Ind. Eng. Chem. Res.

2000, 39, 1423-1429 1423

Design of Solvents for Extractive Distillation

Braam van Dyk and Izak Nieuwoudt*
Institute for Thermal Separation Technology, University of Stellenbosch, Banhoek Road,
Stellenbosch 7600, South Africa

A method is proposed for the computer-aided molecular design of solvents for extractive
distillation. The method is based on a genetic algorithm and uses UNIFAC to estimate relative
volatilities. Joback’s group contribution methods are used to estimate boiling and freezing points.
A number of enhancements, including seeding, evolving fitness functions, and biased gene
selection have been included. The method has been implemented in a computer program that
runs on a standard desktop computer. A number of industrially significant systems were
investigated, and the predicted solvents compare very favorably with those that are currently
in use.

1. Introduction compared to other methods, but unfortunately they are

often very susceptible to local minima traps.
The effectiveness of an extractive distillation process
Fortunately, there exists a family of optimization
relies on the choice of extractive agent. Although many
algorithms that retain the efficiency of mathematical
heuristic methods have been developed to assist in the
programming while still performing well in difficult
choice of solvent, these are mostly qualitative.
search spaces. They are also not susceptible to local
A more effective method for selecting solvents is
minima traps. These are genetic algorithms.
computer-aided molecular design (CAMD). In this method
the required properties of a solvent are specified, and A genetic algorithm (GA) is an optimization method
its structure is then calculated through the use of group- that is based on the assumption that Darwin’s theory
contribution methods. of natural selection is accurate. Individuals that are
Many CAMD algorithms have been proposed. These better adapted to their environment will have a better
may be broadly classified as interactive, combinatorial, chance of survival and thus pass on their genetic
knowledge-based, and mathematical programming meth- material to the succeeding generations. This is the
ods. principle of survival of the fittest.
Interactive methods1 allow the user to manually build GA consist of the following elements:
the solvent, while the relevant properties are continu- Evaluation Scheme. To use GA to solve optimiza-
ously recalculated. These methods rely to a large extent tion problems, a fitness function must be defined. The
on the knowledge of the user for their success. fitness function may take any form, but better solutions
In contrast to this, automated design algorithms will to the problem must have higher fitness values. The
attempt to generate solvents without assistance by the algorithm will attempt to find the individual with the
user. The first of these is the combinatorial method. In highest fitness.
its crudest form, this is a brute-force search. Although Encoding Scheme. Individual solutions are often
this method is guaranteed to find the optimal solution, referred to as chromosomes, in reference to the biological
it is extremely computationally intensive. Various re- equivalent. Like their biological counterparts, GA chro-
finements to the method have been proposed1 to limit mosomes are coded as linear strings. Each separate unit
the search space in order to make the problem more of information on a chromosome is referred to as a gene.
manageable. Many different encoding schemes have been used.
Knowledge-based methods2-6 are generally brute- One of the most popular is the use of binary strings.
force searches that limit search space with a number of Each binary digit in the string then represents a single
rules. These rules are meant to limit the search to gene. Other methods may include enumerated types.
physically viable molecules. Unfortunately, it is ex- This method is very useful when encoding molecular
tremely difficult to find a set of rules for viability that structures because each gene represents a specific,
will be generally applicable. Knowledge-based methods numbered functional group.
closely resemble combinatorial methods, and like the Population Scheme. Reproduction of chromosomes
latter they are generally extremely computationally may be handled in different ways. Two of the most
intensive. commonly used methods are generational and steady-
Most mathematical programming methods employ state reproduction. In generational reproduction, an
variations of mixed-integer nonlinear programming entire new generation is created from the current set
(MINLP) methods to find the number of each of a of chromosomes. This new generation then replaces the
specified set of submolecular groups that should be old generation as a whole. In steady-state reproduction
included in the solvent.7-11 These are then combined new chromosomes will continuously replace their parent
into molecules through an exhaustive search method. chromosomes if they have higher fitness values.
These methods are computationally efficient when To prevent the loss of good solutions in generational
reproduction, an elitist scheme may be used. In these
* To whom correspondence should be addressed. E-mail: cases a number of the best solutions in a generation,
[email protected]. usually 5-10%, are copied to the new generation
10.1021/ie9904753 CCC: $19.00 © 2000 American Chemical Society
Published on Web 03/29/2000
1424 Ind. Eng. Chem. Res., Vol. 39, No. 5, 2000

