c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298

available at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/cherd

Simultaneous optimal synthesis, design and operation of

batch and continuous hybrid separation processes

Tajalasfia M.M. Barakat, Eva Sørensen ∗

Centre for Process Systems Engineering, Department of Chemical Engineering, University College London, Torrington Place,
London WC1E 7JE, United Kingdom

a r t i c l e i n f o a b s t r a c t

Article history: A procedure for the optimal determination of process synthesis, design and operation for
Received 4 July 2007 both batch and continuous separation is presented. The procedure allows the simultane-
Accepted 1 December 2007 ous determination of the optimal process from any specified available process alternatives,
its configuration, its design and corresponding operation in order to perform a given sep-
aration duty. In this work, distillation, pervaporation and hybrid distillation/pervaporation
Keywords: processes are considered as process alternatives.
Process synthesis The optimisation strategy takes into account an overall economic index that encom-
Optimisation passes capital investment, operating costs and production revenues. Furthermore, rigorous
Hybrid separation mathematical models developed from first principles for distillation and pervaporation are
Distillation used.
Pervaporation Case studies are presented and it is found that hybrid distillation/pervaporation pro-
Genetic algorithm cesses are the optimal synthesis solution in most cases and large costs savings can be made
compared to traditional operation based on distillation alone.
© 2007 The Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.

ing for significant savings in terms of energy consumption and

1. Introduction
Distillation is the most commonly used technique for sepa-
rating liquid mixtures within the chemical industries despite
being an energy and capital-intensive process. Many mixtures
commonly encountered in the fine chemical and pharma-
ceutical industries are, however, difficult or impossible to
separate by normal distillation due to azeotropic behaviour,
tangent-pinch or low relative volatilities. Pervaporation has
been hailed as an alternative to distillation for such mixtures
as the separation mechanism is different, relying on differ-
ences in solubility and diffusivity between the components in
the mixture and not vapour–liquid equilibrium as in distilla-
tion. Pervaporation is still, in general, a more costly process
than distillation, however, for some separations, the costs are
comparable or in favour of pervaporation. Recently, hybrid
processes have been proposed where a distillation column
unit and a pervaporation unit are integrated into one process
(Lipnizki et al., 1999). In such a process, the shortcomings of
one method are outweighed by the benefits of the other, allow-
cost. Guerreri (1992) and Van Hoof et al. (2004) reported up to
60% reduction in energy consumption in their studies of sepa-
ration of azeotropic mixtures using hybrid systems compared
to distillation alone. Kreis and Gorak (2006) presented simula-
tion models for hybrid systems with pervaporation and vapour
permeation and predicted that the benefits [of hybrid systems] are
very likely to prevail in the nearest future.
In a hybrid system, the two units can be integrated in differ-
ent ways; the pervaporation unit can be positioned before the
column, after the column, or fully integrated. One example of
a fully integrated system is to position the pervaporation unit
after the column reflux drum outlet and recycle the retentate
stream back to an appropriate location in the column. Further
examples of how the two units can be integrated can be found
in Lipnizki et al. (1999) and will be discussed shortly.
Hybrid distillation/pervaporation processes have been con-
sidered by several authors. Eliceche et al. (2002) carried out
optimisation studies of operating conditions for a continu-
ous hybrid distillation/pervaporation system consisting of an

∗
Corresponding author.
E-mail address: e.sorensen@ucl.ac.uk (E. Sørensen).
0263-8762/$ – see front matter © 2007 The Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.
doi:10.1016/j.cherd.2007.12.004
280 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298

Nomenclature
uo vector of operation variables
A area (m2 ) U heat transfer coefficient (J/m2 K)
Am membrane area (m2 ) UF update factor ($)
ACCk annualised equipment capital costs of process V column vapour flowrate (mol/s)
k ($/year) x vector of state variables
AOCk annualised equipment-operating costs of pro- xi molar fraction of component i in the vapour
cess k ($/year) phase
ci problem constraint i value ximin minimum composition of component i in the
cp heat capacity (J/mol K) mixture
Cex installed heat exchanger cost ($)
Cfeed cost price of feed ($/mol) Greek letters
Ci selling price of product i ($/mol) ı sigma multiplier
Csh installed column shell cost ($)  efficiency
Cut,i utilities cost of equipment i ($/MJ) i penalty term i
CRFi capital recovery factor of i  standard deviation
D column diameter (m)  volumetric flowrate (m3 /s)
f corrected objective function value ˝ objective function value
f(g) objective value of solution g
fi function i Superscripts and subscripts
fs scaled genome fitness anc ancillary
fsum sum of genome fitnesses ave average
Ff feed flow rate (mol/s) b bottom product
Fi product flow rate c column
Fm membrane stage feed rate (mol/s) cond condenser
Fr retentate flow rate (mol/s) ex heat exchanger
Fs side-draw flow rate (mol/s) f, feed feed stream
Fp permeate flow rate (mol/s) h pervaporation membrane system feed heater
h enthalpy (J/mol) hyb hybrid system
K1 Guthrie’s correlation coefficient, column shell in at the inlet
cost L liquid phase
K2 Guthrie’s correlation coefficient, heat m membrane
exchanger cost min minimum
Lr retentate recycle location max maximum
Ls membrane feed location out at the outlet
mf feed amount (mol) p system feed pump
mi product i recovery (mol) r radial direction
mmin minimum product i recovery (mol) reb reboiler
i
M molar holdup (mol) RC reference case
Nc number of components s column side-draw
Nm,s number of membrane modules in stage s t system turbine
Npop number of genomes in genetic population tray tray
Nt number of column trays ut utility
P pressure (Pa) V vapour phase
PA annual profit ($/year) vessel vessel
Pc probability of crossover z axial direction
Pm probability of mutation
Pp permeate pressure (Pa)
azeotropic distillation column connected via a side stream
Ps probability of selection
to a pervaporation unit. They solved the optimisation prob-
Pss population overlap percentage
lem by minimising the operating cost, however, they did
Q heat load (W)
not consider neither the design nor the configuration of the
Qm,h pervaporation heat load (W)
hybrid system. They concluded that the optimum opera-
Rc column internal reflux ratio
tion conditions could result in a 9.7% reduction in operating
Rp permeate offcut ratio
costs.
Rr retentate recycle ratio
Szitkai et al. (2002) optimised the design and operation
tf total batch processing time (min)
of a continuous hybrid dehydration system using an MINLP
ti task i interval time (min)
model to minimise the annual operating costs of a single,
ts batch setup time (min)
post-distillation, hybrid configuration. They concluded that a
T temperature (K)
12% savings in total annual costs can be achieved when using
TA total production time available per annum
a hybrid ethanol dehydration system instead of distillation
(h/year)
alone.
Ts tournament pool size
Kookos (2003) proposed a methodology for the structural
ud vector of design variables
and parametric optimisation of continuous hybrid separa-
 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298
tion systems. He described the superstructure of the hybrid
process using a simplified steady-state mathematical model
where it was assumed that all streams taken from, or returned
to, the distillation column were vapour streams. The method-
ology is therefore not suitable for other membrane processes,
such as pervaporation, or for dynamic systems, such as batch
processes. He outlined that the hybrid systems can potentially
reduce the operating costs by 22% if operated at optimum
conditions. He also proposed work on hybrid process super-
structures, although this only allowed for the optimisation
of the hybrid process and therefore exploring either distilla-
tion or pervaporation alone as the potentially best separation
process, would not be possible.
Van Hoof et al. (2004) compared the economics of
various processes to separate a mixture of IPA and
water (50/50 wt%) to a final IPA concentration of 99.5 wt%.
They considered three configurations: (i) azeotropic distil-
lation, (ii) hybrid distillation–pervaporation, and (iii) hybrid
distillation–pervaporation–distillation. They concluded that
the distillation–pervaporation hybrid process using ceramic
membranes was the most interesting from an economic point
of view and could potentially lead to a saving in total costs
of approximately 50% compared to conventional azeotropic
distillation.
Most studies on hybrid separations in the literature have
been limited to continuous processes. However, most of the
more challenging separations in terms of close boiling or
azeotropic mixtures, where the benefits of a hybrid system
may be the greatest, can be found within the fine chemical
or pharmaceutical industries which are traditionally batch-
wise. The optimal design and operation of batch distillation
columns has received considerable interest in recent years,
particularly in terms of novel column configurations such as
inverted, middle vessel and multi vessel column configura-
tions (Hasebe et al., 1995; Hilmen et al., 1997; Furlonge et al.,
1999; Low and Sørensen, 2002, 2003).
Adding a pervaporation unit to the system, either before,
after or fully integrated, adds complexity to the system but
also more degrees of freedom which, if properly chosen, can
further increase the profitability of the system, particularly for
difficult separations such as that of close boiling or azeotropic
mixtures, and this will be explored in this paper. The design
engineer is faced with a difficult task: to determine not only
the best design and operation of the separation process, but
also which separation technique to use and, if considering a
hybrid system, how the two units should be combined. The
objective of this work is thus to propose an optimal process
synthesis procedure that allows the simultaneous determina-
tion of the optimal process type (e.g. distillation, pervaporation
or a hybrid of the two), as well as its configuration, design and
operation for a given separation duty. The procedure can be
extended to any number of separation process alternatives,
however, the discussion in this work will be limited to dis-
tillation, pervaporation and hybrid distillation/pervaporation
processes. It will be assumed that the decision of operating
batch-wise or continuous has already been made as the inclu-
sion of this important synthesis decision is, for the moment
at least, beyond the scope of this work.
In the next section, the separation synthesis problem is
presented, followed by the objective function formulation and
the optimisation problem definition. The mathematical mod-
els used in this study are then given together with an overview
of the optimisation strategy. Batch separation will be consid-
ered first as this is more general and, in terms of optimisation,
281
is a more challenging problem, and continuous operation will
then be considered thereafter. The separation process synthe-
sis problem and its solution strategy are applied to several
case studies which consider the separation of tangent-pinch
mixtures of acetone and water. It will be shown that the
proposed procedure is a powerful tool which can be applied
in simultaneously determining the optimal separation tech-
nique, its design and its operation for a given separation
problem. Through the case studies, it will be shown that large
cost savings can be made by considering hybrid systems, but
that this depends on the scale of operation.
2. The separation synthesis problem
2.1. Superstructures
The optimal synthesis of the batch separation process super-
structure is presented next followed by the corresponding
continuous superstructure. The superstructures incorporate
three separation processes: distillation, pervaporation and
hybrid distillation/pervaporation processes. The superstruc-
tures proposed here are shown in Fig. 1 and allow not only for
the most economical process to be selected, but also to simul-
taneously determine its optimal design and operation in order
to carry out the required separation duty optimally.
The configuration decision of the batch superstructure is
mainly characterised by the location of the feed and the prod-
uct withdrawal streams. The column reboiler acts as the feed
tank. Material can either flow from the reboiler feed tank to
the distillation column or to the membrane unit or to both. If
the flow is to the distillation column, it can either be directed
to the bottom of the column section or be distributed to the
trays. If the reboiler vapour boilup is zero, there is no flow
to the distillation column and the liquid flow rate from the
reboiler feed tank to the membrane unit is optimised. The
configuration in this case will either be a batch pervaporation
or a pre-distillation hybrid (i.e. a pervaporation unit followed
by a distillation column). If the vapour boilup is zero and the
retentate is recycled back to the reboiler feed tank, the configu-
ration is that of a batch pervaporation unit and the distillation
column does not exist.
If the vapour boilup from the reboiler feed tank is non-zero,
and there is no withdrawal from the column to the mem-
brane unit, including from the reboiler feed tank or the reflux
drum, the configuration is that of a batch distillation col-
umn and the membrane unit does not exist. Otherwise, if the
boilup is non-zero, but there is withdrawal from the column to
the membrane unit, including from the reboiler or the reflux
drum, then the configuration is that of a hybrid process.
The pervaporation membrane separation stage used in this
superstructure consists of a number of identical membrane
modules Nm,s connected in parallel where the membrane
stage feed stream is assumed to be distributed evenly between
the membrane modules and therefore a single mathematical
model can be used to describe the modules. This method, pro-
posed by Marriott and Sørensen (2003b), was found to reduce
the computational time significantly.
A rigorous distillation column tray model is employed and
each tray is modelled to accommodate three (for batch) or
four (for continuous) extra potential streams in addition to
the regular vapour and liquid inlets/outlets (Li and Vi ) to the
neighbouring trays (see Fig. 2). The first stream is the reboiler
vapour inlet stream (dotted lines in Figs. 1 and 2) which allows
282 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298
Fig. 1 – (a) Batch and (b) continuous separation processes
synthesis superstructures.
for the optimisation of the number of trays through a Special
Ordered Set 1 (SOS1) whereby only one out of N options can be
selected as suggested by Viswanathan and Grossmann (1993).
They stated that their suggested MINLP model for the selection
of the optimal number of trays can be improved by introduc-
ing SOS1 methods, although no results to support this were
presented. (This work shows that this is indeed the case.) The
second extra stream in Fig. 2 is a side draw stream to the per-
Fig. 2 – Superstructure tray model.
vaporation unit in a hybrid configuration if the tray is selected
as a membrane feed tray. The third extra stream is an inlet
stream from the pervaporation unit in a hybrid configuration
if the tray is selected as a retentate recycle tray. Note that in
the proposed procedure, only one membrane feed tray and
one membrane retentate recycle tray is allowed to be selected
along the column, although both of these streams are allowed
to exist in the same tray (although this might not be a very
practical option). The fourth stream is the feed stream inlet, if
the tray is selected as a distillation feed tray in a continuous
column.
2.2. Problem definition and objective function
The objective of the separation process synthesis is to deter-
mine the optimal separation process which results in the most
economical benefit when processing a given separation task.
To achieve this objective, the optimal configuration, design
and operation must be considered simultaneously based on an
objective function that encapsulates capital investment, oper-
ating costs and production revenues.
There is, however, a trade-off between capital invest-
ment in terms of equipment and performance on the one
hand, and between operational decisions and performance
on the other. When considering a distillation column for
instance, it is possible to design the column with a low num-
ber of trays operating at high reflux ratio, or alternatively,
design the column with more trays and operating at lower
reflux ratio, and still achieve the same separation require-
ments. The decision will, however, clearly have an impact
on the profitability of the process. The synthesis problem is
further complicated when a batch process is designed for
multiple separation duties. This is commonly the case in
industry as batch processes are utilised as multi-purpose units
because of their flexibility in accommodating changes in sep-
aration duty (Low and Sørensen, 2004). For such cases, the
optimal product synthesis tends to be biased toward those
separation duties that are performed more frequently than
others. High product recovery generally leads to longer batch
times than does low recovery. However, high recovery oper-
ation generally incurs higher operating costs and therefore
the balance between revenue and operating costs must be
considered.
The optimal design and operation of a separation process,
as it is considered in this work, is determined as the most
economical process design and corresponding operating pol-
icy that will satisfy all specified separation requirements and
constraints. The performance index is expressed as an eco-
nomic model of capital cost (annualised), sales revenue and
operating costs. For the simultaneous optimisation of config-
uration, design and operation using a process superstructure,
the optimal solution will be a trade-off between lower cap-
ital and operating costs against higher production revenue,
thus the objective function must be formulated to reflect these
trade-offs. The performance index for the optimal configura-
tion, design and operation of separation systems as used in
this study is an overall profitability function encapsulating all
mentioned costs and revenues and is for a batch system given
by
 Nc 
Cm
i=1 i i
− Cfeed mf
PA = TA − AOCk − ACCk ,
tf + ts
k = c, m, hyb (1)
 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298 283

