Bek-Pedersen, 2000). The DFM is a graphical method. Its authors proved that the
minimum energy requirements corresponds to a maximum in the driving force. The Shortest

us
Stripping Line (SSL) (Angelo Lucia et al., 2008; A. Lucia & McCallum, 2010; Angelo Lucia &
Taylor, 2006) Authors showed that the longest residue curve is related with the highest energy
consumption for a given separation. Then the shortest curve should produce the minimum

an
energy required for the same separation.

Some of the previous methods have been automated, although not all of them can be directly
M
used with an optimization algorithm. In any case, they are valuable tools for obtaining precise
initial values or reliable bounds for the rigorous optimization of distillation columns.

2. Overview of rigorous tray by tray optimization models.

d

As commented in the introduction section, the economic optimization of a distillation column

te

involves continuous decisions, related to the operational conditions and energy involved in the
separation, and discrete decisions related to the total number of trays, and the tray positions of
p

each feed and product streams. A major challenge is to perform the optimization using tray by
tray models that assume phase equilibrium.
ce

The first approach to solve the above commented problem was due to Sargent &
Gaminibandara (1976). In this case, the authors assumed a fixed number of trays, and the goal
Ac

was to select the optimal feed location. To that end, the feed is split into as many streams as
trays has the column (condenser and reboiler are excluded). Figure 1 shows the superstructure.
The model can be written as a Mixed Integer Nonlinear Programming (MINLP) problem by
considering the MESH equations (Mass balances, Equilibrium equations, molar fraction
Summation equals one in all phases, and Enthalpy balances). However, computational
experience shows that this problem is usually solved as a relaxed NLP.

The first model that considers both, the feed tray position optimization and the total number of
trays was due to Viswanathan & Grossmann (1993). The authors used a superstructure that
involves a variable reflux location as shown in Figure 2. The superstructure considers a fixed
feed tray and a column formed by a large enough number of trays above and below the feed.
The reflux (reboil) is returned to all the trays above (below) the feed. The model takes the form

Page 3 of 40
of a MINLP and relies also on MESH equations. A major difficulty with this model is related to
the non-existing trays. In these trays, there is a zero liquid flow (rectifying section) or a zero
vapor flow (stripping section), which can produce numerical problems due to the convergence of
equilibrium equations with a zero value in the flow of one of the phases..

To avoid the numerical problems in MINLP models Yeomans and Grossmann (Yeomans &
Grossmann, 2000a, 2000b) proposed a Generalized Disjunctive Programming model by

t
allowing the bypass of those trays that are not selected. Figure 3 shows the column

ip
representation for this approach. For each existing tray the mass transfer task is accounted for
and modeled with the MESH equations. For a non-existing or inactive tray the task considered

cr
is simply an input-output operation with no mass transfer. Because the MESH equations include
the solution for trivial mass and energy balances, the only difference between existing and non-

us
existing trays is the application of the equilibrium equations. As for the permanent trays, all the
equations for an existing tray apply. The advantage of the disjunctive modeling approach is that
the MESH equations of the non-existing trays do not have to be converged, and no flows in the

an
column are required to take values of zero, making the convergence of the optimization
procedure more reliable. Also, by using Generalized Disjunctive Programming (GDP) as the
modeling tool, the computational expense of solving the problem can be reduced. Barttfeld et
M
al., (2003) considered different representations for the GDP model. Numerical results studies for
separation of ternary mixtures in a single column suggests that the GDP formulation requires
less solution time but is more sensitive toward local optima than MINLP formulations. Even
though, GDP seems to be more reliable than MINLP models both require good initial values and
d

bounds to converge. Barttfeld & Aguirre (2002, 2003), propose to use a reversible distillation
te

model that involves the minimum reflux conditions as well as minimum entropy production to
provide a feasible initial design, and good initial values for the rigorous optimization. Their
p

method is mainly limited by the drawbacks of this so-called “preferred separation”, because, for
azeotropic mixtures, usually nonsharp splits are generated. The extension to the most common
ce

sharp split is not trivial. Kossack et al (2006) proposed to use the Rectification Body Method
(RBM) that can be used in all the cases. However, the initialization procedure is rather complex.
Ac

Another option was proposed by Harwardt & Marquardt (2012): They start calculating the
minimum energy demand and the concentration profile estimated based on pinch points. Based
on these results a simplified model that comprises only component mole balances and
equilibrium relations, but no energy balances, is solved. In subsequent solution steps the
energy balance was included again and the model resolved. To solve the problem, they used
the so called successive relaxed MINLP (SR-MINLP) proposed by Kraemer et al (2007), which
reformulate the MINLP or GDP problems as pure continuous problems with tailored big-M
constraints, where all discrete decisions are represented by continuous variables. The discrete
decisions are enforced by non-convex constraints that force the continuous variables to take
discrete values

Page 4 of 40
Even with all these difficulties, complex problems have been successfully solved, including
reactive distillation (Ciric & Gu, 1994; Jackson & Grossmann, 2001); azeotropic sequences
(Mariana Barttfeld et al., 2004; Bauer & Stichlmair, 1996, 1998), Thermally coupled distillation
sequences (Dunnebier & Pantelides, 1999) or hybrid membrane/distillation systems (Kookos,
2003) among others.

On other side, process simulators are commonly used tools in both academy, even by

t
undergraduate students, and industry to accurately simulate the behavior of complex distillation

ip
systems. Process simulators include thermodynamic and transport models that allow accurately
calculate properties of pure components, mixtures and state of the art algorithms to simulate

cr
distillation columns. Using optimization algorithms with these types of models is a challenging
problem because some variables cannot be accessed or modified directly by the user; which

us
sometimes introduces non-differentiabilities.

One interesting approach that makes use of available process simulators and optimization tools
was proposed by Lang & Biegler (2002) These authors proposed a distributed stream-tray

an
optimization method (DSTO) in which the reflux and the feed flow rates can be distributed and
directed to a set of candidate trays according to a differentiable distribution function (DDF).
Using this DDF, the location of the feed, reflux, and other side streams can be treated as
M
continuous instead of integer variables. The drawback, however, is that the DDF function is
highly non-convex, and the method can easily be trapped in local solutions.
d
Caballero et al. (2005) proposed a GDP model –reformulated as a MINLP- that combines the
capabilities of process simulators, taking advantage of the tailored algorithms designed for
te

distillation and property estimation implemented in these simulators. The model iterates
between two problems: a NLP problem, in which the trays are divided in existing or non-existing
p

and a specially tailored MILP master problem. The basic idea consists of modifying the master
by adding to the objective function and to the constraints the 'extra' contributions due to the
ce

addition or deletion of trays to each section of the column. The model probed to be robust, but
the necessity of a tailored master problem is an important drawback that avoids the inclusion of
the model in general flow-sheets.
Ac

3. Problem statement.

Taking into account all the previous comments the problem we will address can be state as
follows. Given a distillation column (or distillation sequence) for performing a given separation
determine the best column or column sequence.

To that end the designer must specify upper and lower bounds to the total number of trays and
feeds / products tray positions as well as the purity (or other requirements) of the final streams.

The «best column« is defined in economic terms (minimize the Total Annualized Cost –TAC-)
although any other objective can be used.

Page 5 of 40
It is assumed that all the component and thermodynamic properties are known and accurate.

The solution will include the optimal operating conditions, number of trays and feed(s) and
product(s) tray positions as well as all the internal flows, temperatures, etc.

In the rest of the paper we first present a novel algorithm that combines state of the art chemical
TM
process simulators (AspenHysys ) with GDP model, without reformulation to MINLP, and that

t
does not require modifications in NLP or master sub-problems, overcoming some of the

ip
difficulties of the previous approaches. Then a set of examples from single columns to complex
sequences illustrates the approach and its advantages and drawbacks. Finally some

cr
conclusions and directions for future work are included.

us
4. GDP disjunctive model for rigorous tray by tray optimization

In this section we propose a novel algorithm for the rigorous design of distillation columns that

an
integrates a process simulator in a generalized disjunctive programming formulation.

