Computers and Chemical Engineering 23 (1999) 1381 – 1394

www.elsevier.com/locate/compchemeng

Environmentally benign solvent design by global optimization

Manish Sinha, Luke E.K. Achenie *, Gennadi M. Ostrovsky
Department of Chemical Engineering, Uni6ersity of Connecticut, Storrs, CT 06269, USA

Accepted 20 September 1999

Abstract

Changes in environmental regulations have often resulted in the need to replace existing solvents with more environmentally
benign substitutes. For example, solvents such as 1,1,1, trichloroethane are being phased out within the next few years. In general,
‘designer’ compounds, which have desired properties, can be obtained with the help of computer-aided molecular design (CAMD)
approaches. In recent years, the above product design problem has often been posed as a mathematical programming problem.
In this framework, the desired attributes of the compound are posed as performance objective and constraint functions. Nonlinear
functions, such as those associated with solubility parameter models seen in solvent design, can lead to multiple local solutions
(i.e. optimization minima). For these cases, the failure to design the globally optimal compound (using a local optimization
approach) can introduce a source of uncertainty. Thus, there is a need for global optimization techniques in product design. To
address this need, we have developed a new global optimization algorithm that exploits the structure of the CAMD formulation.
We have used this algorithm to study the design of environmentally benign solvents for surface cleaning applications in the
printing industry. © 1999 Elsevier Science Ltd. All rights reserved.

Keywords: Mixed-integer; Global optimization; Computer-aided molecular design

1. Introduction sign solvent alternatives that have a desired set of

Solvents are extensively used as processing materials
in the chemical industry and solvent-based industries
(e.g. resin, coating and paint industry). However, many
organic solvents such as 1,1,1, trichloroethane are being
phased out in response to directives from federal agen-
cies such as the EPA and OSHA (Krishner, 1995). This
has prompted the search for new environmentally be-
nign solvent substitutes. Traditionally, the search for a
substitute solvent has been done through experience,
database searches and bench scale synthesis of solvent
alternatives. All these approaches have severe limita-
tions and are often unable to effectively find the opti-
mal solvent substitute. Computer-aided molecular
design (CAMD), also referred to as product design, has
emerged as a powerful tool for identifying product
alternatives. In this approach, mathematical models
that relate different physico-chemical properties to the
molecular structure are systematically ‘inverted’ to de-
* Corresponding author. Tel.: +1-860-4862756; fax: + 1-860-
4862959.
E-mail address: achenie@engr.uconn.edu (L.E.K. Achenie)
properties.
Constantinou, Bagherpour, Gani, Klein and Glen
(1996) have reviewed the CAMD problem formulation
and its applications in detail. Some solvent applications
for which the CAMD methodology has been proposed
are liquid–liquid extraction (Macchietto, Odele &
Omatsone, 1990; Odele & Machietto, 1993), multicom-
ponent gas absorption processes (Pistikopoulos & Ste-
fanis, 1998), separation processes (Pretel, Lopez, Bottini
& Brignole, 1994) and solvent blends for paint formula-
tion (Klein, Wu & Gani, 1992). Other applications
include polymer design (Vaidyanathan & El-Halwagi,
1994a; Maranas, 1996; Vaidyanathan, Gowayed & El-
Halwagi, 1998), and refrigerant design (Joback &
Stephanopoulos, 1989; Gani, Nielsen & Fredenslund,
1991; Duvedi & Achenie, 1996). A recent issue of
Computers & Chemical Engineering focused on com-
puter-aided design of chemical compounds (Mavrovou-
niotis, 1998).
Two classes of solution approaches, namely enumer-
ation techniques and knowledge based strategies have
been employed for certain CAMD formulations (Gani

0098-1354/99/$ - see front matter © 1999 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 9 8 - 1 3 5 4 ( 9 9 ) 0 0 2 9 9 - 9
1382 M. Sinha et al. / Computers and Chemical Engineering 23 (1999) 1381–1394
& Brignole, 1983; Brignole, Bottini & Gani, 1986;
Joback & Stephanopoulos, 1989; Pretel et al., 1994).
These methods rely significantly on heuristics and user
intervention. Another class of solution approaches are
stochastic based approaches such as genetic algorithms
(Venkatasubramanium, Chan & Caruthers, 1994) and
simulated annealing (Marcoulaki & Kokosis, 1998).
In recent years, the CAMD problem has often been
posed as a mixed integer nonlinear mathematical pro-
gramming (MINLP) formulation in which the user
specifications are modeled as constraints and the desir-
able property is modeled as an objective function. A
CAMD problem may also be formulated as a con-
straint satisfaction problem (i.e. feasibility problem).
Mathematical programming based solution strategy
have been successfully employed for solving the CAMD
problem (Macchietto et al., 1990; Odele & Machietto,
1993; Duvedi & Achenie, 1996; Maranas, 1996; Pistiko-
poulos & Stefanis, 1998; Vaidyanathan et al., 1998).
Many methods for solving mathematical program-
ming based CAMD problems use outer approximation
(OA) algorithm (Viswanathan & Grossman, 1990)
which consists of a sequence of iterations consisting of
NLP (nonlinear programming) subproblem and MILP
(mixed-integer linear programming) ‘master’ problems.
One weakness of such methods is that the NLP sub-
problem may get trapped in a local solution for non-
convex mathematical programming formulations
(nonconvexity in objective function and/or constraints)
as shown in Fig. 1.
Another issue in CAMD is the problem of model
uncertainty. The latter is independent of the search
strategy employed during the design stage. In addition
to model uncertainty, another source of uncertainty is
the fact that a local optimization approach may not
reach the global solution (i.e. that one cannot find a
better solution), if this solution happens to be the best
Fig. 1. Schematic of a linear objective function surface with nonlinear constraints. A local optimization approach may not find the optimal point
from any of the five initial starting (or initial guess (IG)) points.
even in the presence of other uncertainties (see Fig. 1).
Thus, there is a need to develop computationally effi-
cient global optimization algorithms for CAMD.
Global optimization based on interval analysis
(Vaidyanathan & El-Halwagi, 1994b, 1996) has been
employed to solve relatively small sized CAMD prob-
lems. However realistic CAMD problems have several
integer variables describing a molecular structure.
Moreover, most of the variables participate in nonlin-
ear equations. To address this need, we have developed
a new global optimization algorithm that exploits the
structure of the CAMD formulation.
Now we need to consider the issue of the validity of
a global optimization approach in the presence of
model uncertainty. Let us suppose that the CAMD
model has integer variables (i.e. structural variables) as
the only design variables. Then exhaustive enumeration
will yield the same global solution as a global optimiza-
tion approach. Therefore in this regard enumeration
and global optimization have the same deficiency with
respect to model uncertainties. That is, unless model
uncertainties are explicitly accounted for, a global solu-
tion may not necessarily be the best solution in the
presence of model uncertainties. In fact the issue of
model uncertainty is independent of the search strategy
employed.
The paper is organized as follows, in Section 2, some
important models for the CAMD approach to design
environmentally benign solvents are discussed. Next,
the development of a global optimization algorithm for
the CAMD problem is presented in Section 3. In Sec-
tion 4, we develop the relaxation (underestimator) for
the nonconvex solubility parameter constraint. Finally,
in Section 5, the efficacy of the proposed algorithm is
demonstrated via a case study in which the design of
solvent alternatives for use in a commercial cleaning
application is considered.
 M. Sinha et al. / Computers and Chemical Engineering 23 (1999) 1381–1394 1383

Fig. 2. Correlation between environmental measure for bioconcentration factor (log BCF) with octanol – water partition coefficient (log Kow)
(Nendza & Hermens, 1995). Printed with kind permission from Kluwer Academic Publishers.