without modification. The rest of the new generation is fitness, which should be much higher than that of the
then created through reproduction by means of certain second chromosome. Thus, it will have a much better
operators. chance to reproduce and so transfer the desired genetic
Reproduction Scheme. New chromosomes are gen- material to the succeeding generation.
erated by modification of existing chromosomes called A domain of [0; 1] obviates the need for normalizing
parent chromosomes. The operations used may vary but fitness values for different properties before combining
almost always include point mutation and crossover. them into a single fitness value for the molecule.
The point mutation operation consists of replacing a These requirements prompted the use of a sigmoidal
single gene in the chromosome with another. This fitness function for each specified property.
operation requires only one parent chromosome. During
crossover two parent chromosomes are each cut in half, 1

[ ( )]
and one part is joined with a part of the other parent F) (1)
Pi - Pr
chromosome. The child chromosome will thus inherit 1 + exp -β
characteristics of both parents. This is an approximation Pr
of sexual reproduction in nature. Other operations may
include insertions, deletions, and order changes. Here F is the fitness value, Pi is the property value for
Parent chromosomes are usually selected with a the specific chromosome, and Pr is the required value
probability calculated from their fitness values. This for that property. β is a parameter that determines the
means that chromosomes with higher fitness values will slope of the function.
be selected more often, emulating the survival of the A fitness value is calculated for each property with
fittest principle. eq 1, and the weighted mean of these values is taken
GA have been applied to the design of polymers.12 It as the overall fitness for the molecule. The weights
was the purpose of this work to develop a genetic code assigned to each property are chosen by the user to suit
for the design of solvents for extractive distillation. the specific problem and may be changed if the relative
importance of the properties change. In this work the
2. Designing Solvents with GA weights were as follows: relative volatility (40%), boiling
point (30%), and freezing point (30%).
2.1. Basic Algorithm. The basic GA was adapted to
suit the specific problem of designing solvents for Any other properties that can be estimated with group
extractive distillation. contribution methods can also be added to the fitness
function if necessary.
Before this could be done, group-contribution methods
had to be found to estimate the different properties for Encoding Scheme. An apparent problem with using
which requirements were to be specified. In this work, GA to design molecules is the linearity of the chromo-
the UNIFAC13 method was used to calculate the relative somes. Restricting the search space to linear molecules
volatility and the methods of Joback14 were used to is undesirable, and a way had to be found to encode
estimate boiling and freezing points. nonlinear or even cyclic molecules in a linear structure.
Evaluation Scheme. When designing a solvent for Each GA chromosome is constructed from a number
extractive distillation, there are three primary concerns: of encoding units called genes. These genes are linearly
1. The relative volatility of the two key components combined to construct a chromosome. However, the
must be above a certain value. genes themselves need not be linear. Genes of arbitrary
2. The solvent must have a boiling point considerably complexity may be constructed by linearly combining
higher than that of the mixture components to enable different structural groups. UNIFAC and Joback con-
subsequent solvent recovery. stants can then be computed for these genes.
3. The high-boiling solvent must have a freezing point To satisfy the octet rule, two types of genes were
that is sufficiently low to prevent solidification at constructed: start/end genes with only one free bond
process temperatures. and middle genes with two free bonds. To construct
All three of these specifications are one-sided; i.e., the cyclic molecules, only middle genes were used.
required property must be higher or lower than a certain Using these predefined genes has a number of ad-
specified value. Although any type of fitness function vantages:
could be used, a continuous function with a domain in 1. Physically nonfeasible and reactive combinations
[0; 1] is preferred. Eliminating discontinuities in the of functional groups may be avoided to a large extent.
fitness function will improve convergence. This is best This precludes the need for extensive sets of rules to
explained by an example: test for physical viability.
Take as the fitness function a step function with 2. It allows more complex molecules than a simple
values of 1 or zero depending on whether the input is linear combination of UNIFAC groups.
above or below a required value. If the chromosomes of 3. It allows the simultaneous design of aliphatic and
which the fitness is being calculated has a value just aromatic molecules.
below the required value, it will have a fitness of zero. Because of the nature of the UNIFAC groups, there
This will be the same as the fitness of a second is also a third group of genes. These each consist of a
chromosome of which the value in question is much single UNIFAC group with no free bonds. These are
lower than the required value. Although the first dealt with separately.
chromosome will in all probability contain much of the Because molecules may vary greatly in size, the
genetic material necessary for attaining the required chromosomes used to encode them must also be able to
value for the specific property, it has no better chance vary in size. To this end, the number of middle genes
of reproducing than the second chromosome, which in a chromosome are allowed to vary from zero to some
contains none of the required genes. chosen maximum. When this type of variable-length
If a continuous fitness function is used, the first chromosome is used, the method is sometimes referred
chromosome in the above example will have nonzero to as a messy GA.15
 Ind. Eng. Chem. Res., Vol. 39, No. 5, 2000 1425