The basic idea consists of developing a superstructure that uses as basic unit the distillation
M
section. The optimal distillation column, or column sequence, is obtained by selecting among a
set of column sections with different number of theoretical trays. Figure 4 shows an example of
superstructure for a conventional two-section column.
d

In order to write the detailed GDP model let us first introduce the following index sets,
parameters and variables in the model
te

The index sets:

p

REB [ j | j is a reboiler in the superstructure]

ce

COND [ j | j is a condenser in the superstructure]

Sections [ i | i is a column section]
Vessels [v | v is the column vessel]
Ac

DSi [ k | k is a candidate section formed by Nk trays in column section i]

SVv [ Sections that form part of the column vessel v]

Data:

f Charging factor for annualizing the capital cost. It was calculated by the
equation (1), as recommended by Smith (2005) taking into account the
fractional interest rate per year (r) and the horizon time (n).

Page 6 of 40
 r  r  1
n

f  (1)
(r  1)n  1

CH j Heating utility cost in reboiler j

CC j Cooling utility cost in condenser j

t
ip
NTSi,k Number of trays of column section k in disjunction i

Variables

cr
xI Independent variables ( Degrees of freedom ) in the process simulator

us
TAC Total annualized cost

Di Diameter of column section i

DSi,k

NTi an
Diameter of column section k, in disjunction i

Number of trays in column section i

M
Propertyi Any property of section i that must be calculated.

Yi,k Boolean variable. It takes the value True if the column section k is
d

selected in disjunction i, and False otherwise.

te

The GDP model can be written as follows:

p
ce
Ac

Page 7 of 40
 min : TAC
s.t. TAC  f · Capital cost    CH j Qrebj   CC i Qcondi
j REB i COND
Capital cost   Cost Reboilerj    Cost Condenseri  
j REB i COND

 CostVesselv 
v Vessels

t
(Cost Reboilerj ,Cost Condenseri , Qrebj ,Qcondi ,...)  Hy(xI )

ip
 Yi,k 

cr
 
 (DS , NTS ,..., PropertyS )  Hy(x I ) 
 i,k i,k i,k i,k 
 D  DS 
 

us
i i,k
   i  Sections
k DSi  NTi  NTSi,k 
 
  
 
 

an
 Propertyi = PropertySi,k 

Vessel Costv  f (Di , NTi ,...) i  SVv ; v  Vessels

(2)
M
Y  True, False 
m
 
d

where Hy(·) makes reference to the implicit equations solved by the process
te

TM
simulators. In other words, it is a call to AspenHysys .

Without loss of generality, the model given by equation (2) is a conceptual representation of the
p

actual model. The particular details are case dependent. Some remarks are therefore
necessary:
ce

Even though, in the model there is not an explicit logical relationship that force that to a
given section must be assigned exactly a column section, this logical relationship is
Ac

implicitly forced by the set of disjunctions.

It is possible (usually necessary) to add some constraints in order to satisfice some

problem specifications. For example, to force purity or recovery requirements that
cannot be specified in the process simulator. However, these constraints are problem
dependent.

The disjunctions are used to determine all the necessary data to calculate the cost (e.g.
diameter (D), number of trays (NT)) or any other column section property. The cost of
the vessel is calculated in terms of the values assigned to each of the sections that form
the vessel. The sections that form a given column are controlled through the set
'Vessels', that specifies which sections form part of a given column.

Page 8 of 40
 In equation (2) we do not differentiate between dependent and independent variables.
In an equation based environment the solver deals with the variables and the user does
not have to worry about. However, in a process simulator the selection of independent
variables (degrees of freedom) is critical and problem dependent. Besides, the model
given by (2), is a hybrid model, formed by implicit equations (solved by the process
simulator) and explicit equations. The solver can only control the explicit variables –that

t
appear in equation form- and the independent variables in the process simulator. A

ip
critical issue in the optimization with process simulators is that the failure in the
simulator convergence produces the failure in the optimization. Although it is possible to

cr
introduce safeguards –the optimization can be recovered from a simulator convergence
failure-, repeated failures make the optimization impossible. In the particular case of
distillation columns there are sets of specifications that facilitate the convergence (i.e.

us
reflux ratio and boilup ratio or the recovery of key components). However, specifications
like compositions are usually most difficult to converge. In these last case a better
approach consists of selecting a set of 'easy to converge' specifications and add the

an
difficult constraints as external equations in the model.

The direct implementation of the superstructure presented in Figure 4 is not practical even for
M
the case of a single column, at least for two reasons. First, it results in a very complex model in
the process simulator –Figure 5 shows an example of how the model looks like in
TM
AspenHysys -. The NLP optimization takes large CPU times mainly due to the time to
converge the complete flow sheet each time the solver calls it.
d

Fortunately, it is not necessary to use that 'Brute Force Approach'. Instead, the problem can be
te

reformulated as a Disjunctive Problem with Net Structure (only two term disjunctions appear,
and the second term just state that if a Boolean variable is set to False, all the variables related
p

to that disjunction are set to zero). It is worth noting that all disjunctive problems can be
ce

reformulated as disjunctive problems with net structure, and solved using the Logic Based Outer
Approximation (LBOA) algorithm presented by Turkay & Grossmann (1996). The reformulated
problem is as follows:
Ac

Page 9 of 40
min : TAC
s.t. TAC  f · Capital cost    CH j Qrebj   CC i Qcondi
j REB i COND
Capital cost   Cost Reboilerj    Cost Condenseri  
j REB i COND

 CostVesselv 
v Vessels

t
(Cost Reboilerj ,Cost Condenseri , Qrebj ,Qcondi ,...)  Hy(xI )

ip
 Yi,k 

cr
   Yi,k 
 (DS , NTS ,..., PropertyS )  Hy(x I )   
 i,k i,k i,k i,k   DSi,k  0 
 Di  DSi,k   
    i  Sections

us
    NTS i,k  0 
 NTi  NTSi,k    k  DSi
    
    
   PropertyS  0
 Propertyi = PropertySi,k   i,k 

an
 

 Yi,k i  Sections
k DSi
M
(3)
Vessel Costv  f (Di , NTi ,...) i  SVv ; v  Vessels
Y  True, False 
m
 
d
te

The following remarks are important:

p

In the LBOA algorithm (Turkay & Grossmann, 1996), each NLP is formed by a potential flow
sheet configuration, the rest of variables do not appear in the model. Note that strictly speaking
ce

those variables are set to zero –second term in the disjunctions in equation (3)- but in the
practical implementation, when solving the NLP those variables are dropped from the model. In
this case a potential flow sheet is simply a distillation column (or a distillation sequence). This
Ac

fact is really important and what makes the algorithm really useful. Although, from a conceptual
point of view the superstructure is relatively complex (see Figures 4 and 5) the LBOA algorithm
decomposed the problem in feasible NLPs that, as commented, results to be regular distillation
columns. The number of trays in each section at each iteration are «decided» by the algorithm
without user intervention. Therefore, even though there is complex model formulation, the GDP
algorithm results in a sequence of relatively easy to solve sub-problems. As a consequence the
algorithm could be eventually super-imposed to any process simulator without modifying at all
the simulator internal algorithms.

In a single column, with known pressure profile, there are only two degrees of freedom.
Therefore the CPU time needed to solve each NLP is of a few seconds. Even in systems

Page 10 of 40
involving more than one distillation column these NLPs can be efficiently solve, with the same
degree of difficulty that a regular NLP optimization in a process simulator.