2. Important models for solvent design Solute and solvents are represented as points in the
three-dimensional Hansen solubility space. The dis-
2.1. Solubility parameter tance, Rij, of a solute (i ) from a solvent ( j ) is a
measure of the solute–solvent interaction (see Fig. 3).
The amount of solute that a solvent can dissolve Associated with the solute point is a radius of interac-
(capacity) and the ability of the solvent to selectively tion, R*. If Rij is less than R*, then the solvent is likely
dissolve a solute (selectivity) are two important proper- to dissolve the solute. Rij is given by
ties that determine a solvent’s performance. These R ij = 4(dD − d*D )2 + (dP − d*P )2 + (dH − d*H )2 5 (R*)2
properties can be modeled by the equilibrium relation- (2)
ships that govern the distribution of a solute in a
solvent, which in turn can be modeled by an activity where
coefficient model such as UNIFAC (Fredenslund,
Gmehling & Rasmussen, 1977). In addition, there are % ni FDi
i
other models for characterizing solvent performance dD = (2a)
that are prevalent in the solvent industry. Some of these Vo + % ni Vi
models are based on the solubility parameter. i

The solubility parameter has been successfully used

in solvent-based industry. Applications in which the
solubility parameter approach have been employed in-
clude finding compatible solvents for coating resins,
finding an effective mass separating agent, predicting
the swelling of cured elastomers by a solvent, and
estimating liquid–liquid equilibrium in polymer systems
(Grulke, 1989; Archer, 1996). In a recent article, Gess-
ner recommends use of solubility parameter for design-
ing solvents for surface cleaning and surface coating
applications (Gessner, 1998). In this work also the
solvent performance is characterized by the Hansen
solubility parameter (dT) (Hansen, 1969). This parame-
ter in turn is made up of hydrogen bonding interaction
(dH), polar interactions (dp), and nonpolar (dispersive)
interaction (dD) as shown in Eq. (1). Fig. 3. Solute interaction sphere of phenolic resin and the swelling
box of polyisoprene rubber. The design space is the shaded region
dT =
d 2D +d 2P +d 2H (1) (outside the swelling box but inside the interaction sphere).
1384 M. Sinha et al. / Computers and Chemical Engineering 23 (1999) 1381–1394
' % ni (FPi )2
i
dP =
D
Vo +% ni Vi
i
% ni (−UHi )
dH = i (2c)
Vo + % ni Vi
i implicitly constrains BCF and LC50 for the solvent.
Here Fdi, FPi and UHi are the group contribution
parameters associated with the dispersion, polar and
hydrogen bonding solubility parameters, respectively
for the ith group. ni is the number of times i th group is
present in a molecule, Vo is a constant associated with
the liquid molar volume prediction model and Vi is the
group contribution associated with the ith group for
prediction of liquid molar volume.
The group contribution parameters are taken from
van Krevlen and Hoftyzer’s work on the dispersion and
polar components (van Krevelen & Hoftyzer, 1976) and
from Hansen and Beerbower for the hydrogen bonding
component (Hansen & Beerbower, 1971). For predict-
ing the liquid molar volume the model proposed by
Constantinou and Gani (1994) is used.
2.2. En6ironmental health and safety requirements
Environmental, health and safety concerns are im-
portant issues for solvent selection. Solvent properties
targeted by environmental regulations include biocon-
centration factor (BCF), toxicity and flammability.
The environmental and health impact of a solvent
have been correlated to the octanol – water partition
coefficient (Kow) by many researchers (Lyman, Reehl &
Rosenblatt, 1981; Horvath, 1992; van Leeuwen & Her-
mens, 1995). Kow is a measure of hydrophobicity and
relates the equilibrium partition of solvent between
water and a water-immiscible liquid phase. For most
chemicals log Kow ranges from −3 to 7 and may be
estimated by Hansch and Leo’s approach (Hansch &
Leo, 1979), as follows
Log Kow = sum of fragments (x 0) + factors (x 1)
Here empirically derived atomic or group fragment
constants (x 0) and structural factors (x 1) are used.
Several environmental and health metrics such as
(BCF and LC50) depend on the Kow. A solvent with a
high bioconcetration factor has a greater likelihood of
accumulating in and harming living tissue. One of
several correlations for BCF shown in Eq. (4) (Veith &
Konasewich, 1975), also see Fig. 2. (Nendza & Her-
mens, 1995)
log BCF=0.76 (log Kow) − 0.23
LC50 measures the lethal concentration of a chemical
and can be used to quantify the solvent toxicity. The
LC50 value has been correlated by the following rela-
tion (Konemann, 1981)
(2b)
log LC50 = − 0.87(log Kow)− 0.11 (5)
Although the theoretical basis of these models is lim-
ited, they provide practical tools for risk assessment. In
this case study, the environmental impact of the solvent
is considered only by constraining Kow value. This
A solvent that can form potentially explosive vapor
mixtures with air is an occupational hazard and often
results in higher insurance costs. Safety requirements
can be quantified by the solvent’s flash point. Most of
the predictive techniques relate the flash point to the
boiling point. For example, Butler, Cooke, Lukk and
Jameson (1956) proposed a correlation for paraffins,
aromatics and cyloparaffins
Tf = 0.683TB − 119 (6)
Other equations showing direct relationship between
the flash point and the boiling point have been pro-
posed and are reviewed by Lyman et al. (1981). In this
case study we implicitly account for safety by constrain-
ing the boiling point. We note that all correlations
including the ones discussed above have a limited re-
gion of validity.
2.3. Other considerations
One important consideration in solvent substitution
is the desire to use existing equipment and processing
technology. This requires that both the old and the
substitute solvent should have similar transport and
other properties such as viscosity, diffusivity, density,
surface tension, heat capacity (Cp), and vapor pressure
(P vp). Zhao and Cabezas (1988) have reviewed molecu-
lar property considerations for design of substitute
solvents.
One important property in many solvent design
(3)
problems is the heat of vaporization (DH6 ) which is a
good measure of the energy efficiency associated with
the drying operation. The trivial observation that the
solvent has to be a liquid at the operating condition
leads to constraints on the boiling and melting point of
the solvent molecules. Additional constraints are
needed to design structurally feasible molecular struc-
tures (see for example Churi & Achenie, 1996). Con-
straints can also be imposed to eliminate unstable
molecules. For example, a small molecule made up of
three ether group may not exist. In the next section we
(4)
consider the structure of the CAMD problem for sol-
vent design and develop a global optimization al-
gorithm that exploits this structure.
 M. Sinha et al. / Computers and Chemical Engineering 23 (1999) 1381–1394
3. Solvent design: a reduced space branch and bound
approach
3.1. The CAMD problem formulation
A typical molecular design problem may bemodeled
as
minimize (maximize) pobj(U, u)
U
subject to
p Lk 5pk (U, u)5p U
k Ö k =1, 2, 3, ... , m1
hj (U, u)= 0.0 Ö j =1, 2, 3, ... , m2
where U is a vector of variables that define the molecu-
lar structure, u is a vector of group contribution
parameters, pobj is the objective function (some undesir-
able property to be minimized during design), pk is the
specified property requirements constrained between an
upper (p U U constraints.
k ) and a lower value (p k ) and hj are the
equality constraints generally associated with structural