Table 1. Some of the Parameter Sets Tested

parameter A B C D
β 10 20 10 20
point mutation (%) 35 35 40 40
crossover (%) 30 30 20 20
insertion (%) 17.5 17.5 20 20
deletion (%) 17.5 17.5 20 20

Population Scheme. A random initialization tech-

nique may be used to create the initial population. A
larger population would generally decrease the number
of generations required to find good solutions but would
also increase the computer resources and processor time
required. These costs and benefits must be balanced
when determining the population size. In this work a Figure 1. The fitness of the best chromosome in each generation
for the sets of genetic parameters in Table 1.
population of 10 000 chromosomes was used.
Reproduction is generational with a 10% elitist policy.
This means that the best 10% of each generation will gene during mutation, the new gene must be selected
be copied unchanged into the next generation. The rest from the available gene pool. This is also done with the
of the new generation is created by evolution of the roulette wheel method. In this case a selection prob-
present generation by means of the operators described ability for every gene is calculated using the Robbins
below. table.17 As for the operators, the random number used
Generational reproduction was chosen, because it is for selection is generated between zero and the sum of
believed to be superior to the steady-state model.16 The these selection probabilities. This method ensures that
elitist policy helps to speed up convergence and also genes containing functional groups that will favor the
prevents the loss of good solutions due to random required separation will be selected more often than
mutations. others.
All operators and operands are selected with a The entire process may be summarized as follows:
probabilistic method. Each operator is assigned a fixed 1. Initialize a population of chromosomes.
selection probability, while the fitness of each chromo- 2. Evaluate each chromosome in the population by
some is used as its selection probability. doing the following:
Reproduction Scheme. Four operators are used: 2.1. Calculate the relative volatility, boiling point, and
point mutation, crossover, insertion, and deletion. The freezing point.
insertion and deletion operations are limited by the 2.2. Calculate the property fitness values.
chromosome structure, because start and end genes may 2.3. Calculate the global fitness of each chromosome.
not be inserted or deleted. Crossover may only be 3. Choose the best 10%, and copy these chromosomes
applied when there is space for two new chromosomes to the new generation.
in the new generation. 4. Create new chromosomes to fill up the remaining
Operators are selected with the roulette wheel method. 90% of the new generation.
In this method a random number is generated with a 4.1. Choose an operator.
value between zero and the total of all of the selection 4.2. Choose parent chromosome(s).
probabilities. The probabilities are then added one by 4.3. Choose the point on each chromosome where the
one until the subtotal exceeds the value of the random operator will act.
number. The last operator of which the selection prob- 4.4. Apply the operator to the parent chromosome(s).
ability was added to the subtotal is then selected. The 5. Copy the new chromosomes to the new generation
selection probabilities of the operators are point muta- until the generation has been filled.
tion (40%), crossover (20%), insertion (20%), and dele- 6. Replace the current generation with the new
tion (20%). generation.
These selection probabilities were chosen to maximize 7. Repeat from step 2 for the desired number of
both the speed of convergence and the fitness of the final generations.
generation. Various values were tested for each of the 8. Evaluate each chromosome in the final population.
parameters to find the optimal values. The best values 9. Return a list of chromosomes with the highest
were then used in several combinations to find the best fitness.
combination of probabilities. Some of these combinations 2.2. Improvements to the Algorithm. Although the
are shown in Table 1. The fitness of the best chromo- basic GA gives satisfactory results in many cases, there
some in each generation is plotted in Figure 1, using are a number of ways in which it may be improved.
each of these parameter sets. Missing Interaction Parameters. One of the main
Once an operator has been selected, the chromosomes problems experienced in CAMD is the availability of
on which it will act are chosen using the same roulette interaction parameters for the group contribution meth-
wheel method. In this case the fitness value of each ods that are used. The primary concern in this work was
chromosome is used as its selection probability. The the availability of UNIFAC parameters.
random number that is used for selection is generated In some publications missing parameters are taken
between zero and the sum of the fitness values of all of to be zero.14 This often leads to gross errors and is not
the chromosomes in the generation. The point on the acceptable. As an alternative, a penalty system was
chromosome where the operator will act is also chosen introduced. The fitness of a chromosome is penalized
randomly. for each missing interaction parameter. This leads to
When a gene must be inserted into a chromosome or the elimination of all candidate solvents with missing
a gene in the chromosome must be replaced with a new interactions within a few generations. However, if the
1426 Ind. Eng. Chem. Res., Vol. 39, No. 5, 2000