4.1. Implementation details

The logic based outer approximation algorithm has been implemented in Matlab (MATLAB.,
2006.) and completely automatized, i.e. once the model is set there the user intervention is not

t
necessary. For solving the NLP and Master-MILP sub-problems used in the LBOA algorithm we

ip
used the Tomlab optimization environment (Holmström, 1999) that provides a gateway for using
state of the art NLP/MILP solvers. Although it is possible to use the internal NLP optimizers

cr
provided by the process simulator, our experience shows (arguable) that external NLP solvers
like CONOPT (Drud, 1996) or SNOPT (Gill et al., 2002) are more robust and reliable. The MILP
sub-problems were solved using CPLEX (Holmström et al., 2009).

us
Figure 6 shows a scheme of the actual implementation. A description step by step of the
problem is as follows

an
TM
Step 1. In the process simulator (AspenHysys ). Set a distillation column (or distillation
sequence) the total number of trays or the feed(s) tray positions are not important at this stage.
The lack of convergence of the process simulator could eventually produce that the entire
M
algorithm fails. In order to minimize that eventuality the set of specifications must be selected to
make the convergence as easy as possible. For example, usually selecting the reboil ratio and
the boil-up ratio as degrees of freedom results in easy to converge columns (although this is
d

case dependent). At this stage, purity requirements or other constrains can be violated.
te

The selection of thermodynamic models, properties estimation etc., are all in the simulation
environment. All the numerical issues related to the convergence of distillation columns (or
p

column sections) are also maintained in the simulation environment and must be specified in
this stage.
ce

Step 2. Initialization. In the model we specified, for each column section, a set of candidate
column sections with different number of trays. In order to initialize the problem we must solve a
Ac

set of NLP problems that include, at least once, each candidate column section. To this end we
solve a set covering problem to determine the minimum number of feasible flow-sheets that
comply with this condition (Turkay & Grossmann, 1996). For example, consider a single column
with one feed, and two products (distillate and bottoms). Assume that the rectifying section must
be selected among NR different column section, and that the stripping section must be selected
among NS different column sections. The minimum number of initial NLP problems will be equal
to the maximum of NR and NS. Each one of those initial flow-sheets is simply a distillation
column with fixed number of trays and fixed tray position. In general, the number of initial
problems is given by equation (4):

N º Initial NLPs  max  NTSi,k  (4)

Page 11 of 40
If this number is too large, it is possible reduce the number of candidate sections by considering
sections that differentiate in 2, 3, … or n trays. Once the problem has been solved, it is possible
to refine the search by considering values around the optimal solution obtained in previous
iteration.

Step 3. Generate and solve the Master problem. The MILP is generated using the approach
presented in the original LBOA algorithm (Turkay & Grossmann, 1996), with equality relaxation

t
to deal with equality constraints (Kocis & Grossmann, 1987) and the augmented penalty

ip
strategy (Viswanathan & Grossmann, 1990) to ensure that linearizations of non-convex
problems yield always feasible MILP problems. The derivative information for generating the

cr
MILP Master problem is provided directly by the last NLP (or the initial sets of NLPs) and
corresponds to the Jacobian and function gradient in the optimal NLP solution.

us
It is worth mentioning that in the Master problem depends only on independent variables
(degrees of freedom in the flow-sheet) and explicit variables in the model. Those variables
calculated by the flow-sheet (i.e. reboiler and condenser heat loads, compositions, etc.) cannot

an
appear in the Master problem. In other words, the Master is defined in the reduced space of
explicit variables.
M
The master problem includes also an integer cut (Balas & Jeroslow, 1972) to exclude
configurations explored by the algorithm in previous iterations.

Step 4. Solve the new NLP problem. The NLP problem is just a flow-sheet in which all the
d

structural parameters (number of trays, feed and products tray positions) are known. The
te

complexity of these NLPs is equivalent to any regular optimization using a process simulator. As
previously commented, we obtained good and reliable results using an external solver. A
detailed discussion about this topic is out of the scope of this paper. The interested reader can
p

found information, for example in the following references (Biegler & Cuthrell, 1985; Y.D. Lang
ce

& Biegler, 1987)

Step 5. Check convergence. Due to the high non-convexity, the lower bounding property of
the Master does not always hold and therefore, the usual stopping criteria (the last MILP master
Ac

problem and the best NLP upper bound cross each other or the heuristic rule: stop when in two
consecutive major iterations the objective of the NLPs worsen) are likely to provide a local
optimal solution. However, the experience with the outer approximation algorithm for both the
MINLP and logic versions shows that the optimal solution is usually found in the first major
th
iterations (usually in the five first and rarely after the 10 major iteration). So we force the
algorithm to perform at least 10 major iterations. If in these 10 iterations both, worsening of the
objective function in two consecutive NLPs and crossing of the objective values of NLPs and
master problems are simultaneously fulfilled, we stop. Otherwise, the iterations continue until
both conditions hold.

Steps 2 to 5 are completely automatized and do not requires the user intervention.

Page 12 of 40
Some final remarks are of interest:

The problem we are solving in a nonconvex discrete/continuous optimization problem. It is clear

that the approach presented offers no guarantee of convergence to the global optimum. There
are two issues to consider here. The first is that the NLP subproblems are nonconvex, and
therefore local optimizers might get trapped in local solutions. Numerical test, however shows
that most of the NLPs converges to the global optimal solution but if necessary the NLPs can be

t
re-initialized from different starting points to increase the chances of identifying the global

ip
optimum. The second is that the master MILP might cut-off feasible solutions when adding
supporting hyperplanes constructed from nonconvex functions. This limitation is handled (to

cr
some extent) through the addition of slack variables and penalty terms and forcing the stopping
criteria commented in previous paragraphs.

us
As commented the logic based GDP approach generates only feasible NLPs that in this case
result in just a sequence of columns with fixed number of trays and fixed feed position. As a
consequence it is possible super-impose the algorithm in almost any process simulator without

an
modifying the internal algorithms. Eventually it could be included as an "Add in" in any
commercial process simulator.
M
5. Examples
d

Example 1. This first example deals with the separation of a mixture of Methanol, Acetone and
Water (MeOH – Ac – W). Acetone and methanol forms a minimum boiling azeotrope. The
te

objective is to obtain a mixture of acetone and methanol with a combined mole fraction greater
than 0.99 and a combined mole recovery greater than 99%. All relevant data for the example
p

are shown in Table 1.

ce

This case study has been included to provide an example that is easy to reproduce and in
which equations for calculating rigorous costs do not hide the essence of the algorithm. Instead
of rigorous calculation of the total annual cost, we assume that the cost is given by a simple
Ac

expression in terms of the heat loads and number of column trays:

min :0.2Qcond(kW ) QReb(kW )  100(Total number of Trays )

where Qcond and Qreb make reference to the condenser and reboiler heat loads
respectively.

In order to increase the robustness of the algorithm it is convenient to take as much advantage
TM
as possible from the specialized algorithms included in the process simulator (Aspen.Hysys ).
In this case the two constraints (mole fraction and recovery in distillate) can be used as
specifications and therefore the problem is transformed in finding the combination of column
sections with a minimum value of objective function. However, as previously commented, it is

Page 13 of 40
worth remarking that transferring constraints (specifications) to the process simulator is not
always possible because under some specifications the simulator convergence is difficult. Even
though, the NLP solver can usually recover from a convergence failure, in general the lack of
convergence in the simulator makes that all the procedure fail. In those cases a good alternative
consists of selecting a set of specifications that ensure convergence for a large interval of
values (i.e. reboil ratio and boilup ratio), and let the NLP solver to converge the difficult

t
constraints. It is also convenient to carry out a preliminary study to bound the value of variables

ip
to ensure the convergence of the process simulator.

In order to ensure convergence with the current specifications, in this case study the minimum

cr
number of trays for the rectifying sections was fixed to 2; and to one in the stripping section. The
maximum number of trays was fixed to the minimum plus 10 trays. For the case in which the

us
optimal solution lies at one of these limits, the maximum number of trays is increased or the
minimum number of trays is decreased.