and for a continuous system by such that

 0.3

Nc
Am Fm Fp
PA = Ci Fi − Cfeed Ff − AOCk − ACCk , k = c, m, hyb (2) ACCm = ACCm,RC (6)
Am,RC Fm,RC Fp,RC
i=1

where ACCm,RC is the annualised capital costs of the reference

where Ci and Cfeed represent the unit costs of product i
case (see Appendix A), Am and Am,RC represent the membrane
and feed, respectively, mi and mf are the quantities of on-
area of the system and that of the reference case, respectively.
specification product i and of feed, respectively, for a batch
Fm and Fm,RC represent the feed flow rate to the membrane sys-
system and Fi and Ff the flow rates of product i and feed,
tem and that of the reference case, respectively. Fp and Fp,RC
respectively, for a continuous system. TA represents the total
represent the permeate flow rate of the membrane system
available time for production per annum and tf and ts denote
and that of the reference case, respectively. Membrane stage
the processing time and the setup time of the batch process,
flow rate, membrane area and permeate flow rate are the main
respectively.
parameters affecting membrane system performance and are
The annualised capital and operating costs are represented
hence included in the cost equation for comparison with the
by ACCk and AOCk , respectively, where k is the unit index (dis-
reference case. The scale exponent is estimated to be 0.3, the
tillation, membrane or hybrid).
lowest recommended value (IChemE, 1988) for accurate pre-
diction of batch pervaporation capital costs.
2.3. Distillation and pervaporation costs
The feed tank heater and a permeate side cooler, in addi-
tion to the membrane feed pump and the permeate turbine,
The optimisation procedure proposed in this paper can be
are considered to be the main contributions to the operating
used with any costing approach. In the following, a descrip-
cost. The operating costs AOCm for the membrane process are
tion will be given of the distillation and membrane costs used
therefore for the batch system given by
in the following case studies. The derivations of these costing
indices are detailed in Appendix A.
AOCm = (Cut,m,h Qm,h + Cut,m,cond Qm,cond + Cut,p Qm,p
 
2.3.1. Distillation costs TA
+ Cut,t Qm,t ) (7)
The annualised capital costs for a distillation column, ACCc , tf + ts
is found using the Guthrie’s correlation from values of a ref-
where
erence case distillation column as discussed by Longsdon et
al. (1990). These costs are then updated with an update fac- p out
Qm,p = (P − Pin
p ) (8)
tor (UF) assuming a similar trend as that of the Association of p p
Cost Engineering (ACE) index between 1975 and 1985 (IChemE,
1988). The update factor is then estimated to bring the distil- and
lation column costs presented by Longsdon et al. (1990) from  
1990 cost values to 2005 cost values allowing for fair compar- t Pout
t
Qm,t = Pin ln (9)
ison with recently obtained pervaporation membrane costing t t Pin
t
data. The resulting column costs were verified against data
provided by Sulzer Chemtech (2005) and were found to agree where Cut,i represent cost of utilities of equipment i, Qm,h and
with current column costs. The annualised capital cost ACCc Qm,cond represent the pervaporation unit total feed tank heat
is for the distillation column therefore given by duty and permeate condenser cooling duty. Qm,p and Qm,t are
the total energy load of the membrane unit feed pump and per-
ACCc = UF(K1 Nt0.802 V 0.533 + K2 V 0.65 ) (3) meate vacuum turbine, respectively. TA represents the total
available time for operation per annum and ts and tf repre-
The main contributions to the operating costs are assumed sent the process startup and batch time, respectively. i is the
to be the reboiler heating and condenser cooling duties. The flow rate entering and i is the efficiency of the equipment i
operating costs AOCc for the distillation column is therefore (i = pump, vacuum turbine). Pin i
and Pout
i
represent the pressure
for the batch system given by at the inlet and outlet of the relevant equipment i, respectively.