chromosome has some good traits, these may still be

passed on to its offspring before it is completely elimi-
nated.
Better Gene Selection Methods. Many of the
heuristics used to design solvents for extractive distil-
lation give qualitative indications of which functional
groups to use for certain systems. This knowledge was
incorporated into the algorithm to bias gene selection
toward those functional groups that are known to
perform better. This was done by increasing the selec-
tion probability of the genes that contain these preferred
functional groups. The selection probability for each
gene was calculated from the positions of its constituent Figure 2. The effect of solvent to feed ratio on relative volatility
for the acetone/methanol binary mixture.
functional groups on the Robbins chart.17 This method
also allows selected genes to be excluded by setting their
selection probabilities to zero. initialization. The value is then stepped up with suc-
ceeding generations until it reaches the required value.
Physical Feasibility. Although the predefined genes This allows the identification of better combinations of
allow the prevention of many physically nonfeasible genes from the very start.
combinations, there are still some combinations of genes
This is equivalent to the physical case where the
that may lead to nonfeasible or reactive molecules. To
environment is evolving at the same time as the species
prevent this, a small subset of the rules proposed by are evolving.
Gani and Brignole2 is used. Instead of eliminating these
chromosomes outright, their fitness values are penalized
for each violation of the rules. In this way their genetic 3. The SolvGen Program
material may still be used to create other viable
chromosomes. The algorithm was implemented in a computer pro-
gram called SolvGen using the 5th revision of UNI-
Incorporating the User’s Experience. One of the FAC.13 A total of 529 start/end genes and 368 middle
primary advantages of the interactive design methods genes have been defined. For practical purposes the
is that the user’s knowledge and experience is available maximum number of middle genes in a chromosome was
to the program. This is also possible in this method. The limited to 6. This limit may be easily changed. The
initial population may be seeded with molecules that program runs under Windows 95 and NT on a standard
are known or suspected to perform well for the specific desktop computer.
system. Because the fitness values of these chromo- The required values for the relative volatility and
somes would normally be higher than the randomly boiling and freezing points must be specified. The user
generated chromosomes, they are sure to be used in the may also specify which component should be recovered
construction of new chromosomes for the next genera- in the distillate. This ability gives great flexibility in
tion. The functional groups that are responsible for the the design of separation systems, especially for purifica-
performance of these solvents may also be recombined tion, where only small amounts of one component need
in new solvents that may well outperform their prede- be removed. In these cases it would be preferable to
cessors. The seeding process will also help the algorithm recover this component in the distillate.
to identify areas in the solution space that are known Both the mixture composition and the molecular
to contain good solutions. More time will be spent structures of the components that are to be separated
exploring these areas. must be specified. Additionally, a vapor pressure equa-
Evolving Fitness Functions. The sigmoidal fitness tion is specified.
function performs very well in most cases, but there is Figure 2 shows a plot generated by SolvGen to show
at least one disadvantage to using it. the effect of the solvent-to-feed ratio on the relative
If the required value for a certain property is rela- volatility. This information is very important. The
tively high (or low), it is often the case that none of the commonly used practice of selecting solvents based on
chromosomes in the initial, random generation will their resulting relative volatility at infinite dilution may
come close to this specification. Because the fitness cause good solvents to be discarded in favor of others
that do not perform as well at process concentrations.
function is flat in both extremes, there will be very little
This may be clearly seen in Figure 2, where the solvent
difference in the fitness values of chromosomes that are,
that performs better at infinite dilution does not perform
for instance, 10% and 100% below the required value.
as well at the lower concentrations that will actually
As a result, the algorithm will not be able to identify be used in practice.
the better chromosomes and use their characteristics Because it is undesirable to have azeotropes form
to improve the next generation. This situation will between the solvent and any of the mixture components,
continue until point mutation produces a chromosome SolvGen can be used to check for the formation of
with a significantly better value for the specific property. azeotropes. The results of such a check may be seen in
The dependency on random mutations is not satisfac- Figure 3. This is a binary vapor-liquid equilibrium plot
tory. Fortunately, there is a simple solution to the generated by SolvGen for the solvent and a selected
problem. mixture component.
Instead of using a fixed requirement for any specific SolvGen also has an interactive design mode, which
property, the required value is initially set to a value is useful for making alterations to generated solvents
close to that which is attained through the random or for quick tests of available solvents.
 Ind. Eng. Chem. Res., Vol. 39, No. 5, 2000 1427