First we solve the initialization NLPs ensuring that a particular tray in each section is selected at

an
th
least once, see Table 2. The optimal solution was found in the 9 major iteration with an
objective value of 2578.4 (Qreb =1318 kW; QCond =1304 kW; Number of trays = 10). Table 2
th
shows also the results of the major iterations. Note that in the 7 major iteration the MILP
M
master problem and the best NLP crossed each other, using that stopping criterion the solution
would be 2635.8, which is also a good solution because only differs from the optimal in one tray
and only 2.2% in objective function value. Figure 7 shows the optimal solution.
d

This is a small problem therefore it is possible to systematically check all the alternatives that
te

prove that the solution obtained was also the global optimum.
p

Example 2. In this second example the objective is to separate ethanol from propanol and
ce

butanol, to obtain a distillate with at least 0.98 mole fraction in ethanol and at least 99.5%
ethanol recovery. All relevant data for this example are in Table 1.
Ac

The objective function consists of minimizing the total annualized cost (TAC), calculated as the
sum of operating and annualized investment costs. In this example and the following, the sizing
of the distillation column is done following the procedure proposed by Stichlmair & Fair (1998),
and the investment cost is calculated using the correlations presented by Turton et al (2003).

The annualizing factor of the capital cost (f) was calculated by equation (1), as recommended
by Smith (2005) taking into account the fractional interest rate per year (i) and the horizon time
(n). Of course, changing the annualizing period can lead to different optimal columns, due to the
tradeoff between the capital and operating costs.

In this example the number of trays in each section ranges between 5 and 25, which gives a
column with a maximum of 51 trays (25 per section plus the feed tray). With this configuration

Page 14 of 40
we should solve 20 initial NLPs. To reduce the number of initial NLPs we consider only column
sections that differentiate in n-stages (i.e. n = 2, 3 or 4) and then contract the column around the
best solution.

In the first iteration we assume five potential sections (in both stripping and rectifying sections)
that differentiate in 4 trays, the optimal solution for this first iteration, is obtained for a
configuration with 13 stages in the rectifying section and 21 in the stripping section (35 trays –

t
3
including the feed tray- plus condenser and reboiler) with a total annual cost of $ 564.99·10 . In

ip
the second iteration we consider again 5 potential sections (for both rectifying and stripping
sections) that differentiate in two trays centered in the optimal solution obtained in previous

cr
iteration (e.g. 9, 11, 13, 15, and 17 trays in the rectifying section). The optimal solution in this
3
iteration was $562.8·10 , and was formed by 15 trays in the rectifying section, and 21 in the

us
stripping section. In the third iteration, we consider five new sections that differentiate in just one
3
tray. The optimal solution was $562.15·10 , and was formed by 15 trays in the rectifying section,
and 20 in the stripping section (Total number of trays equal to 36).

an
Starting with 10 potential sections that differentiate in 2 trays or starting with different initial
configurations yield solutions that differs from the one presented above in just one tray. Figure 8
shows the optimal configuration.
M
Example 3. In this example, instead of a single column we optimize a given sequence of
d

partially thermally coupled distillation sequences. Information about thermally coupled distillation
te

can be found, for example in (Agrawal, 1996; Blancarte-Palacios et al., 2003; José A. Caballero,
2009; J. A. Caballero & Grossmann, 2001, 2004; José A. Caballero & Grossmann, 2012;
Gutérrez-Antonio et al., 2011; Rong et al., 2001; Shah & Agrawal, 2010):
p

The objective consists of determining the number of theoretical trays, feed positions in each
ce

column and operating conditions for separating a mixture of Benzene, Toluene, p-Xylene and
Bi-Phenyl using the partially thermally coupled sequence of columns presented in Figure 9a.
The molar fraction of each final product must be greater than 0.99. All relevant data for this
Ac

example can be found in Table 1.

In order to facilitate the convergence of each column, It is convenient to transform the sequence
given in Figure 9a into another thermodynamically equivalent (Agrawal & Fidkowski, 1998; J. A.
Caballero & Grossmann, 2003; Hernandez et al., 2006; Rong et al., 2004) that can be
associated to a sequence of conventional columns –each column with a rectifying and a
stripping section- like the sequence shown in Figure 9b. The simulation of thermally coupled
distillation sequences presents the problem that a thermal couple is formed by two side liquid
and vapor streams connecting two column sections. In other words, each thermal couple
introduces a recycle. The usual approach in modular process simulator to converge the cycles
is by using fixed point methods (Biegler et al., 1997), that has only linear convergence. If we

Page 15 of 40
have to converge the flow sheet each time the optimizer calls the simulator the total CPU time
quickly becomes prohibitive. Besides, in slightly numerical noisy systems (J. A. Caballero &
Grossmann, 2008) the recycles act as noise amplifiers. To avoid all these problems we simulate
the sequence of columns using the procedure proposed by Navarro et al. (2012), that substitute
the two streams forming a thermal couple by the equivalent set formed by a saturated stream
and an energy stream.

t
For column sections 1 and 2 (referred to Figure 9) we assume a set of ten column sections

ip
ranging from 10 to 20 trays. For sections 3 to 6 we assume ten column sections ranging from 5
to 15 trays each one. And for column sections 7 and 8 we assume a set of seven column

cr
sections ranging from 1 to 8 trays each.

As independent variables we use the recoveries of key components in each separation. The

us
optimal solution of the problem is obtained in around 3 minutes of CPU time (Intel Core(TM)2 Quad
CPU Q6600 @ 2.40GHz 2.39 GHz under Windows 7). The optimal solution with the most relevant values
is presented in Figure 10

an
Example 4 This example involves the optimization of an extractive distillation process.
Extractive distillation is used to separate close boiling or homogeneous binary azeotropes by
M
adding an entrainer that must be a higher-boiling component. The proposed case study is
adapted from Luyben (2011). The objective is to separate an isomolar mixture of acetone and
methanol using dimethyl sulfur oxide (DMSO) as entrainer. The system acetone-methanol has,
d
at atmospheric pressure, a minimum boiling homogeneous azeotrope at 77.6 mol% acetone.
The normal boiling point of acetone is 239 K and for methanol is 338 K, while for DMSO is 464
te

K. Relevant data for this case study is presented in Table 1.

Extractive distillation comprises two columns. The first one is the extractive column, which has
p

two feeds. The entrainer is fed to the column above the process feed. The second column is the
ce

entrainer-recovery column. The acetone is recovered at the top of the extractive column with
purity higher than 0.9995 mole fraction, while the methanol with the DMSO is obtained as
bottoms product. In the second column the DMSO is separated from the methanol both with
Ac

purities larger than 0.9995 mole fraction.

In extractive distillation, one of the factors that influence the most the cost is the flow of
entrainer introduced in the first column. The flow must be large enough to facilitate the
separation between acetone and methanol, but at the same time, large flows inside a distillation
column increases the heat duties and column diameters.