feasibility requirements. The number of inequality con-
straints is 2m1 and the number of equality constraints is
m2.
pk = j nju 1j /j nju 2j
n
Most of the property constraints are of the form
a
where ‘1’ and ‘2’ correspond to
group contribution parameters associated with a prop-
erty of type ‘1’ and another molecular property of type
‘2’; a is an exponent (generally equal to 1). For exam-
ple, molar density is the ratio of molar weight (‘1’) and
molar volume (‘2’). Other examples of such models
include glass transition temperature and specific heat
capacity (Maranas, 1996). Transformation of such con-
straints into a linear form is straightforward. However
some property constraints are of the form
pk = f 1NL % nju aj /f 2NL % nju bj
j j
1 2
where f NLand f are nonlinear functions. Group
NL
contribution models for solubility parameter are of this
form (Eqs. (2b) and (2c)). Other examples of these
models include Reidels method for prediction of heat of
vaporization and vapor pressure by corresponding
states correlations (Reid, Prausnitz & Poling, 1987). It
is not always possible to reformulate these constraints
into linear or convex forms.
The nonlinear mathematical programming model for
the CAMD problem (PMD) has the following features:
1. It is a mixed integer nonlinear problem (MINLP)
problem involving a large number of variables.
2. The number of linear constrains are much greater
than the number of nonlinear constraints.
1385
We now present the development of a branch and
bound global optimization algorithm that exploits this
problem structure. At first we lay out the steps of a
branch and bound algorithm, discuss its limitation for
the CAMD problem, and then propose modifications.
3.2. Branch and bound
(PMD) Consider a general MINLP problem
f= minimize f(x, y, u) (PORG)
x,y
subject to
gi (x, y, u) 50 i= 1,2, ... , n
hj (x, y, u) =0 j= 1,2, ... , m
x B xBx ,
L U
y(0, 1)
where x is a vector of continuous variables, y is a vector
of integer (binary) variables and u is a vector of
parameters. There are n inequalities and m equality
We employ a branch and bound method for solving
the problem. In a conventional branch and bound
method a region is partitioned into subregions Tk,
which has the form
Tk = {x, y : x L(k) 5 x5x U(k); y L(k) 5 y5y U(k)}
i.e. a subregion (Tk ) is defined by bounds on the design
variables (x, y). We will designate the optimization
problem for subregion Tk as P ORG (k)
.
In order to estimate the lower bound (LBk ) of the
objective function f(x, y, u) in the subregion Tk, we
obtain a relaxation of the problem (P (k)
ORG) by replacing
the nonlinear constraints and objective function (if
nonlinear) by their linear underestimators. In effect we
obtain a mixed integer linear programming (MILP)
problem which can be solved by any standard MILP
solver such as lp – solve (Schwab, 1997) or OSL (IBM,
1992).
If the solution point (x*, y*) of the lower bound
problem satisfies all constraints in problem P (k)
ORG, then
it is a valid solution for the calculation of the upper
bound (UBk = f(x*, y*, u)). In the case of a linear
objective function, LBk = UBk and the solution point
(x*, y*) is a global minimum for f(x*, y*, u) in Tk. If
the solution for the calculation of the lower bound
problem does not satisfy all constraints of P (k)
ORG then
the upper bound (UBk ) for subregion Tk is computed
by replacing the nonlinear constraints and objective (if
nonlinear) by their respective linear overestimators. A
generic branch and bound algorithm for global opti-
mization (Horst & Tuy, 1990) have the following steps.
Step 0: Specify a tolerance (o). Compute the lower
bound (LBo) and upper bound (UBo) for the original
region To and insert it in a list L. Set the overall lower
3. Most of the components of the design vector (U) bound LWBD to−  and overall upper bound UPBD
participate in the nonlinear terms. to − .
1386 M. Sinha et al. / Computers and Chemical Engineering 23 (1999) 1381–1394
Fig. 4. Flowchart for global optimization via reduced space branch
and bound.
Step 1: Remove a region Tk from the list L that has
the lowest LBk (lower bound). If the list L is empty
then the problem is infeasible and has no solution.
Step 2: If LBk \ LWBD set LWBD =LBk and if
UBk B UPBD set UPBD=UBk. Partition the region Tk
into s subregions (T 1k, T 2k, … , T sk) by splitting one or
more of the variables (x, y). In many algorithms only
one of the design variables is bisected.
Step 3: Compute lower and upper bounds (LBik, UBik
Öi =1, 2, … , s) for each subregion generated in step 2.
A subregion T ik is inserted in the list L if it satisfies (i)
the problem of calculation of lower bound resulted in a
feasible solution, and (ii) LBik BUPBD.
If there is a large number of search variables partici-
pating in the nonlinear constraints or objective func-
tion, step 2 may generate a large number of subregions
(or nodes). The above is true for the CAMD problem.
To circumvent this dimensionality problem, we have
developed a branch and bound algorithm that performs
branching based on certain splitting functions, which
are linear functions of x and y.
3.3. Using splitting functions in the branch and bound
algorithm
In our approach, a region is not defined by bounds
on the variables but by bounds (al and bl ) on the
splitting functions (cl ). The problem formulation for
subregion Tk becomes
f= minimize f(x, y, u)
x,y
gi (x, y, u) 50 i= 1,2, ... , n
hj (x, y, u) 50 i= 1,2, ... , m (P. K
ORG)
al 5 cl (x, y, u)5 bl lSF
x L B xBx U, y(0, 1)
where cl (x, y, u) are linear splitting functions and SF is
a set of splitting functions.
In the proposed algorithm, partitioning of a region
into two subregions is performed on one of the splitting
functions as described below.
Modified step 2: If LBk \ LWBD set LWBD=LBk.
If UBk B UPBD set UPBD = UBk. Partition Tk into
two subregions (T Pk and T Q k ) by splitting on one split-
ting function. For example the rth function, such that
T Pk = {x, y: ar 5 cr{x, y)Bcr(x*, y*); al 5 cr(x, y)5bl
Ö(1SF and l" r)}. Similarly T Q k = {x, y: cr(x*, y*)5
cr(x, y)5 br ; al 5 cr(x, y)5bl Ö(1 SF and l"r)},
where x*y* is the solution point for the calculation of
the lower bound.
The proposed algorithm is also shown as a flowchart
in Fig. 4. To construct underestimators and overestima-
tors, we represent a nonlinear function as a tree struc-
ture (see Fig. 6) with the design variables at the lowest
level. The underestimators are constructed in a stepwise
procedure starting from the lowest level going up to the
first level for computation of bounds. This is followed
by a reverse sweep from the first level to the lowest level
Step 4: If UPBD– LWBD \o then go to step 2, else
the algorithm has converged to the global solution.
Note that in order to estimate the lower and upper
bound we need to solve an MILP. As a result, the
termination criteria stated in STEP 4 guarantees that
the final solution is the global solution for the original
problem PORG. We note that the final solution point to
the original problem (PORG) corresponds to the solution
of the upper bound problem for the region j such that
UBj =UPBD.
for construction of the underestimators. In our ap-
proach, the dimensionality of the lower and upper
bound problem is no larger than the dimensionality of
the original problem PORG. It is interesting, and in fact
fortuitous, that there are instances where the number of
branching variables actually decreases. This translates
to a smaller number of partitions and therefore a
smaller number of iterations. The details of the al-
gorithm for constructing the underestimators and over-
estimators are presented in Appendix A.
 M. Sinha et al. / Computers and Chemical Engineering 23 (1999) 1381–1394 1387