Table 2. Generated Solvents and Predicted Properties

for Ethanol (1)/Water (2)
Distillate: Ethanol
R12 Tboil Tfreeze
solvent (UNIFAC) (Joback) (Joback)
hexachlorobutadiene 8.41 523 248
dimethyl sulfoxide 3.44 462 292
industrial solvent
ethylene glycol 2.54 471 258
Distillate: Water
R21 Tboil Tfreeze
solvent (UNIFAC) (Joback) (Joback)
dodecane 4.72 474 224
Figure 3. A VLE plot for a binary mixture of a selected tridecane 4.83 497 236
component and the generated solvent to find azeotropes. tetradecane 4.95 520 247
diethylbenzene 5.81 460 241
4. Case Studies p-cymene 6.17 459 226
butylbenzene 5.85 455 228
Five binary systems were chosen to test the algo- industrial solvent
rithm. In each case equimolar feeds were used with a none available N/A N/A N/A
solvent-to-feed ratio of 4:1 and at a temperature of 350 Experimental Results
K. All properties shown, including those of the industri-
component liquid mole fraction R12
ally used solvents, are as predicted by the group
contribution methods. ethanol (1) 0.163
The aim of these case studies was to evaluate the water (2) 0.181
hexachlorobutadiene 7.26
methodology of designing solvents and not necessarily
the accuracy of the group contribution methods. Com-
paring the predicted properties of the generated solvents Table 3. Generated Solvents and Predicted Properties
for Acetone (1)/Methanol (2)
with the predicted properties of those used in industry Distillate: Acetone
allows the algorithm’s effectiveness to be evaluated
R12 Tboil Tfreeze
without penalizing it for failures in the group contribu- solvent (UNIFAC) (Joback) (Joback)
tion methods.
N,N′-dimethylethylenediamine 6.68 437 262
Group contribution methods are very powerful predic- N,N′-dimethyl-1,3-propanediamine 6.30 460 273
tive tools but are not perfect. All such predictions should dimethyl sulfoxide 2.98 462 292
always be verified experimentally. A number of the industrial solvent
systems were tested with single-stage, total reflux ethylene glycol 3.36 471 258
water 3.38 373 273
distillation experiments. The results of these tests are
included in the following tables. Distillate: Methanol
In each case the algorithm was allowed to run for 10 R21 Tboil Tfreeze
generations. This translates to a total computer time solvent (UNIFAC) (Joback) (Joback)
per system of 10-15 min on a Pentium II 266 MHz PC. hexachlorobutadiene 6.62 523 248
System 1. Ethanol/Water. This mixture forms a dibromoethane 3.50 377 231
minimum-boiling azeotrope, and extractive distillation undecane 1.43 451 213
dodecane 1.45 474 225
may be used as an alternative to azeotropic and pres- tridecane 1.46 497 236
sure-swing distillation. The solvents generated for this industrial solvent
system are listed in Table 2. none available N/A N/A N/A
The solvent listed in the literature for this system is
Experimental Results
ethylene glycol. This solvent will allow the recovery of
ethanol in the distillate with a predicted relative volatil- component liquid mole fraction R12
ity of 2.54 (UNIFAC). In contrast to this, the first acetone (1) 0.065
generated solvent, hexachlorobutadiene, is predicted to methanol (2) 0.086
cause a relative volatility more than 3 times this value. dimethyl sulfoxide 0.848 3.21
This solvent was tested and performed very well (see
Table 2). relative volatility of 2.98. This is somewhat worse than
Solvents were also generated to reverse the relative the solvents currently used, but the boiling point of
volatility of the system to facilitate the recovery of water DMSO is much higher than that of water. This system
in the distillate. These are listed in the second part of was tested experimentally and found to perform com-
Table 2. No solvents could be found in the literature parably with the industrially used solvents (see Table
for this separation. 3).
System 2. Acetone/Methanol. This system also Again, solvents were also designed to recover the
forms a minimum-boiling azeotrope. Solvents listed in other component, methanol, in the distillate. These are
the literature for this system include ethylene glycol and listed in the second part of Table 3.
water. These solvents will result in relative volatilities System 3. Ethanol/Ethyl Acetate. This system also
of 3.36 and 3.38, respectively (UNIFAC), with acetone forms a minimum-boiling azeotrope. Trimethylbenzene
as the distillate. Two of the solvents generated for this is listed in the literature as a solvent to recover ethanol
separation result in relative volatilities in excess of 6, in the distillate. This solvent is predicted to give a
almost double that of the current solvents. The third, relative volatility of 3.84 (UNIFAC). The solvents gener-
dimethyl sulfoxide (DMSO), is predicted to give a ated for this separation (Table 4) either result in higher
1428 Ind. Eng. Chem. Res., Vol. 39, No. 5, 2000