 
TA 2.3.3. Hybrid distillation/pervaporation costs
AOCc = (Cut,reb Qreb + Cut,c,cond Qc,cond ) (4)
tf + ts The annualised capital costs ACChyb and the operating costs
AOChyb for the hybrid batch distillation column are the sum-
and for the continuous system by mation of the cost contributions of the constituent processes
and are therefore given by
AOCc = Cut,reb Qreb + Cut,c,cond Qc,cond (5)
ACChyb = ACCc + ACCm (10)
where Cut,i represents the utility cost of unit i and Qreb and
Qc,cond represent the total reboiler heating and condenser cool- AOChyb = AOCc + AOCm (11)
ing duties, respectively.
2.4. Optimisation problem formulation
2.3.2. Pervaporation costs
The annualised capital cost for a membrane unit is also corre- The objective of the synthesis procedure is to maximise the
lated using the Guthrie’s correlation from values of a reference profitability defined by the objective function PA , subject to
case including pumps and heaters (Sulzer Chemtech, 2005) process type, process model equations and all separation duty
284 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298
constraints. The optimisation problem is therefore given as
follows.
Given a mixture of amount mf with number of components
Nc to be separated, minimum product purities ximin , minimum
product recoveries mmin
i
or Fimin , price structure of feed and
products, Cfeed and Ci and total production time available per
annum TA ; determine the optimum set of design variables ud ,
and the optimum set of operation variables uo , to achieve the
maximum objective function value PA :
max PA
ud ,uo
subject to the mathematical process model of the separation
process; where x is a vector of process state variables, and
ud and uo denote the vectors of design and operating control
variables, respectively:
• material balance, equilibrium, summation of fraction and heat
f (x(t), x(t), t, ud , uo (t)) = 0 (batch);
•
f (x, x, ud , uo ) = 0 (continuous)
and the product purity constraints imposed which must be
satisfied at all times for the continuous system and at the end
of the batch for the batch system:
ci (tf ) ≥ cimin where ci = [xi , mi ] and i = 1, . . . , Nc
and the physical and optimisation bounds of the design and
operating control variables, respectively:
umin
d ≤ ud ≤ umax
d
umin max
o (t) ≤ uo (t) ≤ uo (t) (batch);
umin
o ≤ uo ≤ umax
o (continuous)
In this work, the set of operating variables for the distillation
process includes vapour boilup rate and column reflux ratio
profile, i.e. uco = {V, Rc }, both of which for the batch system can
vary as functions of time in the specified batch task intervals
ti where the total batch time tf = ti . The vapour boilup rate
can subsequently be used to determine the diameter of the
column (e.g. using Guthrie’s correlation, D ∝ V0.5 ) as well as the
reboiler and condenser heat loads. Design variables include
the optimal number of trays Nt , i.e. ucd = {Nt }.
For the pervaporation process, the set of operating vari-
ables include retentate recycle ratio Rr , permeate pressure Pp ,
permeate off-cut ratio Rp and feed tank heat load Qm,h , i.e.
umo = {Rr , Pp , Rp , Qm,h }, all of which for the batch system can
vary as functions of time. Rr is defined as the fraction of reten-
tate recycled back to the column or the feed tank (i.e. Rr = 1
means all retentate is recycled and Rr = 0 means all flows to the
retentate product accumulator). Rp is defined as the fraction
of permeate to the off-cut tank (i.e. Rp = 1 means all perme-
ate flow to the off-cut tank and Rp = 0 means all flows to the
permeate product accumulator). The set of design variables
include number of membrane modules Nm,s per stage s, i.e.
umd
= {Nm,s }.
For the hybrid distillation process, the set of operating vari-
ables and design variables are a combination of the previous
two processes with two additional design variables for the
membrane feed location (from the column) Ls and the reten-
tate recycle location (back to the column) Lr .
3. Process models
The optimisation procedure outlined in this work can be used
with process models of any modelling complexity although
the confidence in the results will of course depend on the
accuracy of the models. The fidelity of the dynamic models
used here is a compromise between accuracy and computa-
tion time. It is assumed that once the optimal solution has
been obtained using the proposed procedure, a more accurate
simulation model is used to verify that the optimal solution is
(12)
indeed the best process alternative.
3.1. Distillation model
The optimisation procedure presented in this work is based on
a rigorous tray distillation column model. The model considers
stage-by-stage column dynamics which is described by a set of
balance (MESH) equations. The main model features include:
(13)
1. Constant molar liquid holdup and negligible vapour holdup on
trays: This eliminates the specification of detailed tray
hydrodynamics and flow characteristics that are usually
not available during the preliminary design screening
stage. Nevertheless, the feature disposes of the common
(14) assumption of negligible tray liquid holdup which may lead
to unrealistic designs.
2. Fast energy dynamics: This energy balance disposes of the
assumption of constant molal overflow which is often
used in distillation modelling. The constant molal over-
(15) flow assumption implies that the heat of vaporisation of
the components in the mixture is the same which is often
unreasonable for difficult separations (e.g. acetone–water:
water = 9.7 kcal/mol and acetone = 7.2 kcal/mol (Perry and
Green, 1984)). Assuming fast energy dynamics instead of a
(16) full energy balance means the energy balance is algebraic
instead of differential which makes the model easier and
faster to initialise and solve.
3. Rigorous thermodynamic models: This disposes of the
assumption of constant relative volatility through the use
of liquid and vapour densities, viscosities, enthalpies and
fugacities.
In order to limit the size of the model to allow for solu-
tion within reasonable computational time and to eliminate
the need to obtain commonly unavailable correlation param-
eters, certain assumptions have been retained. These include
total condensation, no entrainment effects, no down-comer
dynamics, adiabatic operation, phase equilibrium and perfect
mixing. However, these assumptions can easily be eliminated
where relevant information and correlation parameters are
available although the solution time for the optimisation may
then increase considerably.
3.2. Membrane model
The mathematical model used in this study describes the
performance of hollow fibre pervaporation membrane mod-
ules and has the following features (similar to Marriott and
Sørensen, 2003a):
1. Multi-component systems: This allows for multi-component
systems to be considered when investigating multi-
component hybrid distillation/membrane processes.
 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298
2. Non-isothermal flow: This feature considers the tempera-
ture variations within the membrane module. It is required
where significant temperature changes occur, such as that
in pervaporation membranes, as the permeation flux is
highly dependent on these variations.
3. Rigorous thermodynamic models: This disposes of the
assumption of constant physical properties that can oth-
erwise have a large impact on the model accuracy when
significant temperature, pressure or concentration changes
occur.
The model features a 1D plug flow pattern through the
membrane fibres and module shell. Concentration variation
perpendicular to the bulk flow direction is neglected and
prefect mixing throughout is assumed. The membrane char-
acterisation equations reflect the membrane type used and
can easily be modified. In the case studies, a pervaporation
membrane described by Tsuyumoto et al. (1997) has been inte-
grated.
3.3. Hybrid model
The mathematical model of hybrid distillation/pervaporation
is a combination of the distillation and pervaporation models
outlined above.
3.4. Initialisation of batch process
For the batch process, the startup period of operation, i.e.
from when the units (column and membrane module) are
initially dry and cold feed is placed in the feed tank or the
reboiler drum and the mixture is heated up to the boiling point,
is not considered as this leads to an optimisation problem
which is extremely difficult to solve due to discontinuities.
This assumption was indicated by Sadomoto and Miyahara
(1983) to be a reasonably accurate approximation for tray col-
umn operation. All accumulators are assumed to be empty
initially.
4. Solution methodology
The simultaneous consideration of optimal design and oper-
ation of separation processes translates into an optimisation
problem with both discrete (e.g. number of trays and number of
membrane modules) and continuous variables (e.g. reflux and
recycle ratios). Furthermore, the nonlinear dynamic models
used here, as well as the nonlinear objective function defined
earlier (Eq. (12)), transform the problem into a complex mixed
integer dynamic optimisation (MIDO) problem. MIDO prob-
lems are difficult to solve using conventional optimisation
techniques due to the high non-convexity and complex search
space topography nature of these problems and the combina-
tion of integer and continuous variables, and there is much
ongoing research on developing robust and practical solution
algorithms. The proposed separation superstructure is in this
work solved using a steady state genetic algorithm (GA) opti-
misation framework.
GA is a class of evolutionary algorithms and was first
introduced by Holland (1962) as a general model of adap-
tive computer processes, but subsequently has largely been
adopted as a stochastic optimisation method. The GA is
inspired by the natural genetic process, and thus candidate
solutions are encoded as genomes that contain a set of genes
285
representing the decision variables analogous to the DNA in a
natural organism.
In this work, a given solution set consisting of all decision
variables are coded in the genome as direct real and integer
values instead of converted binary bits and mapping coding
which has been found to be less efficient (Coley, 1999). The ini-
tial population is created randomly. Solutions are assigned a
fitness score based on the annual profitability of each genome
as given by Eqs. (1) or (2). A penalty function procedure is
applied when necessary to encourage the GA to drive the pop-
ulation towards feasibility.
The choice of parameters for the GA can have a consider-
able impact on the efficiency and accuracy of the algorithm as
the parameters influence the solution quality and the speed
of convergence by governing the algorithm evolution in using
previous generation knowledge and exploring new unknown
areas in the search space. The parameters affecting the GA
performance are: the population size in each generation Npop ,
the population overlap (steady state GA) Pss , the mutation rate
Pm , the crossover rate Pc , the penalty function implemented
and the termination criterion. In this work, a comprehensive
sensitivity analysis was performed as outlined in Appendix
B to investigate the effect of these parameters on the perfor-
mance of the algorithm used in the optimisation of separation
processes. As a result of this analysis, the GA procedure imple-
mented in the case studies uses a roulette wheel selection
scheme, a population size Npop of 100 genomes, a replacement
rate Pss of 75%, a crossover rate Pc of 75%, a mutation rate Pm
of 10%, and a stopping criterion based on a maximum number
of generations (typically 200 for 15,000 function evaluations).
These GA parameter values have demonstrated to yield good
results in the sensitivity studies as outlined in Appendix B.
The GA frameworks used in this work are implemented
in a C++ programme based on the genetic algorithm library
GALib (Wall, 1999). Function evaluations are performed using
gPROMS gSERVER (PSE, 2005) which solves the process models
for the particular genomes initialised by the genetic algorithm
programme. The value of the objective function, along with
its constraint values, is then passed back to the genetic algo-
rithm programme for assessment. Multiflash (Infochem, 2005)
is used to estimate the thermodynamic model states.
4.1. Genome coding
The set of decision variables representing a solution to the
optimisation problem are translated into genes to form the
solution genome. Where the optimisation of a batch opera-
tion process is considered, the operational decision variables
uo (t) are each parameterised in a piecewise-constant manner
within the batch time intervals ti such that the operational
variable uo,ti represent its value during the relevant batch time
interval ti where the total batch time is tf = ti and ud repre-
sents the vector of time invariant design and configurational
parameters:
genome = (ud , uo,ti , . . . , uo,tf , ti , . . . , tf ) (batch) (17)
The parameterisation of the operational decision variables
uo (t) is not required in the case of a steady state optimisa-
tion problem and time invariant operational parameters uo
are considered instead of continuous operation:
genome = (ud , uo ) (continuous) (18)
286 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298
Each gene has its own bounding range and can represent
a discrete, continuous or logical decision variable which is
particularly helpful where configurational, design and oper-
ational decision variables are considered simultaneously.
4.2. Fitness
The overall algorithm is driven by the fitness of the candi-
date solutions (Eqs. (19) and (21)) which is determined by a
problem-specific objective function (Eq. (12) based on Eqs. (1)
or (2)). The fitness also determines the survival of each can-
didate solution from one genetic generation to the next, and
hence the progression of the optimisation procedure towards
convergence.
The GA is typically initialised with a population of genomes
with random genes from within the search space instead of
a single guess or genome. The populations are then manip-
ulated through the use of three genetic operators; selection,
crossover and mutation, to direct the evolution of the algo-
rithm towards the final population.
4.3. Steady state GA
A steady state genetic algorithm (Syswerda, 1989) is used in
this work which utilises overlapping populations. Firstly, an
initial population of a specified size Npop is generated ran-
domly. Then, in each generation, the fitness of each genome
is evaluated based on the objective function value and con-
straints. Based on the fitness function of each genome, the
algorithm creates a new set of temporary genomes via the
three genetic operators, i.e. selection, crossover and mutation,
and adds these to the previous population. The percentage of
population overlap specified, Pss , governs the number of new
genomes to be created in each generation and is known as the
population criterion or replacement rate. The new genomes
may or may not be promoted to the next generation, depend-
ing on whether they are better or weaker than the rest in
the temporary population. This allows the retention of fitter
genomes for use in the next generation as well as provides
the opportunity to discard new genomes that are weaker than
those of the parents’ generation. This is done by simply rank-
ing the genomes according to their objective function values.
The algorithm thus removes the weaker genomes in order to
return the population to its original size.
4.4. Constraints handling and solution infeasibility
The optimisation problems considered in this work, as most
problems within separation system optimisation, are con-
strained by the (final) product purities and recoveries. The
genome obtained as the optimal result of the GA evolution
must necessarily be feasible, that is, satisfy all constraints
considered. A mechanism is therefore required to check the
feasibility of the evaluated objective value. A method of hard
(absolute) constraint handling is needed when considering
the separation of difficult mixtures. This will ensure that
infeasible solutions are not preferred over those that satisfy
the constraints but resulted in lower objective values (see
Appendix B).
In the optimisations considered in this work, the purity and
recoveries of the final products (xi,tf and mi,tf ), are checked
against the constraints, (ximin and mmin
i
), respectively. If a
solution represented by the genome falls short of these
requirements, a corrected objective function is then assigned
using a penalty function depending on the amount of constraint
violations experienced such that:
⎧
⎪
⎨ PA when feasible
f = Nc 
˘i=1 when PA ≥ 0 (profit but infeasible)
i
⎪
⎩ P (2 − ˘ Nc  ) when P < 0 (loss and infeasible)
A i=1 i A
(19)
where
⎧ 
⎪
⎪ cimin − ci (tf )
⎪
⎨ 1 − if ci (tf ) < cimin
minci
i = (20)
⎪
⎪ where c = [xi , mi ], ∀i = 1, . . . , Nc
⎪
⎩
1 otherwise
where i denotes the penalty function for each of the Nc con-
straints, PA is the original objective function value and f is the
corrected objective value or fitness. Note that the penalty func-
tion and the terms profit and loss of the objective function are
related to the optimisation problems considered in this work
where the ultimate objective is to maximise the objective func-
tion, here the economic performance index PA , and a solution
genome may therefore return a negative value (i.e. loss).
It should be noted that the genomes may represent unre-
alistic or impractical solutions especially when the bounds
set for the design and operational variables are generous. For
instance, the maximum bounds on the boilup rate may cause
the reboiler to run dry during the objective function evalua-
tion and thus will stop the simulation due to the infeasibility
of the solution values chosen. In such a case, a very low or
a very high fitness, depending on whether the objective is to
minimise or maximise the objective function, is automatically
assigned to the genome so that the probability of it being pro-
moted to the next generation is less than those with feasible,
or better infeasible, solutions.
4.5. Fitness scaling
Solution set or genome fitness is a measure of the merit of
that genome and it is proportional to the value of the objec-
tive function achievable by the genome. Fitness scaling may be
required to compensate for the slow convergence during the
later stages of the algorithm where the difference between
the best genome and the worst, or the mean, is not large (e.g.
profit of 12.30 and 12.29 million). The algorithm needs to dif-
ferentiate between these numerical differences for evolution
to continue towards the global optimal. A sigma truncation
scaling technique (Coley, 1999) is utilised in this work to over-
come both premature convergence and the slow finishing
difficulties. Fitness scaling works by pivoting the fitness of
the population members about the average population fitness
value using the standard deviation of the population. The scal-
ing of the fitness score for each genome is therefore given
by
fs = f − (fave − ı f ) (21)
where fs the scaled fitness, f is the corrected objective value
of the genome (Eq. (19)), fave is the average corrected objec-
tive value of the population, ı is the sigma multiplier (1 ≤ ı ≤ 3;
typically 2) and f is the population standard deviation. The
scaled fitness value fs is then utilised by the genetic algorithm
 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298 287