3.4. Illustrati6e example: MINLP test problem proach for the construction of linear underestimator for
this solvent design constraint. Let us introduce the
The following MINLP problem (Kocis & Grossman, following functions
1988) will help illustrate the steps of the proposed
algorithm. The problem has three binary variables and cD = % nj FDj
two continuous variables. Nonconvexitities are present j

in the two equality constraints. The three inequalities

are linear. cP = % nj (FPj )2
j

minimize f=2x1 +3x2 +1.5y1 +2y2 −0.5y3

cH = % nj (− UHj )
subject to: j

2
h1(x, y)=x + y1− 1.25 = 0.0
1
cV = Vo + % nj Vj (7)
h2(x, y)=x 1.5
2 +1.5y2 −3.00 =0.0
j

x1 + y1 51.60 The solvent design constraint (Eq. (2)) can then be

expressed as
1.33x2 +y2 5 0.0
− y1 −y2 +y3 50 4
 cD
− d*D
  2
+

cP
− d*P
 
2
+

cH
− d*H

2
5(R*)2
cV cV
cV
x ] 0, y {0, 1}3 Multiply both sides by c 2V to obtain
The equality constraints are broken into two inequali-
f 01 = 4(cD − d*DcV)2 + (
cP − d*c
p V)
2
ties. Equality hl (x, y) is decomposed to g1(x, y)and
g2(x, y). Similarly h2(x, y) is decomposed to g3(x, y)and + (
cHcV − d*HcV)2 − R 2c 2V 5 0.
g4(x, y). Here
fo is made up of four separable terms. The first and the
g1(x, y)=x 21 + y1− 1.25 5 0.0, fourth terms are squares of linear equations. The sec-
g2(x, y)=x 21 + y1− 1.25 ] 0.0 ond and the third terms are relatively more compli-
cated. Next, we will describe a tree representation of the
g3(x, y)=x 1.25
2 +1.5y2 −3.00 50.0 third term. This representation will then be used to
g4(x, y)=x 1.25 +1.5y2 −3.00 ]0.0. construct its underestimator. The third term is
2