Table 4. Generated Solvents and Predicted Properties Table 5. Generated Solvents and Predicted Properties
for Ethanol (1)/Ethyl Acetate (2) for Methanol (1)/Methyl Acetate (2)
Distillate: Ethyl Acetate Distillate: Methyl Acetate
R21 Tboil Tfreeze R21 Tboil Tfreeze
solvent (UNIFAC) (Joback) (Joback) solvent (UNIFAC) (Joback) (Joback)

dimethyl sulfoxide 4.55 462 292 N,N′-dimethylethylenediamine 6.58 391 240

ethylene glycol 2.97 471 256 N,N′-dimethyl-1,3-propanediamine 5.92 414 251
diethylene glycol 2.46 498 278 diethylenetriamine 4.12 486 354
1,4-diaminobutane 2.83 436 301
diethylenetriamine 2.49 531 380
1,3-diaminopentane 2.61 458 297
hexamethylenetriamine 2.28 532 376
ethylene glycol 1.84 471 258
industrial solvent dimethyl sulfoxide 3.45 462 292
none available N/A N/A N/A water 4.22 373 273
Distillate: Ethanol industrial solvent
2-methoxy ethanol 1.95 383 206
R12 Tboil Tfreeze
solvent (UNIFAC) (Joback) (Joback) Distillate: Methanol

butylthiophene 5.85 454 292 R12 Tboil Tfreeze

solvent (UNIFAC) (Joback) (Joback)
dodecane 3.54 474 225
tridecane 3.60 497 236 hexachlorobutadiene 4.16 523 248
tetradecane 3.65 520 247 1,4-dibromobutane 3.60 423 254
industrial solvent ethylthiophene 5.48 408 270
trimethylbenzene 3.84 442 242 tetrachloro ethylene 5.13 394 251
dodecane 2.62 474 225
Experimental Results tridecane 2.62 497 236
industrial solvent
component liquid mole fraction R21
none available N/A N/A N/A
ethanol (1) 0.149
ethyl acetate (2) 0.069 component liquid mole fraction R12
diethylene glycol 0.782 3.05 methanol (1) 0.112
methyl acetate (2) 0.111
component liquid mole fraction R21 tetrachloroethylene 0.776 2.46
ethanol (1) 0.055
ethyl acetate (2) 0.019 component Liquid mole fraction R12
dimethyl sulfoxide 0.926 2.89 methanol (1) 0.005
methyl acetate (2) 0.023
tetrachloroethylene 0.972 5.44
relative volatilities or will give approximately the same
relative volatility but have higher boiling points. This Table 6. Generated Solvents and Predicted Properties
will result in easier solvent recovery in the second stage for Cyclohexane (1)/Benzene (2)
of the extractive distillation process. Again solvents Distillate: Cyclohexane
were also generated to reverse the relative volatility of R12 Tboil Tfreeze
the mixture to facilitate the recovery of ethyl acetate solvent (UNIFAC) (Joback) (Joback)
in the distillate. acetonyl acetone 2.58 444 257
Two of these systems were tested with single-stage, dimethyl maleate 2.79 457 236