Fig. 3 – Roulette wheel selection.

selection operator as detailed next. Fig. 4). This is required where the genes hold variables of dif-
ferent relevance in the problem domain (e.g. number of trays,
4.6. Selection reflux ratio, etc.).

The selection operator is used to generate offspring from cur-
rent genomes. In order to maximise the chance of selecting
good genomes at the expense of worse ones, a fitness-based
selection criterion, such as roulette wheel (Coley, 1999), is
adopted.
In a roulette wheel procedure (Fig. 3), the probability of
individual selection is proportional to its fraction of the sum-
of-fitnesses (fsum = fs ) of the whole population. The slot of
which a particular genome occupies within the roulette wheel
is defined as
2fs
slot =
fsum
4.8. Mutation
The mutation operator is used to introduce new genetic mate-
rial into the offspring genomes (see Fig. 4). Offspring genomes
undergo Gaussian type mutation with a probability of Pm .
For a randomly generated number Randm ∈ {0, 1}, the gene
is mutated if Randm ≤ Pm using a Gaussian function around
the current value, otherwise the gene is not mutated. In case
the mutated gene results in a value outside the original gene’s
bounding range, the mutated gene value is reset to the violated
bound. Rounding to the nearest integer is applied to discrete
(22) genes.

4.9. Termination criteria

For every parent selection step, the roulette wheel is spun and
where the ball lands, that individual is selected for reproduc-
During the sequential evolution from one generation to the
tion in the new generation. The position of the ball landing is
next, the fitness of the best genome and the average fitness
determined through the following steps:
of the population in each generation increase towards the
global optimum. The GA will continue evolving until there is
1. Choose a random number Rands ∈ {0, fsum }.
no longer diversity in the population, i.e. the best and the aver-
2. Add fitness, one at a time, until fsum ≥ Rands .
age are the same. There are a number of ways to terminate the
3. The genome whose fitness was last added before
fsum ≥ Rands is the selected genome.
4. Repeat to select the second parent genome.

The parent selection is repeated until enough offspring has

been produced to fill the next genetic generation. Because of
the stochastic nature of this selection, there is no guarantee
that a good genome will be selected for crossover, there is,
however, a higher chance of their selection as they occupy a
bigger proportion in the wheel.

4.7. Crossover

After selection, the genes of selected parents are then

combined using the crossover operator. The probability of
crossover Pc determines the amount of genetic material to
be exchanged in order to generate the offspring. For each
pair of genes, a random number Randc ∈ {0, 1} is gener-
ated. If Randc ≤ Pc , the genes are exchanged, otherwise the
pair proceed without crossover. Uniform crossover reserve
the structure of the genome, allowing the genetic material
exchange to occur only between the same gene positions (see Fig. 4 – Uniform crossover and mutation operators.
288 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298

algorithm. One of the criteria is termination when a specified

Table 1 – Unit specifications and operating conditions for
number of generations has been generated and tested. Other batch case study
methods, such as termination based on the best genome, aver-
Property Value
age genome or a combination of the two, can be used. The
effects of two termination criteria, number of generation and Feed composition, xi,feed (mole fraction)
best genome convergence, are discussed in Appendix B. In Acetone 0.70
the case studies presented later, a stopping criterion based on Water 0.30

a maximum number of generations (typically 200 for 15,000 Batch size, mf (mol) 20,000
function evaluations) is used. Product purity specifications, xi,f (mole fraction) ≥0.97
Product recoveries, mmin
i
≥0.70
Condenser holdup (mol) 50.0
5. Batch case study Tray holdup (mol) 1.0
Column operating pressure, P (atm) 1.0
Available production time, TA (h) 7920
In this section, the procedure outlined in the previous sec- Batch setup time, ts (min) 30
tions is now demonstrated by considering separations of
Cost, Ci ($/mol)
acetone–water mixtures in a batch separation system of either
Acetone 0.606a
batch distillation alone, batch pervaporation alone or a hybrid Water 0.0038a
batch distillation/pervaporation. (The acetone–water mixture Feed 0.15
is a tangent pinch mixture. Note that if the mixture consid- Offcut 0.01
ered forms a minimum or maximum boiling azeotrope, then Utilities ($/MJ) 0.019b
traditional distillation alone would obviously not be a feasible K1 1500c
option.) Multiflash (Infochem, 2005) is used to estimate the K2 180c
thermodynamic model states based on the Wilson model. UF 4.31c
The objective for each case is to determine the optimal a
Adopted from Mujtaba (1999).
configuration, design and operation of the most economi- b
Adopted from Sinnott (1993).
cally favourable process for the given separation task. Based c
Adopted from Longsdon et al. (1990).
on the annual profit given by the objective function in Eq.
(1), the results are then verified by comparing the optimal
process with the optimal design and operation of the other Table 2 – Optimisation variable bounds for batch and
two process alternatives. Note that only a single pervapora- continuous case studies
tion membrane separation stage is considered in this work Decision variable Bounds
although multistage processes are possible and may in some
Number of distillation trays, Nt [1, 30]
cases have higher economical potentials. The effects of feed Number of membrane modules, Nm [1, 10]
specifications and batch size on the optimal solution are also Retentate recycle tray, Lr [1, 30]
investigated to determine if these factors have any impact on Time, ti , tf (min) [0.01, 1000]
which process is the optimal. Recycle ratios, Rr , Rc , Rp [0, 1]
Column boilup rate, V (mol/s) [0.5, 5.0]
Permeate pressure, Pp (Pa) [300, 4000]
5.1. Batch base case description Membrane feed rate, Ffeed (mol/s) [0.5, 5.0]

The separation of a 70:30 mol% acetone–water mixture of

20,000 mol (approximately 9000 l) batch size is considered
first. The objective is to determine the most appropriate
process (distillation alone, pervaporation alone or hybrid dis-
tillation/pervaporation) for the separation of the mixture
assuming product constraints on both products. Tables 1 and 2
summarise the unit specifications and operating conditions
and the optimisation decision variable bounds used, respec-
tively. It is assumed that all energy duties Qi carry the same
utility cost index Cut,i although different indices could have
easily been used and their impact on which process is the
optimal could have been investigated.
The batch distillation process operation is split into three
intervals, a total reflux period followed by an acetone prod-
uct collection period and an off-cut period for water product
purification remaining in the reboiler. The optimal solution
will have an appropriate balance between the time intervals
(Table 3). (The membrane characterisation equations reported
by Tsuyumoto et al. (1997) are used although they described
the separation of an ethanol–water mixture. It is considered
a reasonable assumption to use these equations to describe
the separation of the acetone–water mixture as the objective
of the case studies is to demonstrate the applicability of the
optimisation procedure and not to obtain an accurate solution
for this particular mixture.)
The batch pervaporation process operation is separated
into three intervals, a water product collection, followed by
Table 3 – Hollow fibre module specifications (Tsuyumoto
et al., 1997)
Property Value

as well as determine the total batch time which is the sum of Membrane material Polyion complex
these intervals. Support material Polyacrylonitrile
For the batch pervaporation process, the optimisation pro- Module radius (×10−3 m) 97.4
cedure allows any membrane characterisation equation to Fibre inner radius (×10−3 m) 0.251
Fibre outer radius (×10−3 m) 0.395
be used. In this case study, a polyion complex membrane
Membrane thickness (×10−3 m) 0.015
type reported by Tsuyumoto et al. (1997) is assumed. The
Fibre length (m) 1.0
membrane preferentially permeates water and the module Number of fibres 3800
is constructed as a highly porous (75%) hollow-fibre bundle
 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298 289