The nonlinearities are only associated with the two g o1 = (

cHcV − d*HcV)2
terms, x 21 and x 1.5
2 , involving variables x1 and x2. For Let us introduce a function such that g 11 = (
cHcV −
this problem, the linear splitting functions are c1,= x1 d*HcV). Then we obtain g o1 ,= (g 11)2 and
and c2 =x2. The problem is solved with the initial
bounds of {05 c1 5100} and {0 5 c2 5100}. The al- 1
g 11 = [(
cH +
cV)2 − (
cH −
cV)2 − 4d*HcV].
gorithm took 39 iterations to converge. At the global 4
solution f *= 7.667, x* =[1.118, 1.310] and y*=
[0, 1, 1]. The reduction in the number of branching Next, we introduce the second level functions (g 21 and
g 22) such that g 21 =
cH +
cV and g22 =
cH −
cV
dimensions translates to smaller number of nodes vis-
and third level function such that g 31 = cH and g 32 =cV.
ited during the branch and bound operation. To verify
Note that the third level functions are linear with
this claim, the same problem was again solved with all
respect to the design variables nj. Now g 11 =(1/
the five variables as branching variables. This con-
verged to the same global solution, but after 227 itera- 4)[(g 21)2 − (g 22)2 − 4d*HcV] or g 0 = {(1/4)[(g 21)2 −(g 22)2 −
tions. This illustrative example demonstrates the 4d*HcV]}2 or g 01={(1/4)[(
cH+
cV)2−(
cH−
cV)2
computational advantage of the proposed strategy, es- − 4d*HcV]}2. The functions g 21 and g 21 are themselves
pecially when many variables participate in the nonlin- related to g 31 and g 32 by the square root. The third level
ear terms. functions in turn are linear functions with respect to the
design variables (nj ). The bounds on cH and cV are
used to construct bounds on the second level functions,
4. Construction of linear estimator for the solubility g 21, and g 22. This is further used to construct the bounds
parameter design constraint on the first level function, g 11.
The linear underestimators are constructed in a re-
The solvent performance constraint illustrated in Eq. verse sweep that starts at the first level and goes down.
(2) makes the solvent design problem nonconvex and First, the linear underestimator of g o1 is constructed in
results in multiextremality. We now present an ap- terms of g 11 such that L[g o, S 1]= m1(g 11)+ m2. The sign
1388 M. Sinha et al. / Computers and Chemical Engineering 23 (1999) 1381–1394
of m1 (which depends on the bounds on g 11, determines
whether the function m1(g 11) is convex or concave with
respect to g 21 and g 22. The underestimator now has the
following form
L[g o,S 2]=m3(g 21)+m4(g 22) +m5(CV) + m6
or
L[g o, S 2]= (m3 −m4)(
cH) + (m3 +m4)(
cV) +m5(cV) to find environmentally benign solvent substitutes for
+ m6.
The signs of (m3 −m4) and (m3 + m4) are subsequently
used to construct the underestimator with respect to cH
and cV. After rearranging the terms the linear underes-
timator is represented as
L[g o, S 2]= m8CH +m9(CV) + m10.
Thus, the underestimator will be linear with respect to
CD, CP, CH and CV; therefore, it will be linear in terms
of the search variables nj. At any iteration in a branch
and bound strategy, the subregion is represented by the
bounds on the splitting functions. Based on this infor-
mation the coefficients m1, (m13 −m4) and (m3 + m4) are
calculated and a decision about construction of the
underestimator is taken at two levels as shown in Fig.
5. Similarly, the overestimators can be constructed in a
similar fashion. Further details can be found in Ap-
pendix A.
In the next section, we apply the global optimization
algorithm and underestimator for the solubility
parameter model in a case study to design solvent
molecules in a surface cleaning application.
5. Case study: single component blanket wash design
In a lithographic printing process, ink is carried to
the impression plate by means of a train of rubber
rollers commonly called ‘blankets’. The cleanliness of
the blanket is of primary concern for production of
high-quality images. One of the most used solvents in
Fig. 5. Decision tree for construction of underestimator for the solubility parameter constraint.
the printing industry is called the ‘blanket wash’ which
is specially formulated to clean residue ink from litho-
graphic printing presses. There are more than 52 000
lithographic printers in the United States (Adrian,
1991). Many solvents used in commercial blanket
washes are to be phased out in near future due to
environmental regulations. The ‘Printing Industry of
America’ (PIA) and EPA have taken a major initiative
this application (Design for the Environment Program,
1997).
In this section we present a case study CAMD prob-
lem for to design blanket wash substitute solvents. The
detailed model development has been reported in an
earlier work (Sinha & Achenie, 1998). In short, an ideal
blanket wash solvent should have the following at-
tributes: (a) ability to rapidly dissolve dried ink; (b)
have minimal drying time; (c) should not swell the
rubber blanket; (d) be nonflammable; and (e) be envi-
ronmentally benign.
Drying time is correlated with the heat of vaporiza-
tion (DHV) of the solvent, which is minimized during
design stage. Low DHV translates to higher evaporation
rate and hence shorter drying time. It also translates to
lower utility costs associated with the drying operation.
Ink residue is assumed to consist of phenolic resin
(Super Bakacite® 1001, (Cunningham, 1998)) with dd =
23.3, dp = 6.6 and dh = 8.3 MPa1/2 and R* =19.8
MPa1/2. Solvents that can effectively dissolve the ink
residue have the following solute–solvent interaction
constraint
R ij = (4(idD − 23.3)2 + (idP − 6.6)2 + (idH − 8.3)2)1/2
B 19.8
A lithographic blanket is generally made of polyiso-
prene rubber. Natural rubber (polyisoprene) has 25%
swelling range between polar solubility parameter (dp)
values 0 − 6.3 MPa1/2 (i.e. a solvent that does not have
dp in this range will not significantly swell the blanket).
To ensure minimal blanket swelling we add a constraint
 M. Sinha et al. / Computers and Chemical Engineering 23 (1999) 1381–1394 1389

on dp to be greater than 6.3. Thus, the ability to % % Uij (2−6j )= 2 (Octet rule)
i j
dissolve dried ink but at the same time not swell the
rubber blanket is ensured by constraining the solvent
exp((% % Uij tmj )/102.425) 5223
design to a specified region in the solubility parameter i j
space as shown in Fig. 3.
(Melting point constraint)
We note that the correlations for both log BCF and
log LC50 depend linearly on log Kow. Therefore con-
exp((% % Uij tbj )/204.4) ] 323
straining log Kow is sufficient in the sense that we need i j
not explicitly constrain the log BCF and the log LC50.
(Boiling point constraint)
In a nutshell, a solvent’s environmental impact is in-
ferred from its Kow value. Here we constrain the log Kow
% % Uij (x o)j + % % Uij (x o)j 5 4.0
of the solvent to a small value (i.e. less than 4.0). To i j i j
ensure that the solvent is a liquid at ambient tempera-

'
(Octanol-water partition constraint)
ture, the boiling point (Tb) should be larger than 323 K
2
and the melting point (Tm) should be less than − 223 Á % % U*F Â Á % % U*(F )2
à ij Dj à à ij Pj
K. The presence or absence of structural group in a i j i j
4Ã − 23.3Ã + Ã
molecule is represented by a binary matrix U (Churi &
ÃV o + % % Uij Vj à à V o + % % Uij Vj
Ä Å Ä
! '
Achenie, 1996) with elements Uij such that i j i j
2 2
1 Â Á % % U*(−U Â
Uij = Ã Ã ij Hj ) Ã

'
i j
− 6.6Ã + Ã − 8.3Ã 5 (19.8)2
if ith position in a molecule has jth structural gro à à Vo + % % Uij Vj Ã
Å Ä Å
up i j
0 otherwise

  ' 
(Solubility parameter constraint)
Note that nj =i Uij. The size of the molecule is limited
to a maximum of five structural groups i j Uij 5 5 % % Uij (FPj )2
i j
] 6.3 (Swelling constraint)
in this case study. The octet rule for acyclic molecules
as proposed by Odele et al. is used (Odele & Machietto, Vo + % % Uij Vj
i j
1993).
cD 5 % % U*F
ij Dj 5 cD (Splitting function 1)
% % Uij (2−6j )=2 i j
i j

where 6j is the valency of jth structural group. cP 5 % % U*F

ij Pj 5 cP (Splitting function 2)
i j
The basis set consists of twelve groups: CH3-, CH2-,
Ar-, Ar= , -OH, CH3CO-, -CH2CO-, -COOH, cH 5 % % U*(−U
ij Hj )5 cH (Splitting function 3)
CH3COO-, -CH2COO-, -CH3O, -CH2O-. Ar- refers to i j

benzene ring with one less hydrogen (valency of one)

and Ar = refers to a benzene ring with two less hydro-
gen atoms (valency of two). Many chlorinated com-
pounds (e.g. 1,1,1, trichloroethane) are being phased
out for their ozone depleting potential (ODP). We have
addressed this by not including any structural group
that has chlorine, in our basis set. This basis set can be
used to design straight chain hydrocarbons, aromatics,
alcohols, ketones, esters, ethers and acids. Heat of
vaporization, boiling point and melting point are calcu-
lated by the Constantinou and Gani (1994) method.
The resulting CAMD formulation for blanket wash
solvent design is shown below.
cV 5 Vo + % % Uij Vj 5 cV (Splitting function 4)
i j
where h6j, tbj and tmj are group contribution parameters
associated with DHv, Tb, and Tm. Solubility parameter
and swelling constraints are nonlinear and nonconvex.
The splitting functions are cD, cP, cH and cV. The
subregion is thus represented by four pairs of addi-
tional linear constraints that confine the search between
the maximum and minimum values of these splitting
functions.
5.1. Results and discussion

minimize % % Uij hVj (P3) The problem has 60 binary variables and 15 con-
Uij i j straints. The algorithm took 31 iterations. Since maxi-
subject to: mums of two MILPs are solved per iteration, the total
number of MILPs solved is no more than 62. The
% % Uij 55 (Molecular size constraint) globally optimal solution found is methyl-ethyl ketone
i j (MEK) with molecular structure CH3-CH2-CO-CH3.
1390 M. Sinha et al. / Computers and Chemical Engineering 23 (1999) 1381–1394