total reflux distillations. The performances of diethylene dimethyl phthalate 3.67 576 325
N-methyl-2-pyrrolidone 3.76 475 250
glycol and DMSO were found to closely match the industrial solvent
predicted values. aniline 2.65 436 267
System 4. Methanol/Methyl Acetate. As with the
previous systems, a minimum-boiling azeotrope is formed.
in the distillate. The solvents generated for this separa-
Ethylene glycol monomethyl ether (2-methoxyethanol)
tion are listed in Table 6. The only solvent not to result
is listed in the literature as a possible solvent for the
in a higher relative volatility is acetonyl acetone;
separation of this mixture. This will result in the
however, this solvent has a higher boiling point than
recovery of methyl acetate in the distillate with a
aniline.
relative volatility of 1.95 (UNIFAC). A number of
solvents were generated that are predicted to result in
higher relative volatilities (Table 5). The only one of 5. Conclusions
these with a slightly lower predicted relative volatility
is ethylene glycol. In this case the boiling point of In this work a flexible and efficient algorithm was
ethylene glycol is much higher than that of 2-methoxy- proposed for designing molecules with specific proper-
ethanol, which will result in easier solvent recovery in ties.
the second stage of the process. As with the previous This algorithm was applied to the design of solvents
systems, solvents were generated to recover the second for extractive distillation. This is, however, by no means
component in the distillate. They are all predicted to the only problem to which this algorithm may be
perform very well. applied. Wherever the properties of a substance may
One of these solvents, tetrachloroethylene, was tested be calculated from its molecular structure, CAMD
and found to perform better than predicted (see Table algorithms may be employed to find structures that will
5). yield certain required values of those properties.
System 5. Benzene/Cyclohexane. This system also This algorithm, like all CAMD algorithms, is limited
forms a minimum-boiling azeotrope. Aniline is given in by the accuracy of the group contribution methods that
the literature as a possible solvent, resulting in a are used. The availability of accurate parameters for
relative volatility of 2.65 (UNIFAC) with cyclohexane these methods is of crucial importance.
 Ind. Eng. Chem. Res., Vol. 39, No. 5, 2000 1429