The extractive distillation sequence can be divided in five column sections (see Figure 11). Due
to the number of trays in each column section can be very different, and in particular due to in
section 2 the number of trays is considerably larger than in the rest of the sections, we follow a
sequential approach similar to that in example 2. Initially, for section 1 we select a column
section among a set of five column sections that differentiate in two trays ranging from 3 to 11

Page 16 of 40
trays. For sections 2 to 5 we assume a set of 5 column sections for each one that differentiate in
four trays: Between 20 and 40 for section 2, between 10 and 30 for section 3 and between 2
and 18 for sections 4 and 5. Then in successive iterations we consider set of columns that
differentiate in 1 or 2 trays by contracting around the optimal solution of the previous iteration.
Table 3 shows the details of the three major iterations needed to obtain the optimal solution.
Figure 11 shows the optimal solution obtained.

t
The importance of the DMSO flow rate is evident by a sensitivity analysis of the optimal solution

ip
(Figure 12). This Figure shows that the reboiler heat load in the first column decreases when
the DMSO flow rate increases. However, the condenser heat load presents a minimum for

cr
values around 365 kmol/h of DMSO (DMSO to feed ratio equal to 0.73). In the second column,
as expected, the reboiler and condenser heat loads increases with the DMSO flow rate. The

us
minimum total annualized cost is obtained for a DMSO flow rate equal to 344.5 kmol/h.

an
6. Conclusions

This paper has proposed a systematic method for the rigorous design of distillation columns in
which operational conditions (reflux ratio, internal and external flows, etc.) as well as structural
M
parameters (number of trays in each column section and consequently location of feed and
product streams, etc.) are simultaneously optimized.
d

The rigorous optimization of a distillation column, or column sequence, can be performed by

considering a column as a succession of column sections, separated by feeds, products or heat
te

streams. Therefore, the optimal distillation column, or column sequence, is obtained by

selecting, for each column section, among a set of column sections with different number of
p

theoretical trays. This problem is formulated as a generalized disjunctive problem, in which each
m-term disjunction is related with selection among a set of m candidate column sections.
ce

The model is solved using a Hybrid simulation optimization approach by taking advantage of the
effective and reliable numerical methods included in process simulators for converging
Ac

distillation columns as well as the thermodynamic packages, property estimation etc., and at the
same time the flexibility of an equation based environment. Difficult constraints can be
transferred to the explicit equation part increasing the robustness of the optimization process.

The model is solved using the Logic Based Outer Approximation algorithm. The major
advantages of this algorithm are

!! It allows the use of commercial process simulators to perform the rigorous design of
distillation columns or column sections without the necessity of special algorithms but
standard logic based GDP solvers.

Page 17 of 40
 !! Due to the fact that the LBOA algorithm or its modifications solves a series of NLPs that
must correspond to feasible alternatives, and that in the context of distillation
columns/sequences this is a column/sequence with fixed number of trays and fixed
feeds positions, these NLPs correspond to the continuous optimization of a regular
distillation column that can be done by internal (included in the simulator) or external
solvers –in this paper we follow this second approach-. The complex superstructure is

t
reduced, from the point of view of the final user, to a single column or column sequence

ip
in the process simulator.

!! As a consequence the implementation of this algorithm in a process simulator is

cr
relatively easy because there is no needed any modification, or MINLP reformulation.

!! The examples presented, show that the approach is robust and reliable with CPU times

us
lower than 5 minutes in the worst case. However, due to the extremely non-convex
nature of the problem, we can ensure just an optimal solution. Although the experience

an
shows that usually, if not the global optimal, very good solutions are obtained.

!! The simplicity, from the point of view of a final user, make that the algorithm can be
used by designers without an optimization background.
M
Two drawbacks should also be mentioned:
d

!! Even though the model usually get near global optimal solutions, this fact cannot be
te

guaranteed and correspond to the designer the critical analysis of the final solution.

!! Some knowledge about the system is needed. In general it would not be a good idea try
p

to perform a 'blind optimization' from scratch. Adequate bounds on the number of trays
ce

in each section, reasonable initial values and feasible constraints on purity should be
provided. In that sense this algorithm can be considered an excellent tool to
complement other approaches (conceptual design, shortcut methods or any other of
Ac

those outlined in the introduction), and can be used to 'get a rigorous design' from a
preliminary design.

Future directions include the extension to the synthesis of distillation sequences where a full
column can completely disappear in a superstructure, and the integration in a general
superstructure synthesis framework.

Acknowledgements

The author wish to acknowledge financial support from the Spanish Ministry of Science and
Innovation (CTQ2012-37039-C02-02).

References

Page 18 of 40
Agrawal, R. (1996). Synthesis of distillation column configurations for a multicomponent
separation. Industrial & Engineering Chemistry Research, 35, 1059-1071.

Agrawal, R., & Fidkowski, Z. T. (1998). More operable arrangements of fully thermally coupled
distillation columns. AIChE Journal, 44, 2565-2568.

Bagajewicz, M. J., & Manousiouthakis, V. (1992). Mass/heat-exchange network representation

of distillation networks. Aiche Journal, 38, 1769-1800.

Balas, E., & Jeroslow, R. (1972). Canonical Cuts On Unit Hypercube. Siam Journal On Applied

t
Mathematics, 23, 61-69.

ip
Barbosa, D., & Doherty, M. F. (1988). Design and minimum-reflux calculations for single-feed
multicomponent reactive distillation columns. Chemical Engineering Science, 43, 1523-

cr
1537.

Barttfeld, M., & Aguirre, P. A. (2002). Optimal Synthesis of Multicomponent Zeotropic Distillation
Processes. 1. Preprocessing Phase and Rigorous Optimization for a Single Unit.

us
Industrial & Engineering Chemistry Research, 41, 5298-5307.

Barttfeld, M., & Aguirre, P. A. (2003). Optimal Synthesis of Multicomponent Zeotropic Distillation
Processes. 2. Preprocessing Phase and Rigorous Optimization of Efficient Sequences.
Industrial & Engineering Chemistry Research, 42, 3441-3457.

an
Barttfeld, M., Aguirre, P. A., & Grossmann, I. E. (2003). Alternative representations and
formulations for the economic optimization of multicomponent distillation columns.
Computers & Chemical Engineering, 27, 363-383.
M
Barttfeld, M., Aguirre, P. o. A., & Grossmann, I. E. (2004). A decomposition method for
synthesizing complex column configurations using tray-by-tray GDP models. Computers
& Chemical Engineering, 28, 2165-2188.

Bauer, M. H., & Stichlmair, J. (1996). Superstructures for the mixed integer optimization of
d

nonideal and azeotropic distillation processes. Computers & Chemical Engineering, 20,
S25-S30.
te

Bauer, M. H., & Stichlmair, J. (1998). Design and economic optimization of azeotropic distillation
processes using mixed-integer nonlinear programming. Computers & Chemical
Engineering, 22, 1271-1286.
p

Bausa, J., von Watzdorf, R., & Marquardt, W. (1998). Shortcut Methods for Nonideal
ce

Multicomponent Distillation: 1. Simple Columns. AIChE Journal, 44, 2181-2198.

Biegler, L. T., & Cuthrell, J. E. (1985). Improved infeasible path optimization for sequential
modular simulators—II: the optimization algorithm. Computers & Chemical Engineering,
9, 257-267.
Ac

Biegler, L. T., Grossmann, I. E., & Westerberg, A. W. (1997). Systematic methods of chemical
process design. Upper Saddle River, N.J.: Prentice Hall PTR.

Blancarte-Palacios, J. L., Bautista-Valdes, M. N., Hernandez, S., Rico-Ramirez, V., & Jimenez,
A. (2003). Energy-efficient designs of thermally coupled distillation sequences for four-
component mixtures. Industrial & Engineering Chemistry Research, 42, 5157-5164.

Caballero, J. A. (2009). Thermally Coupled Distillation. In C. A. O. d. N. Rita Maria de Brito

Alves & B. Evaristo Chalbaud (Eds.), Computer Aided Chemical Engineering (Vol.
Volume 27, pp. 59-64): Elsevier.

Caballero, J. A., & Grossmann, I. E. (1999). Aggregated models for integrated distillation
systems. Industrial and Engineering Chemistry Research, 38, 2330-2344.

Page 19 of 40
Caballero, J. A., & Grossmann, I. E. (2001). Generalized disjunctive programming model for the
optimal synthesis of thermally linked distillation columns. Industrial & Engineering
Chemistry Research, 40, 2260-2274.

Caballero, J. A., & Grossmann, I. E. (2003). Thermodynamically equivalent configurations for

thermally coupled distillation. AIChE Journal, 49, 2864-2884.