Table 1
Result of solvent design case study, optimal molecules and their properties

Methyl ethyl ketone (MEK) Diethyl ketone (DEK) Ethylene glycol mono methyl ether Comments

Structure

DH6 (KJ/mol) 35.47 40.12 47.6 Computed

(objective function) 34.5 38.37 – Experimental
TB (°C) 81 112.5 114.2 Computed
79.6 101.0 124.6 Experimental
TM (°C) −80.0 −66.4 −68.9 Computed
−86.0 −39.0 −85.1 Experimental
dD (Mpa1/2) 15.13 15.33 15.25 Computed
16.0 15.8 – Experimental
dP (Mpa1/2) 8.64 7.34 7.76 Computed
90 7.6 – Experimental
dH (Mpa1/2) 4.65 4.28 17.08 Computed
5.1 4.7 – Experimental
R ij 16.86 8.9 11.9 R a =19.2
log Kow 1.59 2.37 −0.65 Computed
0.29 0.91 −0.77 Experimental
log BCF 0.97 1.556 −0.724 Computed
0.0 0.46 −6.152 Computedb
0.0 0.46 −6.8 Experimental
Tf (°C) −16.4 4.9 6.2 Computed
−9.0 13.0 43.0 Experimental
Evaporation ratec 3.8 2.3 0.53 Experimental
Solvent type Ketone Ketone Glycol ether

a
Computed via correlation from experimental log Kow value.
b
Computed via correlation from predicted log Kow value.
c
Experimental value of the evaporation rates relative to evaporation rate of butyl actate. Note that this value has a good correlation with the
objective function (DHv).