Table 4 – Optimal solution sets for batch case study

Design variables Operation variables Profit (M$/year)

Hybrid (optimal) ud = {Nt , Nm , Ls , Lr } uo = {t

f , ti , Rc , R Fm } 
r , Rp , Pp , V,    19.024
81.3 1.0 0.16 0.0
ud = {30, 2, 1, 5} uo = 104.5, 15.0 , 0.99 , 0.16 , 0.0 , 300, 4.94, 2.72
8.1 0.99 0.15 0.0
Distillation (fixed) ud = {Nt } uo = {t
f , ti , Rc , V}    17.77
0.02 0.75
ud = {30} uo = 117.8, 117.0 , 0.57 , 4.70
0.06 0.71
Pervaporation (fixed) ud = {Nm } uo = {t
f , ti , Rr , Rp , P p , Fm }
    2.185
797.5 0.99 0.0
ud = {8} uo = 857.5, 25.4 , 0.37 , 0.0 , 300, 4.84
34.6 0.27 0.0

a possible off-cut period and a final acetone product-dumping
period (into the retentate product tank) when acetone reaches
the required product purity.
The hybrid distillation/pervaporation process operation is
also separated into three task intervals, the first period is a
total reflux period similar to the batch distillation process, fol-
lowed by a second interval for water product collection, and
the final interval for acetone product collection.
5.2. Optimal solution of batch base case
The optimum solution set for the base case problem
outlined above is shown in Table 4. A batch hybrid distilla-
tion/pervaporation process is found to be the most profitable
processes alternative that meets all separation requirements
for this case study (see Table 4, top section), with an estimated
profit of $19.024 million per annum.
The optimal superstructure solution, an integrated hybrid
of distillation and pervaporation, indicates the optimal num-
ber of trays Nt to be 30 and membrane modules Nm to be 2.
The optimal number of trays (30) is at the upper bound, how-
ever, the vapour load V is 4.94 mol/s, which is below the upper
bound (5 mol/s). The optimal side-draw location Ls was found
to be from tray 1, i.e. from the top of the column, with a flow
rate Fm of 2.72 mol/s. The optimal retentate recycle stream
location Lr was found to be at tray 5.
The optimal batch time is found to be 104.5 min with inter-
val split of 81.3, 15.0 and 8.1 min, respectively. The interval
splits are, however, insignificant as the optimal column reflux
ratio (Rc ), retentate recycle ratio (Rr ) and the permeate off-cut
ratio (Rp ) remain almost constant (the reflux ratio is slightly
reduced from 1.0 to 0.99 during the last two intervals and the
retentate recycle ratio is only reduced from 0.16 to 0.15 during
the last interval while the permeate off-cut ratio Rp remains
constant at 0 which means no off-cut is required).
The membrane permeate side pressure is found to be at the
lower bound of 300 Pa as expected, as the minimum permeate
side pressure will allow for faster retentate stream purifica-
tion and hence lower overall batch time. It is observed that for
the optimal hybrid process, variations in reflux and retentate
recycle ratios are very small with no variations in the perme-
ate off-cut. This indicates that constant operation settings can
be adopted with minimal losses which makes the process easy
to operate.
The product purity profiles are shown in Fig. 5. It can be
noted that the acetone product purity reaches the required
specification after approximately 30 min, whilst the water
purity in the reboiler continues to increase until all the acetone
product is collected on the top accumulator.
5.3. Comparison with fixed configurations
The optimal configuration for the base case is a hybrid process.
In the following, the optimal hybrid process will be compared
with the best possible batch distillation process and the best
possible pervaporation process (this is done using the same
optimisation procedure as before but with a fixed configura-
tion of either distillation or pervaporation).
For comparison (see Table 4), when the process super-
structure is fixed to optimise a batch distillation column
configuration only, the optimum design results in an esti-
mated annual profit of $17.77 million which is 7% less
than that of the optimal hybrid solution. When the process
superstructure is fixed to optimise a batch pervaporation con-
figuration only, the optimal design and operation results in the
least profitable alternative with an estimated annual profit of
only $2.185 million (88% less). This is significantly lower than
for the batch hybrid and batch distillation processes because of
the long batch time required (857.5 min) for the pervaporation
process due to the process scale, as pervaporation membrane
units are normally better suited to small scale processes with
the module and costing structures used in this case study.
5.3.1. Batch distillation process
Now consider the designs in detail. In comparison, the optimal
number of trays and the reboiler vapour load for a fixed batch
distillation configuration are found to be 30 and 4.70 mol/s,
respectively, i.e. the optimal number of trays is the same as
that of the optimal configuration. The vapour load is, how-
ever, lower than that of the optimal configuration, indicating
a slightly smaller column diameter is preferred which results
Fig. 5 – Product compositions in optimal hybrid
(70:30 mol%, 20,000 mol feed).
290 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298

in lower capital and operation costs. The batch time is found

to be longer at 117.8 min with interval splits of 0.02, 117.0 and
0.06 min and corresponding constant reflux ratio values for
these intervals of 0.75, 0.57 and 0.71, respectively. Note that the
optimal total reflux period is found to be almost nonexistent
(0.02 min) which is reasonable as the column was assumed to
be at total reflux conditions initially. The final time is also very
short (0.06 min) and in practice, the column would be operated
with one time interval only (117.8 min) instead of three.

5.3.2. Batch pervaporation process

The batch pervaporation process has an optimal number of
eight membrane modules. The batch time is significantly
longer than that of the batch distillation and hybrid pro-
cesses with 857.5 min and interval splits of 797.5, 25.4 and
34.6 min. The optimal retentate recycle ratio remained at 0.99
for the first interval (797.5 min) for the purification of the ace-
tone product to the required purity. The recycled ratio is then
reduced to 0.37 and 0.27 for the subsequent two intervals,
respectively, to allowing for more product collection. The per-
meate pressure is found to be 300 Pa and the feed flow rate is
4.84 mol/s. As expected, the optimal permeate pressure is at its
lower bound allowing for the maximum possible flux through
the membrane, and thus purification of the acetone retentate
stream, in the shortest possible time.

5.3.3. Hybrid process

Comparing the optimal superstructure solution for the hybrid
base case (H,0.7 in Fig. 6 which shows the capital and oper-
ating costs for different cases) with the optimal distillation
and pervaporation solutions (D,0.7 and P,0.7, respectively), it
was found that the capital cost of the hybrid process (H,0.7)
is the highest with the majority of the cost contributed by
the distillation column. The pervaporation capital cost is the
second highest and the distillation process capital cost is the
lowest. This is in-line with the general assumption that mem-
brane processes are more costly than distillation processes for
large-scale production (the feed was 20,000 mol/batch).
Fig. 6 – Capital costs (a) and operating costs (b) as a
The operating cost of the hybrid process (H,0.7) is also
function of acetone feed composition.
the highest of the three processes with the major contribu-
tion from the distillation column accounting for almost two
thirds of the total operating costs. The pervaporation process
resulted in the lowest operating costs, almost half that of the assumed to be of equal cost (0.15 $/mol) and all other spec-
distillation process which had the second highest operating ifications and constraints are the same as in the base case
costs. (Tables 1 and 2).
It should be noted that high performance equipment which The results in Table 6 and Fig. 7 show the merit of the hybrid
is operating for high product recovery will cost more in terms batch distillation/pervaporation process over the stand-alone
of capital and operating costs, however, will achieve more processes. The hybrid distillation system was in all cases
product revenue and result in shorter batch times. This trade- found to be more profitable than the other two and perva-
off is considered in the proposed procedure and for this case poration was found to be the least profitable. This is mainly
study, the optimal superstructure solution is biased toward because pervaporation processes tend to have significantly
higher product recoveries and shorter batch times instead of
lower capital and operating costs given the costing parameters
in Table 1.
Table 5 – Cases considered for batch case study
5.4. Effect of feed composition Case Feed size (mol) Feed composition (mol%)
(acetone–water)
The results presented for the base case was for a 70–30 mol%
A 20,000 [0.5, 0.5]
mixture with a 20,000 mol feed. In this section, the effect
B 20,000 [0.7, 0.3]
of the feed composition on the optimal synthesis is con- Ca 20,000 [0.9, 0.1]
sidered in order to investigate its impact on the optimal D 2,000 [0.9, 0.1]
solution. The comparison studies are based on 20,000 mol E 200 [0.9, 0.1]
feed of acetone–water mixture of various feed specifications a
Base case.
as shown in Table 5 (cases A–C). All the feed mixtures are
 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298 291

Table 6 – Optimal solution sets for cases A–C for batch case study
Case Distillation Pervaporation Hybrid

A B C A B C A B C

Feed composition (mol%) (acetone–water) [0.5, 0.5] [0.7, 0.3] [0.9, 0.1] [0.5, 0.5] [0.7, 0.3] [0.9, 0.1] [0.5, 0.5] [0.7, 0.3] [0.9, 0.1]
Number of trays, Nt 30 30 30 – – – 28 30 30
Number of modules, Nm – – – 9 8 10 1 2 3
Batch time, tf (min) 97.83 117.80 164.45 887.17 857.5 410.5 70.37 104.53 134.37
Permeate pressure, P (Pa) – – – 300 300 300 340 300 300
Vapour load, V (mol/s) 4.46 4.70 4.70 – – – 4.99 4.94 4.94
Stage feed Fm (mol/s) – – – 4.94 4.84 5.0 2.40 2.72 2.97
Capital cost ($/year) 645,352 664,500 664,500 865,102 778,385 657,502 803,416 867,158 939,554
Operating cost ($/year) 120,540 126,275 137,976 55,542 52,236 64,074 182,534 191,419 203,207
Revenue (M$/year) 11.831 18.578 19.970 1.667 3.015 8.880 15.041 20.306 23.306
Annual profit (M$/year) 11.065 17.770 19.168 0.747 2.185 8.158 14.055 19.024 22.163