Caballero, J. A., & Grossmann, I. E. (2004). Design of distillation sequences: from conventional
to fully thermally coupled distillation systems. Computers & Chemical Engineering, 28,
2307-2329.

t
ip
Caballero, J. A., & Grossmann, I. E. (2008). An algorithm for the use of surrogate models in
modular flowsheet optimization. Aiche Journal, 54, 2633-2650.

cr
Caballero, J. A., & Grossmann, I. E. (2012). Synthesis of complex thermally coupled distillation
systems including divided wall columns. Aiche Journal, 59, 1139-1159.

Caballero, J. A., Milan-Yanez, D., & Grossmann, I. E. (2005). Rigorous Design of Distillation

us
Columns: Integration of Disjunctive Programming and Process Simulators. Industrial &
Engineering Chemistry Research, 44, 6760-6775.

Ciric, A. R., & Gu, D. (1994). Synthesis of nonequilibrium reactive distillation processes by
MINLP optimization. Aiche Journal, 40, 1479-1487.

an
Drud, A. S. (1996). CONOPT: A System for Large Scale Nonlinear Optimization. Reference
Manual. Bagsvaerd, Denmark: ARKI Consulting and Development A/S.

Dunnebier, G., & Pantelides, C. C. (1999). Optimal design of thermally coupled distillation
M
columns. Industrial & Engineering Chemistry Research, 38, 162-176.

Edmister, W. C. (1943). Design of hydrocarbon absorption and stripping. . Industrial &

Engineering Chemistry, 35, 837-839.
d

Fenske, M. R. (1932). Fractionation of Straight-Run Pennsylvania Gasoline. Industrial &

Engineering Chemistry, 24, 482-485.
te

Fidkowski, Z. T., Doherty, M. F., & Malone, M. F. (1993). Feasibility of separations for distillation
of nonideal ternary mixtures. Aiche Journal, 39, 1303-1321.
p

Fidkowski, Z. T., Malone, M. F., & Doherty, M. F. (1991). Nonideal multicomponent distillation:
Use of bifurcation theory for design. Aiche Journal, 37, 1761-1779.
ce

Gani, R., & Bek-Pedersen, E. (2000). Simple new algorithm for distillation column design. Aiche
Journal, 46, 1271-1274.

Gill, W. P. E., Murray, W., & Saunders. (2002). SNOPT: an SQP algorithm for large-scale
Ac

constrained optimization. SIAM J. Optimization, 979-1006.

Gilliland, E. R. (1940). Multicomponent Rectification Estimation of the Number of Theoretical

Plates as a Function of the Reflux Ratio. Industrial & Engineering Chemistry, 32, 1220-
1223.

Gutérrez-Antonio, C., Briones-Ramírez, A., & Jiménez-Gutiérrez, A. (2011). Optimization of

Petlyuk sequences using a multi objective genetic algorithm with constraints.
Computers & Chemical Engineering, 35, 236-244.

Harwardt, A., & Marquardt, W. (2012). Heat-integrated distillation columns: Vapor

recompression or internal heat integration? Aiche Journal, 58, 3740-3750. .

Hernandez, S., Gabriel Segovia-Hernandez, J., & Rico-Ramirez, V. (2006). Thermodynamically

equivalent distillation schemes to the Petlyuk column for ternary mixtures. Energy, 31,
2176.

Page 20 of 40
Holmström, K. (1999). The Tomlab Optimization Environment in Matlab. Adv. Model Optim., 1,
47-69.

Holmström, K., Anders, O. G., & Edwall, M. (2009). User's guide for TOMLAB/CPLEX v.11.2. In
http://tomopt.com/docs/TOMLAB_CPLEX.pdf (Ed.).

Humphrey, J. (1995). Separation processes: playing a critical role. Chemical Engineering

Progress, 91, 43-54.

Jackson, J. R., & Grossmann, I. E. (2001). A disjunctive programming approach for the optimal

t
design of reactive distillation columns. Computers & Chemical Engineering, 25, 1661-

ip
1673.

Julka, V., & Doherty, M. F. (1990). Geometric behavior and minimum flows for nonideal

cr
multicomponent distillation. Chemical Engineering Science, 45, 1801-1822.

Kamath, R. S., Grossmann, I. E., & Biegler, L. T. (2010). Aggregate models based on improved
group methods for simulation and optimization of distillation systems. Computers &

us
Chemical Engineering, 34, 1312-1319.

Kocis, G. R., & Grossmann, I. E. (1987). Relaxation Strategy For The Structural Optimization Of
Process Flow Sheets. Industrial & Engineering Chemistry Research, 26, 1869-1880.

an
Kookos, I. K. (2003). Optimal Design of Membrane/Distillation Column Hybrid Processes.
Industrial & Engineering Chemistry Research, 42, 1731-1738.

Kossack, S., Kraemer, K., & Marquardt, W. (2006). Efficient Optimization-Based Design of
Distillation Columns for Homogenous Azeotropic Mixtures. Industrial & Engineering
M
Chemistry Research, 45, 8492-8502.

Kraemer, K., Kossack, S., & Marquardt, W. (2007). An efficient solution method for the MINLP
optimization of chemical processes. . Computer Aided Chemical Engineering, 105-110.
d

Kremser, A. (1930). Theoretical analysis of absortion process. . Natural Petroleum News, 22,
43-49.
te

Kunesh, J., Kister, H., Lockett, M., & Fair, J. R. (1995). Distillation: still towering over other
options. Chemical Engineering Progress, 91.
p

Lang, Y. D., & Biegler, L. T. (1987). A unified algorithm for flowsheet optimization. Computers &
Chemical Engineering, 11, 143-158.
ce

Lang, Y. D., & Biegler, L. T. (2002). Distributed stream method for tray optimization. AIChE
Journal, 48, 582-595.

Levy, S. G., & Doherty, M. F. (1986). Design and synthesis of homogeneous azeotropic
Ac

distillations. 4. Minimum reflux calculations for multiple-feed columns. Industrial &

Engineering Chemistry Fundamentals, 25, 269-279.

Levy, S. G., Van Dongen, D. B., & Doherty, M. F. (1985). Design and synthesis of
homogeneous azeotropic distillations. 2. Minimum reflux calculations for nonideal and
azeotropic columns. Industrial & Engineering Chemistry Fundamentals, 24, 463-474.

Lucia, A., Amale, A., & Taylor, R. (2008). Distillation pinch points and more. Computers &
Chemical Engineering, 32, 1342-1364.

Lucia, A., & McCallum, B. R. (2010). Energy targeting and minimum energy distillation column
sequences. Computers & Chemical Engineering, 34, 931-942.

Lucia, A., & Taylor, R. (2006). The geometry of separation boundaries: I. Basic theory and
numerical support. Aiche Journal, 52, 582-594.

Page 21 of 40
Luyben, W. L. (2011). Principles and Case Studies of Simultaneous Design. Hoboken, New
Yersey, USA.: John Wiley & Sons.

MATLAB. (2006.). The Language of Technical Computing. In: The Mathworks Inc.

Mix, T., Dweck, J., Weinberg, M., & Armstrong, R. (1978). Energy conservation in distillation.
Chemical Engineering Progress, 74.

Navarro, M. A., Javaloyes, J., Caballero, J. A., & Grossmann, I. E. (2012). Strategies for the
robust simulation of thermally coupled distillation sequences. Computers & Chemical

t
Engineering, 36, 149-159.

ip
Rong, B. G., Kraslawski, A., & Nystrom, L. (2001). Design and synthesis of multicomponent
thermally coupled distillation flowsheets. Computers & Chemical Engineering, 25, 807-

cr
820.

Rong, B. G., Kraslawski, A., & Turunen, I. (2004). Synthesis and optimal design of
thermodynamically equivalent thermally coupled distillation systems. Industrial &

us
Engineering Chemistry Research, 43, 5904-5915.

Rudd, D. F., & Watson, C. C. (1968). Strategy of process engineering. . New York: Wiley.

Sargent, R. W. H., & Gaminibandara, K. (1976). Optimum Design of Plate distillation Columns.

an
In L. W. C. Dixon (Ed.), Optimization in Action (pp. 267): Academic Press, London.