The DHv (also the objective function) associated with
MEK is predicted as 35.471 KJ/ mol. This molecule
was found at the 10th iteration with a overall upper
bound (UPBD) of 35.471 and an overall lower bound
(LWBD) of 33.99. The algorithm finally converged with
MEK as the global solution after 21 more iterations.
To find the second best molecule, additional constraint
was added to eliminate MEK from the design space.
The next best molecule designed is diethyl ketone
(DEK). This process was repeated to find the third best
solvent namely ethylene glycol monomethyl ether
(EGME). The results are presented in Table 1. This
table also lists the predicted properties along with the
experimental values. The model error associated with
DHv, Tb, Tm and solubility parameters is relatively low.
The predicted value of log Kow shows a considerable
model error and consistently predicts a larger value
compared to the experimental value. This error also
results in an error in log BCF. However, it is interesting
to note that if the experimental value of log Kow is used
in Eq. (4), then the model prediction for log BCF is
very accurate for the three molecules designed. Table 1
also lists the evaporation rate of each solvent. There is
an inverse relationship between DHv and the evapora-
tion rate. Hence in effect, minimization of DHv (objec-
tive function) results in the maximization of the
evaporation rate. Thus the molecules designed have
high evaporation rate and consequently will have mini-
mal drying time.
6. Conclusions
This paper presents a framework that incorporates
environmental and health considerations in solvent de-
sign. Solvent design models based on solubility parame-
ters are nonlinear and often result in a nonconvex
CAMD formulation. We note that many nonlinear
correlations for CAMD have a structure similar to that
of solubility parameter models and are a source of
nonconvexity. To solve these CAMD problems, we
have developed a global optimization algorithm that
exploits the problem structure. We have also presented
a strategy for the construction of linear underestimators
for nonlinear models often seen in solvent design. The
underestimators are used in the reduced space branch
 M. Sinha et al. / Computers and Chemical Engineering 23 (1999) 1381–1394
and bound strategy to solve the case study solvent
design problem. We have introduced the idea of split-
ting functions that result in a smaller number of branch-
ing nodes. For example in the solvent design case study,
instead of branching on all 60 binary variables (that
participate in nonlinear terms), we only need to branch
on four splitting functions.
We are continuing to address other issues. In this
regard, more complete models for molecular structure
that account for connectivity information (i.e. how
different structures are connected to each other, Churi
& Achenie, 1996) are being considered. In addition
other property requirements on viscosity, surface ten-
sion, ozone depletion potential, etc. are being
considered.
In the proposed global optimization algorithm we
only need to solve a sequence of MILPs. Therefore we
are investigating methods for solving MILPs more effi-
ciently. It is interesting to note that, in our limited
experience with the proposed algorithm, the computa-
tional effort, measured in terms of the number of
iterations, grows linearly as the problem size increases.
However, as the problem size increases Ip – solve is not
very robust. As a result, we are currently replacing
Ip – solve by more computationally efficient MILP
solvers such as OSL. This modification is not expected
to reduce the number of iterations. Instead, it is ex-
pected to reduce the computational effort per iteration.
Fig. 6. A multilevel representation of a nonlinear function as a tree
function.
1391
In this paper, only pure component product design is
considered. Mixture design problems are often posed as
MINLP problems. This algorithm can be used for
solving nonlinear and nonconvex mixture design prob-
lem as well. Incorporation of model uncertainties is also
being considered.
Appendix A. Multilevel approach to construction of
underestimator
McCormick (1976) suggested a method for construc-
tion of convex underestimators and concave overesti-
mators for functions that are the product of two
univariate functions, f(x)g(y). Extension for products
of univariate functions was developed by Maranas and
Floudas (1995). In this paper, we discuss another ap-
proach for construction of underestimator and
overestimator.
The first step involves representation of a complex
function in terms of simpler one-dimension functions.
The only arithmetic operations relating these functions
to the original function are addition, subtraction,
square, and square root. Such a representation has a
tree structure as shown in Fig. 6. The original function
represented as 8 o1, corresponds to the root node (zeroth
level). The child nodes emanating from the root node
correspond to the first level functions, represented as
8 (1)
i (i
th
function at the first level). The next level corre-
sponds to the functions at the second level (8 (2) i ) and so
on. As shown in Fig. 6, the number of nodes at kth
level is nk. The functions at each level are repeatedly
decomposed till the functions at the lowest level child
nodes are linear function of the search variables. Hence
this multilevel scheme attempts to represent the original
function as a tree function. A mathematical representa-
tion of tree function is shown in Eq. (A1).
nk + 1
8 ki = % mj (8 kj + 1)tj (A1)
j=1
where mj is the coefficient for the jth child function
8 (K + 1), and there are m such child functions. We
consider the case when the power tj for the jth function
takes the value 0.5, 1, or 2. For a linear group contribu-
tion model, at the lowest level in the tree structure, mj
corresponds to the group contribution parameter uj and
tj is 1.
We can obtain linear estimators of the kth level
function 8 (k) in terms of the function at the level
(k+ 1), 8 (ki
+ 1)
. There are standard techniques for con-
structing linear underestimators for separable functions
that can be represented as Eq. (A1). Only the functions
for which tj equals 2 (square) or 0.5 (square root) need
to be linearized. Depending on the sign of the coeffi-
cient, mj corresponding to the jth function at the (k +
1)th level, the function 8 (k)
i can be concave or convex
1392 M. Sinha et al. / Computers and Chemical Engineering 23 (1999) 1381–1394
Fig. 7. A multilevel representation of the function f o1 =x0x1x2x3, as a tree function.
with respect to the functions, 8 (k + 1)
. For concave
i f 11 = ( f 21)2 + ( f 22)2 + ( f 23)2 + ( f 24)2
functions the linear underestimator for region S ki has
the form
k k
[ f(8 ) − f(8 )] k
i i
L[ f(8 ki );S ki ]= f(8 ki )+ (8 i −8 ki ) (A2)
8 ki −8 ki
where S ki = {c ki : c ki 5c ki 5c ki }. c( and c(( correspond
to the minimum and maximum values for ci.
For convex functions, the linear underestimator in
region S ki has the form
L[ f(8 ki );S k]= f %(8̂ ki )(8 ki −8̂ ki ) + f(8̂ ki )
where 8 ki is the mid point of 8 ki in the region S k, f %(8 ki )
is the derivative of function f(8 ki ) at the midpoint.
If the bounds on the search variables are known,
then the bounds or range on 8 oi , 8 1i , 8 2i … 8 N
i can be
computed. This results in a bound propagation at
different levels of the functions beginning from the
lowest level. For example, let us consider the following
function
f 01 = x0x1x2x3
Let us introduce two functions f 11 =x0x1 +x2x3 and
f 12 = x0x1 −x2x3. Now, using the following identity
1 1
ab= (a+b)2 − (a−b)2
4 4
we can represent the original function as f 01 = 1/
4( f 11)2 −1/4( f 12)2. At this level let us introduce the
functions corresponding to the second level f 21 = x0 +
x1,f 22 =x0 − x1, f 23 = x2 +x3 and f 24 =x2 −x3. Using
these second level functions, we can represent the first
level functions f 11 and f 12 as
1 1 1 1
4 4 4 4
and
1 1 1 1
f 12 = ( f 21)2 + ( f 22)2 − ( f 23)2 − ( f 24)2.
4 4 4 4
Here in three levels (see Fig. 7) we have reduced the
original function and represented it as a superposition
of univariate functions. The only operations are plus,
minus and the square. Given the bounds on x0, x1, x2
(A3) and x3, bounds on the second level functions f 21, f 22, f 23,
and f 24 can be computed via interval arithmetic opera-
tors (Moore, 1966). Similarly, the bounds on the first
level functions f 11 and f 12 can be computed. Thus the
estimation of bounds in the tree function starts from
the bottom level and goes up. Care must be taken to
compute the true range, as the interval operations may
result in overestimation of the bounds.
The underestimators are constructed in the reverse
order. First the original function ( f0) is linearized with
respect to the set of first level function (here f 11 end f 12).
Next, the resulting linear expression of f 11 and f 12 is
considered. Underestimators are constructed for each
term of this expression with respect to the second level
functions f 21, f 22, f 23, and f 24. At this stage the expression
for the underestimator is a linear function of f 21, f 22, f 23
and f 24. In this way we obtain the linear relations which
connect linear underestimators of the kth level with the