longer batch time at this scale (20,000 mol). Distillation pro-
cesses are expected to be economically favourable at lower
acetone feed composition, as membrane processes are gen-
erally inefficient at high water content feed streams with the
membranes used in this study.
In order to consider the optimal number of distillation col-
umn trays, the upper bound for Nc was extended to 50 for case
A to investigate the effect of this on the optimal solution. The
optimal superstructure solution was again found to be a hybrid
process with the optimal number of trays and the reboiler
vapour load found to be 32 and 4.93 mol/s, respectively, com-
pared to 30 and 4.94 mol/s with the optimal solutions obtained
with upper Nc bound of 30. This configuration was, however,
found to result in a sub-optimal solution with lower profit
value ($15.51 million per annum) than the solution obtained
with upper Nc bound of 30 ($19.024 million per annum). This
is due to the fact that with higher bound for Nc , the search
solution space is further extended and hence it is made more
difficult (and time consuming) to obtain the global solution
and this was not taken into account in the solution and the
procedure may have terminated prematurely as the maximum
number of generations was fixed to 200.
In the hybrid configuration it is observed that the side-draw
flow rate increases as the hybrid becomes more dependant
on the membrane stage (2.40 for case A, 2.72 for case B and
2.97 mol/s for case C), this is also apparent as more modules
are used as the water content in the feed is reduced (1 for case
A, 2 for case B and 3 for case C).
The batch processing time for the hybrid distillation sys-
tem is generally less than those of the batch distillation and
pervaporation processes, allowing for more batches to be pro-
cessed per year in a 24-h production plant. The pervaporation
processes have significantly longer batch time, resulting in a
lower number of batches that can be processed in the available
annual production time. This resulted in lower revenues and
therefore a lower overall profitability. However, pervaporation
processes are expected to out-perform distillation and hybrid
processes at low process scales as will indeed be discussed in
the next section.
The annual capital costs (see Fig. 6) seem to have an
increasing trend in batch distillation and hybrid configura-
tions as the acetone feed concentration increases. However,
for the pervaporation processes, the annual capital cost shows
a decrease with an increase in acetone feed purity, which
is mainly because pervaporation membrane processes have
higher efficiency for lower water content in the feed.
5.5. Effect of batch size
The effect of batch size is investigated at a feed concentration
of 90 mol% acetone, i.e. for mixtures with low water content
which should be preferable for the pervaporation processes.
Three different batch sizes of 20,000, 2000 and 200 mol are
explored (cases C–E in Table 5) and the results are given in
Table 7 and Fig. 8 (note the semi-log scale). The column liquid
holdup is kept constant in each case to allow for an easy com-
parison, however, in reality, this would be a function of batch
and column size.

Fig. 7 – Profit vs. feed concentration for all processes Fig. 8 – Profit vs. scale for different batch size (90:10 mol%
(20,000 feed). feed).
292 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298

Table 7 – Optimal solution sets for cases C–E for batch case study
Case Distillation Pervaporation Hybrid

C D E C D E C D E

Feed size (mol) 20,000 2,000 200 20,000 2,000 200 20,000 2,000 200
Number of trays, Nt 30 28 23 – – – 30 30 19
Number of modules, Nm – – – 10 4 1 3 1 2
Batch time, tf (min) 164.45 16.73 4.67 410.5 58.17 11.50 134.37 16.00 5.85
Permeate pressure, P (Pa) – – – 300 300 300 300 370 710
Vapour load, V (mol/s) 4.70 4.90 5.0 – – – 4.94 4.72 1.66
Stage feed, Fm (mol/s) – – – 5.0 3.96 3.33 2.97 2.31 2.22
Capital cost ($/year) 664,500 657,769 476,118 657,502 325,939 180,853 939,554 736,045 323,751
Operating cost ($/year) 137,976 60,151 17,545 64,074 30,761 11,841 203,207 67,109 10,750
Revenue (M$/year) 19.970 8.303 1.134 8.880 4.315 0.924 23.306 8.430 1.049
Annual profit (M$/year) 19.168 7.585 0.641 8.158 3.958 0.731 22.163 7.627 0.714

With a batch size of 200 mol, the optimal superstructure same utility cost index Cut,i although different indices could
is a pervaporation process with a profit 12% higher than the have easily been used.
batch distillation process and 2% higher than the hybrid pro- The optimum solution sets of the superstructure and that
cess. At this batch size, a distillation configuration results in of a comparative distillation case are shown in Table 9. A fully
the least profitable operation although the differences are not integrated hybrid distillation/pervaporation process is again
large. This conclusion does not hold, however, at batch sizes found to be the most profitable process alternative that meets
of 2000 and 20,000 mol where the hybrid process is found to all separation requirements for this case study with an esti-
be the optimal process, with the pervaporation process the
least profitable option and the distillation process second best.
It is therefore concluded that pervaporation processes are
more suited to small-scale separations as expected. A cap-
ital and operating cost comparison (see Fig. 9) indicates a
consistent decrease in costs as batch size decreases, also as
expected.
It is interesting to note that a hybrid process always seem to
perform better than a pure batch distillation process for these
cases (cases C–E). This is because the hybrid process resulted
in lower batch times than distillation in cases C and D, hence
more batches can be performed in the annual available time
TA and therefore higher annual revenue, and consequently
higher profit, can be achieved. For case E, the hybrid process
has resulted in lower energy consumption, hence in lower
operating costs, and therefore higher overall profitability than
batch distillation alone.

6. Continuous case study

The optimal process synthesis procedure developed in this

work is next demonstrated by considering the continuous sep-
aration of an equimolar tangent-pinch mixture of acetone
and water. The separation process specifications are shown
in Table 8. It is assumed that all energy duties Qi carry the

Table 8 – Unit specifications and operating conditions for

continuous case study
Property Value

Feed composition xi,feed (mole fraction)

Acetone 0.50
Water 0.50

Feed rate, Ffeed (mol/h) 18,000

Product purity specifications, xi,f (mol fraction) ≥0.97
Product recoveries, mi,f ≥0.70
Tray/condenser/reboiler holdup (mol) 1/100/500
Column operating pressure, P (atm) 1.0
Fig. 9 – Capital costs (a) and operating costs (b) as a
Available production time, TA (h) 7920
function of batch size.
 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298
Table 9 – Optimal solution sets for continuous case study
Optimal
Design variables ud = {Nt , Nm , Lf , Ls , Lr } = {18, 1, 16, 1, 5}
Operation variables uo = {Rc , Rr , Pp , V, Fm } = {0.92, 0.23, 370, 4.32, 2.72}
Revenue (M$/year) 23.55
Capital cost ($/year) 503,000
Operating costs ($/year) 195,000
Profit (M$/year) 22.85
mated profit of 22.85 M$ per annum. The optimum hybrid
process is found to be of a comparable profit to the optimal
distillation process and only a 1% increase in profitability can
be achieved with an optimised hybrid column compared to
distillation alone. It is noted, however, that the capital cost of
the hybrid column (distillation and membrane) is considerably
lower (14%) than that of the conventional distillation process.
Similar savings, although not as large, in operating costs can
be achieved where the hybrid process shows a 4% decrease
compared to the optimal distillation process.
The optimal number of trays and membrane modules are
found to be 18 and 1, respectively. The optimal feed stream
location is found to be at tray 16 as expected for a tangent-
pinch mixture where the separation is more difficult at the
top of the column, and the feed is therefore directed towards
the bottom of the column. The optimal side-draw location in
the hybrid column is found to be at tray 1 (i.e. at the top of the
column), with the optimal retentate return location to tray 5.
Optimal reflux and recycle ratios are found to be 0.92 and
0.23, respectively. Optimal reboiler vapour load is found to be
4.32 mol/s with optimal side-draw flow rate of 2.72 mol/s. The
optimal permeate side pressure is 370 Pa.
Note that it is beyond the scope of this work to compare
batch versus continuous operation, hence this is not discussed
here.
7. Conclusions
In this work, an optimisation procedure based on genetic
algorithm for the optimal synthesis of batch and continu-
ous separation processes has been proposed. The synthesis
problem is solved through simultaneous consideration of opti-
mal configuration, design and corresponding operation of all
process alternatives through a process superstructure. The
superstructure encompasses three different process alter-
natives, distillation, pervaporation and a hybrid of the two
although other process alternatives can also be integrated
into the superstructure. The complexity will, however, greatly
increase and accurate solutions may become difficult to obtain
with current computer capabilities. The proposed objective
function reflects the various trade-offs between design and
operation decision variables versus production revenue as
well as that of capital investments versus operating costs.
The separation of a mixture of acetone and water was
used to demonstrate the procedure. The hybrid batch distilla-
tion/pervaporation configuration was found to be the optimal
synthesis solution for most process scale separations, and
distillation processes were second best. Pervaporation pro-
cesses were found to have significantly higher batch time
which generally resulted in the least overall profitability. Nev-
ertheless, pervaporation processes were found to be most
suitable at small scale, out-performing both distillation and
hybrid processes at this scale. It is therefore concluded that
batch pervaporation processes are most suited for small-scale
293
Hybrid Distillation
ud = {Nt , Lf } = {23, 22}
uo = {Rc , V} = {0.58, 4.50}
23.42
572,000
203,000
22.64
operations. For large-scale separations, optimal hybrid con-
figurations were found to be the most suitable for the case
studies considered here.
The proposed synthesis procedure can also be applied
to continuous separation processes. From the case study,
a hybrid distillation/pervaporation configuration was again
found to be the optimal synthesis solution for the separation
of an equimolar acetone–water mixture. Although the hybrid
process was found to result in significant capital and operat-
ing cost savings, it was only marginally more profitable than
the distillation alone process.
An unlimited feed availability and product demand has
been assumed in this work. A supply and demand model
can be integrated in the objective function where adequate
information is available. This will reflect a more accurate rep-
resentation of the number of batches that can be performed
per year as well as the cost incurred and the revenue achieved
from these batches. This may ultimately have an impact on
the optimal process found by the synthesis procedure. The
same applies to the specific costing data used.
Appendix A. Derivation of cost correlations
This appendix details the costing data for the distillation, per-
vaporation and the hybrid system as used in this work. The
distillation column cost correlations are first outlined followed
by the derivations of the pervaporation and the hybrid unit
costs.
A.1. Distillation column
A.1.1. Capital costs
The annualised capital costs associated with the installed
equipment costs for the column shell can be described using
Guthrie’s correlations (Douglas, 1988) as follows:
 N 0.802  D 1.066
Csh = Csh,RC (A.1)
NRC DRC
where N is the number of trays, D the diameter of the column
and RC represents the reference case column on which the
Guthrie’s correlation is based. Assuming the column diameter
varies as the square root of the column vapour loading, D ∝ V0.5
(Douglas, 1988), the equation can be written as
 N 0.802  V 0.533
Csh = Csh,RC (A.2)
NRC VRC
Another contribution to the column cost is the reboiler and
condenser installation cost. Guthrie (Douglas, 1988) proposed
that the annual installed cost of heat exchangers can be
294 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298

described by
Table 10 – Membrane reference case capital cost
 A 0.65 estimates (Sulzer Chemtech, 2005)
Cex = Cex,RC (A.3) Property Value
ARC
Reference membrane area, Am,RC (m2 ) 6
where the heat exchanger area A is calculated from Reference feed flowrate, Fm,RC (mol/s) 0.73
Reference permeate flowrate, Fp,RC (mol/s) 0.0365
Q = Vcp T = UA Tmean Purchase cost of reference module ($) ≈350,000
Equipment service cost ($/year) ≈20,000
where T is the temperature difference of the exchanger fluid Capital recovery factor, CRFm ≈1/3
(reboiler or condenser) and Tmean is the mean temperature
difference across the heat exchanger. Assuming constant heat
capacity cp and overall heat transfer coefficient U and substi- such that
tuting for area A gives
 0.3
 V T Tmean,RC
0.65 ACCm = ACCm,RC
Am Fm Fp
(A.12)
Cex = Cex,RC (A.4) Am,RC Fm,RC Fp,RC
VRC TRC Tmean