Literature Cited (10) Maranas, C. D. Optimal Computer-Aided Molecular De-

sign: A Polymer Design Case Study. Ind. Eng. Chem. Res. 1996,
(1) Joback, K. G.; Stephanopoulos, G. Designing Molecules 35, 3403.
Possessing Desired Physical Property Values; FOCAPD’89, Snow- (11) Churi, N.; Achenie, L. E. K. Novel Mathematical Program-
mass, CO, 1989; p 363. ming Model for Computer Aided Molecular Design. Ind. Eng.
(2) Gani, R.; Brignole, E. A. Molecular Design of Solvents for Chem. Res. 1996, 35, 3788.
Liquid Extraction based on UNIFAC. Fluid Phase Equilib. 1983, (12) Venkatasubramanian, V.; et al. Computer-Aided Molecular
13, 331. Design using Genetic Algorithms. Comput. Chem. Eng. 1994, 18
(3) Brignole, E. A.; et al. A Strategy for the Design and Selection (9), 833.
of Solvents for Separation Processes. Fluid Phase Equilib. 1986, (13) Hansen, H. K.; et al. Vapor-Liquid Equilibria by UNIFAC
29, 125. Group Contribution. 5. Revision and Extension. Ind. Eng. Chem.
(4) Nielsen, B.; et al. Computer-Aided molecular Design by Res. 1991, 30, 2352.
Group Contribution Method. Comput. Appl. Chem. Eng. 1990, 227. (14) Reid, R. C.; et al. The Properties of Gases and Liquids, 4th
(5) Gani, R.; et al. A Group Contribution Approach to Computer- ed.; McGraw-Hill: New York, 1987.
Aided Molecular Design. AIChE J. 1991, 37 (9), 1318. (15) Koza, J. R. Genetic Programming: On the Programming
(6) Pretel, E. J.; et al. Computer-Aided Molecular Design of of Computers by Means of Natural Selection; MIT Press: Cam-
Solvents for Separation Processes. AIChE J. 1994, 40 (8), 1350. bridge, MA, 1992.
(7) Macchietto, S.; et al. Design of Optimal Solvents for Liquid- (16) Davis, L., Ed. Handbook of Genetic Algorithms; Van
Liquid Extraction and Gas Absorption Processes. Trans. Inst. Nostrand Reinhold: New York, 1991.
Chem. Eng. 1990, 68A, 429. (17) Robbins, L. A. Liquid-Liquid Extraction: A Pretreatment
(8) Naser, S. F.; Fournier, R. L. A System for the Design of an Process for Wastewater. Chem. Eng. Prog. 1980, 76 (10), 58.
Optimum Liquid-Liquid Extractant Molecule. Comput. Chem.
Eng. 1991, 15 (6), 397. Received for review June 29, 1999
(9) Vaidyanathan, R.; El-Halwagi, M. Computer-Aided Syn- Revised manuscript received January 20, 2000
thesis of Polymers and Blends with Target Properties. Ind. Eng. Accepted January 26, 2000
Chem. Res. 1996, 35, 627.
IE9904753