Shah, V. H., & Agrawal, R. (2010). A matrix method for multicomponent distillation sequences.
AIChE Journal, 56, 1759-1775.
M
Smith, R. (2005). Chemical Process Design and Integration. Chichester: John wiley & Sons,
Ltd.

Soave, G., & Feliu, J. A. (2002). Saving energy in distillation towers by feed splitting. Applied
Thermal Engineering, 22, 889.
d

Stichlmair, J. G., & Fair, J. R. (1998). Distillation: Principles and Practice. New York.: Wiley-
te

VCH.

Turkay, M., & Grossmann, I. E. (1996). Logic-based MINLP algorithms for the optimal synthesis
of process networks. Computers & Chemical Engineering, 20, 959-978.
p

Turton, R., Bailei, R. C., Whiting, W. B., & Shaeiwitz, J. A. (2003). Analysis Synthesis and
ce

Design of Chemical Processes. New York: McGraw-Hill.

Underwood, A. (1948). Fractional distillation of multicomponent mixtures. Chemical Engineering

Progress, 44, 603-614.
Ac

Viswanathan, J., & Grossmann, I. E. (1990). A combined penalty function and outer-
approximation method for MINLP optimization. Computers & Chemical
Engineering, 14, 769-782.

Viswanathan, J., & Grossmann, I. E. (1993). An Alternate MINLP Model for Finding the Number
of Trays Required for a Specified Separation Objective. Computers & Chemical
Engineering, 17, 949-955.

von Watzdorf, R., Bausa, J., & Marquardt, W. (1999). Shortcut Methods for Nonideal
Multicomponent Distillation: 2. Complex Columns. AIChE Journal, 45, 1615-1628.

Yeomans, H., & Grossmann, I. E. (2000a). Disjunctive programming models for the optimal
design of distillation columns and separation sequences. Industrial & Engineering
Chemistry Research, 39, 1637-1648.

Page 22 of 40
Yeomans, H., & Grossmann, I. E. (2000b). Optimal design of complex distillation columns using
rigorous tray-by-tray disjunctive programming models. Industrial & Engineering
Chemistry Research, 39, 4326-4335.

t
ip
cr
us
an
M
d
p te
ce
Ac

Page 23 of 40
Figure Captions

Figure 1. Superstructure by Sargent and Gaminibandara (left) and a possible solution (right).

Figure 2. Superstructure by Viswanathan and Grossmann (left) and a possible solution (right).

t
Figure 3. Superstructure by Yeomans and Grossmann.

ip
Figure 4. Superstructure for a single distillation column. Each column section must be selected
among a set of column sections that differentiate in the number of trays.

cr
TM
Figure 5. Direct Implementation of the proposed superstructure in Hysys . The complexity of
the resulting superstructure makes the direct implementation difficult for more than a

us
single column.

Figure 6. Scheme of the general modeling framework and the hybrid simulation-optimization

an
solution algorithm.

Figure 7. Optimal solution for example 1.

M
Figure 8. Optimal solution for example 2.

Figure 9. Sequence of partially thermally coupled distillation sequence and its

thermodynamically equivalent configurations using to section columns.
d

Figure 10. Optimal solution for example 3.

te

Figure 11. Optimal solution for example 4.

p

Figure 12. Results of the sensitivity analysis for the optimal solution of example 4 in terms of
ce

DMSO flow rate introduced to the first columns.

Ac

Page 24 of 40
Table 1. Data for the Examples

Heat Exchangers
2
Reboiler: U = 820 W/(m K)
2
Condenser: U = 1000 W/(m K)
Utilities
Low Pressure Steam (254 ºC) 17.70 $/GJ
High Pressure Steam (130 ºC) 14.05 $/GJ

t
Cold Water (20-40 ºC) 0.354 $/GJ

ip
Columns
Calculated based on sieve trays.

cr
Stainless steel.
Tray separation 0.609 m.
Sizing following the procedure by Stichlmair & Fair (1998).

us
Economics
Annual interest rate ( i ) 10%
Time horizon (n) 8 years

an
Capital cost correlations from Turton et al (2003)

Example 1 Example 2
Feed 100 kmol /h Feed 100 kmol/h
M
Composition (m.f.) Composition (m.f.)
Methanol 0.4 Ethanol 0.3
Acetone 0.3 1-propanol 0.4
Water 0.3 1-butanol 0.3
d

Pressure 101.325 kPa Pressure 101.325 kPa

Thermal state Saturated liquid Thermal state Saturated liquid
te

Thermodynamics NRTL Hysys default Thermodynamics NRTL Hysys default

Specifications Specifications
Acetone +
p

Ethanol mole
Methanol mole >0.99 >0.98
fraction
fractions
ce

Acetone +
Ethanol
Methanol >99% >99.5%
recovery
recovery
Example 3 Example 4
Ac

Feed 200 kmol/h Feed

Composition (m.f.) Composition (m.f.)
Benzene 0.25 Acetone 0.5
Toluene 0.25 Methanol 0.5
p-Xylene 0.25 Entrainer
Bi-phenyl 0.25 DMSO 1

Pressure 101.325 kPa Pressure 100 kPa

Thermal state Saturated liquid Thermal state Saturated liquid
Thermodynamics Peng Robinson Thermodynamics NRTL Hysys default
Specifications Specifications
Mole fraction of Mole fraction of
each pure >0.99 each pure >0.9995
component component

Page 25 of 40
Table 2. Step by step iterations of example 1.

Initialization
Sub-problem Objective CPU time
Type Nº of trays Feed tray Solver
function (s)++

t
1 NLP 6 4 4920.7 0.998 CONOPT

ip
2 NLP 8 5 2857.8 0.374 CONOPT
3 NLP 10 6 2688.8 0.328 CONOPT

cr
4 NLP 12 7 2739.4 0.265 CONOPT
5 NLP 14 8 2857.7 0.296 CONOPT

us
6 NLP 16 9 3006.0 0.218 CONOPT
7 NLP 18 10 3170.4 0.218 CONOPT
8 NLP 20 11 3344.5 0.312 CONOPT

an
9 NLP 22 12 3525.0 0.187 CONOPT
10 NLP 24 13 3709.8 0.176 CONOPT
Major Iterations
M
1 Master MILP ---- ---- 2625.8 0.078 CPLEX
1 NLP 10 5 2930.4 0.296 CONOPT
2 Master MILP ---- ---- 2650.3 0.280 CPLEX
d
2 NLP 9 5 2862.1 0.280 CONOPT
3 Master MILP ---- ---- 2657.7 0.062 CPLEX
te

3 NLP 11 5 3019.7 0.218 CONOPT

4 Master MILP ---- ---- 2714.6 0.078 CPLEX
p

4 NLP 12 5 3115.9 0.203 CONOPT

5 Master MILP ---- ---- 2785.0 0.156 CPLEX
ce

5 NLP 13 5 3214.6 0.687 CONOPT

6 Master MILP ---- ---- 2863.4 0.093 CPLEX
6 NLP 14 5 3314.1 0.328 CONOPT
Ac

7 Master MILP ---- ---- 2896.3 0.094 CPLEX

7 NLP 9 6 2635.8 0.468 CONOPT
8 Master MILP ---- ---- 2947.1 0.062 CPLEX
8 NLP 15 5 3414.0 0.468 CONOPT
9 Master MILP ---- ---- 2971.4 0.109 CPLEX
9 NLP 10 7 2578.4** 0.250 CONOPT
10 Master MILP ---- ---- 2998.0 0.156 CPLEX
10 NLP 11 6 2773.0 0.234 CONOPT

** Optimal solution.

++Intel Core(TM)2 Quad CPU Q6600 @ 2.40GHz 2.39 GHz under Windows 7

Page 26 of 40
Table 3. External iterations in example 4.