If the stream temperatures are assumed fixed, a simple model The membrane stage flowrate Fm , membrane area Am and
for the heat exchangers costs in terms of flows can then be permeate flow rate Fp are the main parameters affecting mem-
obtained: brane system performance and are hence included in the cost
equation for comparison with a reference case with costs
 V 0.65
ACCm,RC . The scale exponent is estimated to be 0.3, the low-
Cex = Cex,RC (A.5)
VRC est recommended value (IChemE, 1988), in order to accurately
predict the pervaporation capital costs.
The equations for the column shell costs and the heat Based on the reference case specifications given in Table 10
exchanger costs can now be written respectively as: (Sulzer Chemtech, 2005), the annualised capital cost of the
reference case ACCm,RC can be calculated as
Csh = K1 N0.802 V 0.533 (A.6)
1
Cex = K2 V 0.65 (A.7) ACCm,RC ≈ $350, 000 × + $20, 000 ≈ $136, 000 (A.13)
3

The values of the correlation coefficients K1 and K2 can be A.2.2. Operating costs
calculated from the reference case column: The feed stream heater and the permeate side cooler, in addi-
tion to the membrane feed pump and the permeate vacuum
Csh,RC turbine, are considered to be the main contributions to the
K1 = 0.802 0.533
(A.8)
NRC VRC operating cost of the membrane unit. The operating costs for
the membrane process are therefore given by
Cex,RC
K2 = 0.65
(A.9)
VRC
AOCm = (Cut,m,h Qm,h + Cut,m,cond Qm,cond + Cut,m,p Qm,p

The total annualised capital costs ACCc for the column are TA
+ Cut,m,t Qm,t ) (A.14)
then given by ts + tf

ACCc = K1 Nt0.802 V 0.533 + K2 V 0.65 (A.10) where

p out
A.1.2. Operating costs Qm,p = (P − Pin
p ) (A.15)
p p
The main contributions to operating costs in a distillation col-
umn are the reboiler heating and condenser cooling duties.
The annual operating costs AOCc for the distillation column and
is therefore given by  
t in Pout
t
Qm,t = P ln (A.16)
AOCc = Cut,reb Qreb + Cut,cond Qc,cond (A.11) t t Pin
t

where Cut,reb and Cut,cond represent the unit cost of utility
in the reboiler and condenser, respectively. Qreb and Qc,cond
represent the reboiler heating and condenser cooling duties,
respectively.
A.2. Pervaporation costs
A.2.1. Capital costs
The annualised capital cost for the pervaporation membrane
process can also be correlated using a Guthrie-type correlation
from values of a reference case including pumps and heaters
where Cut,i represent cost of utilities of equipment i (i = pump,
turbine), Qm,h and Qm,cond represent the pervaporation unit
total feed tank heat duty and permeate condenser cooling
duty. Qm,p and Qm,t are the total energy load of the membrane
unit feed pump and permeate vacuum turbine, respectively.
TA represents the total available time for operation per annum
and ts and tf represent the process startup and operating
times, respectively. i is the flow rate entering the ancillary
equipment i and i is the efficiency of equipment i. Pin
i
and Pout
i
represent the pressure at the inlet and outlet of the relevant
equipment i, respectively.
 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298 295

A.3. Hybrid process costs

Table 11 – GA parameters used in sensitivity study
Parameter Value
The annualised capital costs ACChyb and the operating costs
AOChyb for the hybrid system are the summation of the cost Population size, Npop 100 (50, 100, 120)b
a

contributions of the constituent processes and are therefore
given by
ACChyb = ACCc + ACCm
AOChyb = AOCc + AOCm
Appendix B. Genetic algorithm parameters
tuning
The sensitivity of the GA parameters used in this work is anal-
ysed in this appendix. The base case process for the batch
hybrid separation of acetone–water mixture described earlier
is used to conduct the analysis (Tables 1 and 2).
The main parameters within the GA framework are the
population size Npop , the population overlap percentage Pss ,
the crossover rate Pc , and the mutation rate Pm . Consider-
able theoretical studies have been carried out in order to
link the population size Npop with the number of variables
of the optimisation problem and to better tune mutation and
crossover rates. The tuning operation, because of the stochas-
tic nature of the algorithm, nevertheless should be performed
for each specific problem at hand, especially where the genetic
algorithm approach has not been applied before (e.g. GA opti-
misation of hybrid separation processes as considered in this
work) as the choice of the values of these parameters may
influence the outcome of the optimisation.
Another important determining factor in obtaining a good
final solution is the termination criterion employed and, in
the case of constrained optimisation problems, the penalty
method used. This appendix outlines the effects of the GA
population size Npop , mutation percentage Pm , penalty func-
tion and termination criteria employed on the algorithm’s
performance as employed in this work. (The values of the
other GA parameters are within typical ranges previously
reported which have been shown to be effective in most
genetic algorithm applications on a wide range of problems
including those outside of the chemical engineering field.)
The general procedure in tuning GA parameters is first to
obtain good initial guesses for their values based on general
heuristics and rules of thumb (Goldberg, 1989) or from past
experience of similar chemical engineering optimisation prob-
lems where GA has been applied (e.g. Marriott and Sørensen
(2003b) for membrane processes). For instance, most of the
applications in various fields, including outside chemical engi-
neering, points to high crossover rates and low mutation
rates, e.g. above 70% and below 20%, respectively. This is then
followed by fine-tuning of the parameters through either a
sensitivity analysis as presented in this study, or paramet-
ric optimisation although this adds further complexity to the
problem. The analysis conducted in the following uses as base
case parameters those given in Table 11.
B.1. Population size Npop
The size of the population Npop should be sufficiently large to
provide diversity within the GA populations. Lewin et al. (1998)
report that genetic algorithm are generally insensitive to the
population size provided that this is not very small. Garrard
Crossover rate, Pc 75%a
Mutation rate, Pm 0%a (0%, 5%, 10%, 15%, 20%)b
Population overlap percentage, Pss 25%a
Termination criteria Population besta ≥ 95%
(A.17)
(population best ≥ 95%, number
of generations ≥ 200)b
(A.18)
a
Base case.
b
Sensitivity.
and Fraga (1998) studied the effect of population size and per-
centage overlap for a range Npop ∈ [10, 100] and Pss ∈ [25%, 75%],
respectively. They concluded that higher population size and
percentage overlap result in a better standard deviation of the
population from the optimum solution value, but with higher
computational costs. The value of percentage overlap Pss is
thus a trade-off between reliability of the final solution and
computational costs as long as the critical population size is
established.
Table 12 shows the effect of varying the population size
Npop on the average and best solutions with all the other
parameters at their base case values (Table 11). It can be seen
that a population size of 120 yields the best average and best
solutions compared to the other sizes. (Note that a popula-
tion size of 150 resulted in an infeasible final solution. This
is mainly because, when using a roulette wheel selection on
large populations, there is reduced chances that good solu-
tions will be selected as there are more to choose from.)
With a small population size (50) for the optimisation, the
final solution is infeasible as there is not enough diversity in
the genetic population to evolve effectively. If a population size
of 100 is considered, a 21% decrease on the best solution value
is observed compared to 120 generations. The improvement
of using the extra 20 genomes in the population, however,
increases the computational time, as the population size Npop
of 120 converged in 68 generations compared to 59 genera-
tions with Npop = 100. The decision of which size to use will
depend on the required accuracy and the available computa-
tional time which is essentially decided by the purpose of the
optimisation procedure (e.g. rigorous or initial screening). As
the work presented in this paper considers initial screening of
configuration alternatives, a population size of 100 was used
in the case studies.
B.2. Mutation rate Pm
The mutation rate Pm is the likelihood that a gene in an off-
spring genome will change and it is used to control the rate of
introducing new genetic material into the population. A high
Table 12 – Effects of population size on GA performance
Population size, Npop
50 100 120
Final best solution (×106 $/year) 0.92a 11.8 15.0
Final average solution (×106 $/year) 0.85a 2.77 12.1
Constraint violation in final best (%) 8.5a 0 0
Generations to convergence 48a 59 68
a
Infeasible solution.
296 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298

Table 13 – Effects of mutation rate on GA performance

Mutation rate Pm (%)

0 5 10 15 20
6 a b b
Final best solution (×10 $/year ) 0.98 0.99 11.8 7.25 11.7
Final average solution (×106 $/yeara ) 0.98b 0.98b 2.77 0.73 0.30
Constraint violation in final best (%) 2.3b 0.36b 0 0 0
Generations to convergence 86b 45b 59 59 93

a
Infeasible solution unit is $.
b
Infeasible solution.