linear underestimator for the (k+ 1)th level (k=
0, 1, … , N−1). These second level functions are them-
selves linear functions of the original variable x0 x1, x2,
and x3. At the next stage, we can obtain underestimator
of all the levels as linear functions of x0 x1, x2, and x3.
 M. Sinha et al. / Computers and Chemical Engineering 23 (1999) 1381–1394
In turn we have computed the linear underestimator of
the original function f 01. This operation must be per-
formed from the bottom level to the top level.
In a similar fashion linear overestimators can be
constructed. For this only Eqs. (A2) and (A3) has to be
modified. If the function is concave then the right hand
side of Eq. (A3) describes the functional form of the
overestimator. If the function is convex then the right
hand side of Eq. (A2) describes the functional form of
the overestimator. The remaining steps are same as
those for construction of underestimators. This
methodology has also been demonstrated via construc-
tion of linear underestimator for the solubility parame-
ter constraint in Section 4.
In contrast to the McCormick’s approach, in our
approach the dimensionality of the lower and upper
bound problem is no larger than the dimensionality of
the original problem PORG. However, there are in-
stances where the number of branching variables actu-
ally decreases. This translates to smaller number of
partitions and therefore a smaller number of iterations.
References
Adrian, J. R. (1991). Managing solvents and wipers. Pollution Pre-
6ention Re6iew, 419 – 425.
Archer, W. L. (1996). Industrial sol6ent handbook. Marcel Dekker.
Brignole, E. A., Bottini, S., & Gani, R. (1986). A strategy for the
design of solvents for separation processes. Fluid Phase Equilibria,
29, 125 – 132.
Butler, R. M., Cooke, G. M., Lukk, G. G., & Jameson, B. G. (1956).
Prediction of flash points of middle distillates. Industrial and
Engineering Chemistry, 48, 808–812.
Churi, N., & Achenie, L. E. K. (1996). Novel mathematical program-
ming model for computer aided molecular design. Industrial and
Engineering Chemistry Research, 35 (10), 3788–3794.
Constantinou, L., & Gani, R. (1994). New group contribution
method for estimating properties of pure compounds. American
Institute of Chemical Engineers Journal, 40, 1697–1710.
Constantinou, L. K., Bagherpour, R., Gani, R., Klein, J. A., & Glen,
R. C. (1996). Computer-aided product design: problem formula-
tions, methodology, and applications. Computers & Chemical
Engineering, 20 (6), 685.
Cunningham, H. W. (1998). Typical ink composition, WCC Inc.
Personal communication.
Design for the Environment Program. (1997). Cleaner technologies
substitute assessment: lithographic blanket washes. EPA 744-R-
97-006, USEPA.
Duvedi, A. P., & Achenie, L. E. K. (1996). Designing environmen-
tally safe refrigerants using mathematical programming. Chemical
Engineering Science, 51, 3727–3739.
Fredenslund, A., Gmehling, J., & Rasmussen, P. (1977). Vapor– liquid
equilibria using unifac. Amsterdam: Elsevier.
Gani, R., & Brignole, E. A. (1983). Molecular design of solvents for
liquid extraction based on UNIFAC. Fluid Phase Equilibria, 13,
331.
Gani, R., Nielsen, B., & Fredenslund, A. (1991). A group contribu-
tion approach to computer-aided molecular design. American
Institute of Chemical Engineers Journal, 37, 1318–1331.
Gessner, A. W. (1998). Switch materials to prevent pollution. Chemi-
cal Engineering Progress, December, 59–64.
1393
Grulke, E. A. (1989). Solubility parameter values. In J. Brandrup, &
E. H. Immergut, Polymer handbook (pp. 519 – 559). John Wiley &
Sons.
Hansch, C., & Leo, A. J. (1979). Substituent constants for correlation
analysis in chemistry and biology. New York: John Wiley.
Hansen, C. (1969) The universality of the solubility parameter, Indus-
trial Engineering Chemistry Product Research and De6elopment,
2 – 11.
Hansen, C. M., & Beerbower, A. (1971). Solubility parameters. In A.
Standen, Kirk-Othmer encyclopedia of chemical technology. New
York: Interscience.
Horst, R., & Tuy, H. (1990). Global optimization: deterministic ap-
proaches. Heidelberg: Springer-Verlag.
Horvath, A. L. (1992). Molecular design. Elsevier.
IBM. (1992). Optimization subroutine library: guide and reference.
Kingston, NY.
Joback, K. G., & Stephanopoulos, G. (1989). Designing molecules
possessing desired physical property values, Proceedings FO-
CAPD 1989, Snowmass Village, Colorado.
Klein, J. A., Wu, D. T., & Gani, R. (1992). Computer aided mixture
design with specified property constraints, Computers & Chemical
Engineering, 16 (Suppl. S229 – S236).
Kocis, G. R., & Grossman, I. E. (1988). Global optimization of
nonconvex mixed-integer nonlinear programming (MINLP) prob-
lems in process synthesis. Industrial and Engineering Chemistry
Research, 27, 1407 – 1421.
Konemann, H. (1981). Quantitative structure – activity relationships
in fish toxicity studies. 1. Relationship for 50 industrial chemicals.
Toxicology, 19, 209 – 221.
Krishner, E. M. (1995). Environment, health concerns force shift in
use of organic solvents, Chemical and Engineering News, June,
13 – 20.
Lyman, W. J., Reehl, W. F., & Rosenblatt, D. H. (1981). Handbook
of chemical property estimation methods. McGraw-Hill.
Macchietto, S., Odele, O., & Omatsone, O. (1990). Design of optimal
solvents for liquid – liquid extraction and gas absorption processes.
Transactions of the Institute of Chemical Engineers, 68, 429.
Maranas, C. D., & Floudas, C. A. (1995). Finding all solutions of
nonlinearly constrained system of equations. Journal of Global
Optimization, 7, 143 – 182.
Maranas, C. D. (1996). Optimal computer-aided molecular design: a
polymer design case study. Industrial and Engineering Chemistry
Research, 35, 3403.
Marcoulaki, E. C., & Kokosis, A. C. (1998). Molecular design
synthesis using stochastic optimisation as a tool for scoping and
screening. Computers & Chemical Engineering, 22, S11 –S18.
Mavrovouniotis, M. L. (1998). Design of chemical compounds. Com-
puters & Chemical Engineering, 22 (6), 713 – 715.
McCormick, G. P. (1976). Computability of global solutions to
factorable nonconvex programs. Part I — convex underestimat-
ing problems. Mathematical Programming, 10, 147 – 175.
Moore, R. E. (1966). Inter6al analysis. Englewood Cliffs, New Jersey:
Prentice-Hall.
Nendza, M., & Hermens, J. (1995). Properties of chemicals and
estimation methodologies. In C. J. van Leeuwen, & J. L. M.
Hermens, Risk assessment of chemicals: an introduction (pp. 239–
292). Dordrecht, The Netherlands: Kluwer Academic Press.
Odele, O., & Machietto, S. (1993). Computer aided molecular design:
a novel method for optimal solvent selection. Fluid Phase Equi-
libria, 82, 47 – 54.
Pistikopoulos, E. N., & Stefanis, S. K. (1998). Optimal solvent design
for environmental impact minimization. Computers & Chemical
Engineering, 22 (6), 717 – 733.
Pretel, E. J., Lopez, P. A., Bottini, S. B., & Brignole, E. A. (1994).
Computer-aided molecular design of solvents for separation pro-
cesses. American Institute of Chemical Engineers Journal, 40 (8),
1349 – 1360.
1394 M. Sinha et al. / Computers and Chemical Engineering 23 (1999) 1381–1394
Reid, R. C., Prausnitz, J. M., & Poling, B. E. (1987). The properties
of gases and liquids. New York: McGraw Hill.
Schwab, H. (1997). Documentation of lp – sol6e. Heidelberg.
Sinha, M., & Achenie, L. E. K. (1998). Computer aided solvent
design for lithographic blanket wash system. Topical Conference
on Pollution Pre6ention and Risk Reduction, Annual American
Institute of Chemical Engineers Meeting, Paper 13e, Miami, FL.
Vaidyanathan, R., & El-Halwagi, M. M. (1994a). Computer-aided
design of high performance polymers. Journal of Elastom Plastics,
26 (3), 277.
Vaidyanathan, R., & El-Halwagi, M. M. (1994b). Global optimiza-
tion of nonconvex nonlinear programs via interval analysis. Com-
puters & Chemical Engineering, 18 (10), 889–897.
Vaidyanathan, R., & El-Halwagi, M. M. (1996). Computer-aided
synthesis of polymers and blends with target properties. Industrial
Engineering Chemistry Research, 35 (2), 627–634.
Vaidyanathan, R., Gowayed, Y., & El-Halwagi, M. (1998). Com-
puter-aided design of fiber reinforced polymer composite prod-
ucts. Computers & Chemical Engineering, 22 (6), 801–808.
van Krevelen, D. W., & Hoftyzer, P. J. (1976). Properties of polymers:
their estimation and correlation with chemical structure. Amster-
dam: Elsevier.
van Leeuwen, C. J., & Hermens, J. L. M. (1995). Risk assessment of
chemicals: an introduction. Kluwer Academic Publishers.
Veith, G. D., & Konasewich D. E. (1975). Structure – activity correla-
tions in studies of toxicity and bioconcentration wiht aquatic
organisms. Proceedings of a Symposium held in Burlington, On-
tario, at the Canada Center for Inland Water, Ontario, Canada.
Venkatasubramanium, V., Chan, K., & Caruthers, J. M. (1994).
Computer-aided molecular design using genetic algorithms. Com-
puters & Chemical Engineering, 18, 833 – 844.
Viswanathan, J., & Grossman, I. E. (1990). A combined penalty
function and outer-approximation method for MINLP optimiza-
tion. Computers & Chemical Engineering, 14 (7), 769 – 782.
Zhao, R., & Cabezas, H. (1988). Molecular themodynamics in the
design of substitute solvents. Industrial and Engineering Chemistry
Research, 37, 3268 – 3280.