Nº of trays in Nº of trays in
Iteration 1
candidate sections optimal solution
Column 1
Section 1 3, 5, 7, 9, 11 5 DMSO Feed Tray 6

t
ip
Section 2 19, 23, 27, 31, 35 31 Feed Tray 37
Section 3 14, 18, 22, 26, 30 18 Nº of total trays 56

cr
Column 2
Section 1 2, 6, 10, 14, 18 6 Feed Tray 7
Section 2 2, 6, 10, 14, 18 6 Nº of total trays 13

us
6
TAC ($10 /yr) 4.827
Nº of trays in Nº of trays in
Iteration 2
candidate sections optimal solution

an
Column 1
Section 1 3, 4, 5, 6, 7 3 DMSO Feed Tray 4
Section 2 27, 29, 31, 33, 35 29 Feed Tray 33
M
Section 3 14, 16, 18, 20, 22 16 Nº of total trays 50
Column 2
Section 1 2, 4, 6, 8, 10 6 Feed Tray 7
d

Section 2 2, 4, 6, 8, 10 6 Nº of total trays 13

te

6
TAC ($10 /yr) 4.818
Nº of trays in Nº of trays in
Iteration 3
candidate sections optimal solution
p

Column 1
ce

Section 1 1, 2, 3, 4, 5 3 DMSO Feed Tray 4

Section 2 27, 28, 29, 30, 31 29 Feed Tray 33
Section 3 14, 15, 16, 17, 18 16 Nº of total trays 50
Ac

Column 2
Section 1 4, 5, 6, 7, 8 6 Feed Tray 7
Section 2 4, 5, 6, 7, 8 6 Nº of total trays 13
6
TAC ($10 /yr) 4.818 (same as in iteration 2)

Page 27 of 40
 Ac
ce
pt
ed
M
an
us
cr
i
Page 28 of 40
Figure(s)

i
cr
us
Distillate Distillate

an
M
Feed

ed
pt
ce
Ac

Bottoms Bottoms

Page 29 of 40
Figure 1
 i
cr
us
an
M
ed
pt
ce
Ac

Figure 2 Page 30 of 40
 i
cr
us
an
M
Conditional trays

ed
Fixed Tray
pt
Conditional trays
ce
Ac

Figure 3

Page 31 of 40
 i
cr
us
an
M
ed
pt
ce
Ac

Figure 4
Page 32 of 40
 i
cr
us
an
M
ed
pt
ce
Ac

Figure 5
Page 33 of 40
 i
cr
Set up HYSYS flow-sheet:
Modeling language
Column or columns sequence

us
Variable declarations MATLAB (interpreter) Thermodynamics
Disjunctions Convergence parameters …
Logical relationships…

an
Set up communication
MATLAB-HYSYS

M
(Windows COM)
HYSYS
FLOWSHEET

ed
Logic Based Outer Approximation Algorithm

MASTER
Solver
pt
Generation of initial
NLP Solver
(CONOPT/
xI variables
NLP problems
(CPLEX) SNOPT) xD variables
ce
Ac

MILP Master
Generation

NLP sub-problem

Convergence?
Results
Page 34 of 40
Figure 6
 i
cr
min : QReb (kW )  0.2QCond (kW ) 100( N ºof Trays)

us
Q = 1304 kW 70 kmol/h

an
56.9 ºC
MeOH (m.f.) 0.5614

M
Acetone (m.f.) 0.4286
Water (m.f.) 0.01

ed
Methanol (m.f.) 0.4 6 trays
Acetone (m.f.) 0.3
Water (m.f.) 0.3
pt 7th Tray
100 kmol / h
ce
P = 1 atm
3 trays
Ac

96.2 ºC
MeOH (m.f.) 0.0233
Acetone (m.f.) 0.000
Water (m.f.) 0.9767
Q = 1318 kW

xMeOH  xAcetone  0.99(mol fraction)

recovery ( MeOH  Acetone)  0.99
Page 35 of 40
Figure 7
 i
cr
us
Q = 10.24 MW 30.6 kmol/h

an
78.5 ºC
Ethanol (m.f.) 0.98 0

M
1-Propanol (m.f.) 0.019
1-Butanol (m.f.) 0.001

ed
Ethanol (m.f.) 0.3 15 trays
1-Propanol (m.f.) 0.4
1-Butanol (m.f.) 0.3 pt 16th tray
100 kmol / h
ce
P = 1 atm
20 trays
Ac

104.2 ºC
Ethanol (m.f.) 0.002
1-Propanol (m.f.) 0.567
1-Butanol (m.f.) 0.431
Q = 10.27 MW

Page 36 of 40
Figure 8
 i
cr
us
an
Benzene Benzene
( > 0.99 m.f.) ( > 0.99 m.f.)
3
3

M
4 Toluene 4
( > 0.99 m.f.) 1

ed
Toluene
1 ( > 0.99 m.f.)

p-Xylene 2 5
5 pt ( > 0.99 m.f.)
2
7
p-Xylene
ce
6 6 7 ( > 0.99 m.f.)
Ac

8 8
Bi-Ph Bi-Ph
( > 0.99 m.f.) (a) (b) ( > 0.99 m.f.)

Figure 9 Page 37 of 40
 i
cr
2366 kW

us
Benzene
( > 0.99 m.f.)
3

an
D = 1.89 m
154.0 kmol/h 11th Tray
24 Trays

M
83.1 kmol/h D = 1.26 m
200 kmol/h D = 1.47 m 4
1

ed
(isomolar) Toluene
29 Trays ( > 0.99 m.f.)
Feed 15th Tray
2
pt
D = 1.53 m 5 D = 1.26 m
ce
145.1 kmol/h 12th Tray 551.9 kW
27 Trays
274.2 kmol/h p-Xylene
Ac

D = 2.02 m
6 7 ( > 0.99 m.f.)
D = 0.98 m
th
TAC = M$ 2.489 231.3 kmol/h 7 Tray
15 Trays
331.2 kmol/h
D = 2.45 m
8
Bi-Ph
3843 kW ( > 0.99 m.f.)
Figure 10 Page 38 of 40
 i
cr
us
6792 kW
Acetone

an
( > 0.9995 m.f.)
D = 3.0 m
3 Trays

M
DMSO makeup 344.5 kmol/h
4th Tray
D = 2.5 m

ed
2687 kW
29 Trays
Methanol (m.f.)
pt 0.5
Acetone (m.f.) 0.5 Methanol
500 kmol / h 33rd Tray ( > 0.9995 m.f.)
ce
P = 100 kPa D = 1.4 m 6 Trays
16 Trays
D = 2.3 m
7th Tray
Ac

594.5 kmol/h
DMSO (m.f.) 0.5795 6 Trays
Methanol (m.f.) 0.4203
5697 kW Acetone (m.f.) 0.0002 DMSO
( > 0.9995 m.f.)
D = 1.4 m

3848 kW

Figure 11 Page 39 of 40
Figure(s)
4.92

Total annualized cost ($106 /yr)

4.9

4.88

4.86

t
ip
4.84

cr
4.82

us
4.8
300 320 340 360 380 400 420 440 460
DMSO flow rate (kmol/h)

5900 6900

an
Reboiler heat load first column (kW)

Condenser heat load first column (kW)

5850 6880

5800
6860

5750

5700
M 6840

6820
5650
ed
6800
5600

5550 6780
pt

5500 6760
300 320 340 360 380 400 420 440 460 300 320 340 360 380 400 420 440 460

DMSO flow rate (kmol/h) DMSO flow rate (kmol/h)

ce

4200 2780
Reboiler heat load second column (kW)

Condenser heat load first column (kW)

Ac

4150
2760
4100

4050 2740

4000
2720
3950

3900 2700

3850
2680
3800

3750 2660
300 320 340 360 380 400 420 440 460 300 320 340 360 380 400 420 440 460

DMSO flow rate (kmol/h) DMSO flow rate (kmol/h)

Page 40 of 40
Figure 12