mutation rate will override the effect of genome crossover, but where
a value set too low will cause the algorithm to converge pre- ⎧ 
maturely. It is therefore important that an appropriate value ⎪
⎪ cmin − ci (tf )
⎨ 1 − i min if ci (tf ) < cimin
for the mutation rate be chosen. ci
i = (B.2)
The mutation rate Pm , as all genetic operators, is depen- ⎪
⎪ where c = [xi , mi ], ∀ i = 1, . . . , Nc
⎩
dent on the type of problem being solved. Garrard and Fraga 1 otherwise
(1998) and Low and Sørensen (2003) found that a mutation
rate of 10% resulted in better solutions than other values for
Soft constraints handling:
mass exchanger network synthesis and optimisation of batch
distillation processes, respectively. A mutation rate of 20%, 
Nc 
˝˘i=1 when ˝ ≥ 0 (profit)
i
however, has been found to be better for the optimisation f = Nc  )
(B.3)
˝(2 − ˘i=1 i when ˝ < 0 (loss)
of pervaporation membrane systems (Marriott and Sørensen,
2003b). It is therefore expected that a value between these
where
two will be appropriate for the hybrid processes considered
⎧ p
here. ⎪ cmin − ci (tf ) i
Table 13 shows the effect of different mutation rates on
⎪
⎨ 1 − i min if ci (tf ) < cimin
ci
the average and best solutions, respectively, with all other i = (B.4)
⎪
⎪ where c = [xi , mi ], ∀i = 1, . . . , Nc
parameters fixed as in the base case. It is observed that with ⎩
1 otherwise
low mutation rates (0% and 5%), the final population con-
verged to infeasible solutions (using as termination condition
The difference between the two methods is that the hard
a convergence setting of ≥95% of the best solution com-
constraints handling method assigns very low fitness values
pared to the best solution within the previous 30 generations).
to infeasible solutions (a value between 0 and 1) relative to
The largest constraints violation was found to be 2.3% for
the constraint violations whereas the soft method multiplies
the case with a 0% mutation. Lower violations (0.36%) and
the objective value (i.e. profit) by the constraint violations
hence solution closer to feasibility, resulted for the case of
(a value between 0 and 1) and thus reduces the objective
5% mutation rate. Mutation rate of 10% was found to yield
value proportional to the violation. The problem with this
the best solution in terms of the final solution’s best and
method is that if a feasible solution can achieve, for instance,
average values. A mutation rate of 15% was found to result
$1 × 106 profit, an infeasible solution with a 9% violation but
in a lower average and a lower best solution compared to a
an uncorrected profit figure of $10 × 106 will be preferred by
10% mutation rate but with convergence at the same pop-
the optimisation procedure as the corrected profit will then
ulation (after 59 generations). A similar final solution was
be 10 × 106 × (100 − 9%) = $9.1 × 106 . It should be noted that the
found, although a slightly lower best solution, using a 20%
objective value in this work is proportional to the amount of
mutation rate but at a significant increase in the required
product collected and capital and operating costs incurred,
population for convergence compared to 10% mutation rate
regardless of the final product purities and recoveries, as these
(93 generations).
two are usually formulated as constraints to the optimisation
problem.
The effects of using either constraint handling method
B.3. Penalty function are now considered. As shown in Table 14, the absolute

A penalty function is applied to infeasible solutions to drive

the algorithm to feasibility and subsequently to optimality. A
method of hard (absolute) constraints’ satisfaction is proposed Table 14 – Effects of constraint handling method on GA
performance
in this work. This method is a variant of the soft constraints
handling equation presented by Low and Sørensen (2003). Constraint handling
Absolute constraints handling: Absolute Soft
6
Final best solution (×10 $/year) 11.8 19.4a
⎧ Final average solution (×106 $/year) 2.77 17.1a
⎨˝ when feasible Constraint violation in final best (%) 0 5.2a
f = Nc 
˘i=1 when ˝ ≥ 0 (profit but infeasible) Generations to convergence 59 101a
i
⎩ ˝(2 Nc  )
− ˘i=1 when ˝ < 0 (loss and infeasible)
i a
Infeasible solution.
(B.1)
 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298
Table 15 – Effects of termination criteria used on GA
performance
Termination criteria
Population
best
Final best solution (×106 $/year) 11.8
Final average solution (×106 $/year) 2.77
Constraint violation in final best (%) 0 0
Generations to convergence 59
a
Fixed.
constraint handling method results in better constraint satis-
faction and convergence time (i.e. populations required). The
soft constraints handling method resulted in an infeasible
final solution and a small violation was accepted for lower
operating costs and thus higher profit value (without feasibil-
ity).
The hard constraint handling method proposed here will
also eliminate the need to specify the value of the param-
eter pi in Eq. (B.4). (The comparison reported in Table 14
is based on a parameter value pi = 8 as used by Low and
Sørensen (2003) for the optimisation of batch distillation
processes).
B.4. Termination criteria
The accuracy of the final solution of the GA procedure depends
on the termination criterion employed. Common convergence
criteria used are genome, population or generation based.
Genome and population based criteria have similar charac-
teristics, with the former depending on the similarity of the
best solution from one generation to another and the latter
depending on the similarity of the genomes within the pop-
ulation. The population convergence typically takes longer to
converge than does the genome convergence. Another com-
mon technique is to terminate the GA when a particular
number of generations or/and objective function evaluations
has been exceeded. The effects of these termination criteria
are now investigated with all other parameters as in the base
case (Table 11).
Table 15 shows the results for the genome based (popu-
lation best) and maximum number of generations methods
applied to the optimisation of hybrid processes. It is noted
that both methods result in solutions that satisfy all the con-
straints. The maximum number of generation convergence is
based on a maximum of 200 generations. The genome con-
vergence criterion is based on that the best solutions should
have ≥95% resemblance compared to the best solution during
the previous 30 generations. The maximum number of gener-
ations method resulted in a better solution (61% higher) than
the genome convergence criterion. Which method to use will
depend on the required accuracy and the available computa-
tional resources.
B.5. Summary
Based on the sensitivity studies presented here, general guide-
lines are now available for appropriate parameter values for
the optimisation of hybrid separation systems using GA. Note
that the sensitivity trials presented were based on a single
hybrid distillation/pervaporation problem (base case study),
hence further parameter value adjustment may be needed
297
should the size or complexity of the optimisation problem
vary.
Maximum
references
generations
19.02
Coley, D., (1999). An Introduction to Genetic Algorithm for Scientists
15.2
and Engineers (1st ed.). (World Scientific Publishing).
Douglas, J., (1988). Conceptual Design of Chemical Processes.
200a
(McGraw-Hill Book Company Inc, New York).
Eliceche, A., Daviou, M., Hoch, P. and Uribe, I., 2002,
Optimisation of azeotropic distillation columns combined
with pervaporation membranes. Comput Chem Eng, 26:
563–573.
Furlonge, H., Pantelides, C. and Sorensen, E., 1999, Optimal
operation of multivessel batch distillation columns. AIChE J,
45: 781–801.
Garrard, A. and Fraga, E., 1998, Mass exchange network synthesis
using genetic algorithms. Comput Chem Eng, 22(12): 1873.
Goldberg, D., (1989). Genetic Algorithms in Search, Optimization and
Machine Learning. (Addison-Wesley).
Guerreri, G., 1992, Membrane alcohol separation
process-integrated pervaporation and fractional distillation.
Trans IChemE (Part A), 70: 501–508.
Hasebe, S., Kurooka, T. and Hashimoto, I., 1995, Comparison of
the separation performances of a multi-effect batch
distillation system and a continuous distillation system, In
Proceeding of the DYCORD’95 (p. 249).
Hilmen, E., Skogestad, S., Doherty, M. and Malone, M., 1997,
Integrated design, operation and control of batch extractive
distillation with a middle vessel, In Proceedings of the AIChE
Annual Meeting 1997
Holland, J., 1962, Outline for a logical theory of adaptive systems.
J ACM, 4: 297.
IChemE., (1988). A Guide to Capital Cost Estimating (3rd ed.).
(IChemE).
Infochem, 2005, Multiflash User Manual, Infochem Computer
Services Ltd.
Kookos, I., 2003, Optimal design of membrane/distillation column
hybrid processes. Ind Eng Chem Res, 42: 1731–1738.
Kreis, P. and Gorak, A., 2006, Process analysis of hybrid separation
processes: combination of distillation and pervaporation.
Chem Eng Res Des, 84(A7): 595–600.
Lewin, D., Wang, H. and Shalev, O., 1998, A generalized method
for HEN synthesis using stochastic optimisation. I. General
framework and MER optimal synthesis. Comput Chem Eng,
22(10): 1503.
Lipnizki, F., Field, R. and Ten, P., 1999, Pervaporation-based hybrid
process: a review of process design, applications and
economics. J Membr Sci, 153: 183–210.
Longsdon, J., Diwekar, U. and Biegler, L., 1990, On the
simultaneous optimal design and operation of batch
distillation. Trans Inst Chem Eng, 68(A): 434.
Low, K.H. and Sørensen, E., 2002, Optimal operation of extractive
distillation in different batch configurations. AIChE J, 48(5):
1034–1050.
Low, K.H. and Sørensen, E., 2003, Simultaneous optimal design
and operation of multivessel batch distillation. AIChE J, 49(10):
2564–2576.
Low, K.H. and Sørensen, E., 2004, Simultaneous optimal design
and operation of multipurpose batch distillation columns.
Chem Eng Proc, 43: 273–289.
Marriott, J.I. and Sørensen, E., 2003a, A general approach to
modelling membrane modules. Chem Eng Sci, 58(22):
4975–4990.
Marriott, J.I. and Sørensen, E., 2003b, The optimal design of
membrane systems. Chem Eng Sci, 58(22): 4991–5004.
Mujtaba, I.M., 1999, Optimization of batch extractive distillation
processes for separating close boiling and azeotropic
mixtures. Trans Inst Chem Eng, 77: 588.
Perry, R. and Green, D., (1984). Perry’s Chemical Engineers’ Handbook.
(McGraw-Hill).
298 c h e m i c a l e n g i n e e r i n g r e s e a r c h a n d d e s i g n 8 6 ( 2 0 0 8 ) 279–298
PSE., (2005). gPROMS User Guide. (Process Systems Enterprise Ltd).
Sadomoto, H. and Miyahara, K., 1983, Calculation procedure for
multicomponent batch distillation. Int Chem Eng, 23: 56.
Sinnott, R.K., (1993). Coulson and Richardson’s Chemical Engineering
(Pergamon Press, Oxford).
Sulzer Chemtech, 2005, Personal communication with Sulzer
Chemtech Ltd., Switzerland.
Syswerda, G., 1989, Uniform crossover in genetic algorithms, In
Proceedings of the 3rd International Conference on Genetic
Algorithms
Szitkai, Z., Lelkes, Z., Rev, E. and Fonyo, Z., 2002, Optimization of
hybrid ethanol dehydration systems. Chem Eng Process, 41:
613–646.
Tsuyumoto, M., Teramoto, A. and Meares, P., 1997, Dehydration of
ethanol on a pilot plant scale, using a new type of hollow-fibre
membrane. J Membr Sci, 13(1): 83–94.
Van Hoof, V., Van der Abeele, L., Buekenhoudt, A., Dotremont, C.
and Leysen, R., 2004, Economic comparison between
azeotropic distillation and different hybrid systems
combining distillation with pervaporation for the dehydration
of isopropanol. Sep Purif Technol, 37(1): 33–49.
Viswanathan, J. and Grossmann, I., 1993, An alternate MINLP
model for finding the number of trays required for a specified
separation objective. Comput Chem Eng, 17(9): 949–955.
Wall, M., 1999, GAlib: C++ Library of Genetic Algorithm
components, Version 2.4.5, http://lancet.mit.edu./ga.