Fluid Phase Equilibria, 38 (1987) 245-272 245

Elsevier Science Publishers B.V., Amsterdam - Printed in The Netherlands

PHASE BEHAVIOR OF AQUEOUS TWO-POLYMER SYSTEMS

C.H. KANG and .%I. SANDLER

Department of Chemical Engineering, University of Delaware, Newark, DE 19716 (U.S.A.)

(Received February 17, 1987; accepted in final form June 1, 1987)

ABSTRACT

Kang, C.H. and Sandler, S.I., 1987. Phase behavior of aqueous two-polymer systems. Fluid
Phase Equilibria, 38: 245-272.

In this work we consider the thermodynamic description of the liquid-liquid phase

behavior of dextran-polyethylene glycol-water systems which are of interest for biological
separations. In this effort, we use the Flory-Huggins and UNIQUAC models, and develop a
new numerical procedure to estimate the interaction parameters.
Also, when using interaction parameters for the Flory-Huggins model obtained from
osmotic pressure and plait point data, qualitatively reasonable binodal curves were obtained.
Due to the asymmetry in the experimental binodal curves for these polymer systems, there is
a small difference between the calculated interaction parameters for each of the polymers
with water. The location of a binodal curve is found to be sensitively dependent on the
interactions between the unlike polymers.
Multiple sets of interaction parameters were found for the UNIQUAC model which could
describe liquid-liquid phase splitting in a reasonable manner. In particular, an intercorrela-
tion was found in which any set of parameters in a broad range could predict the phase
behavior equally well. Also, the volume and surface area parameters of the polymers were
found to play an important role in determining phase behavior of these polymer systems. The
shift of the binodal curve with molecular weight of the polymers was also investigated.

INTRODUCTION

On adding small amounts of two water-soluble, but incompatible poly-

mers, such as dextran and polyethylene glycol to water, two water-rich
phases may be formed. As is well known, these aqueous two-phase polymer
systems are suitable for the purification or isolation of biological materials
which would be denatured by the thermal or chemical environments encoun-
tered in typical chemical engineering separation techniques (Albertsson,
1971).
Although these aqueous two-phase systems have been widely used in the
separation of biochemical materials, there has been little effort devoted to
the thermodynamic description of the phase equilibrium. Such a description

0378-3812/87/$03.50 0 1987 Elsevier Science Publishers B.V.

246

is important as it could provide the basis for predicting phase behavior in

cases where experimental data are not available. The classical Flory-Hug-
gins model has been used in the determination of the polymer-polymer
interaction parameters in an organic solvent (Allen et al., 1960; Robledo-
Muniz et al., 1985). Recently, the model has also been used to interpret the
phase separation of an aqueous polymer solution containing dextran and
polyethylene glycol (PEG) (Gustafsson et al., 1986). A simple model for the
dextran-PEG systems was derived by Ogston (1962), and this model was
used to predict the phase behavior (Edmond and Ogston, 1968). However,
Ogston’s expression was derived based on an assumption of dilute solutions,
so its utility for the phase behavior of the water-dextran-PEG system is
unproven (Gustafsson et al., 1986).
Once the physical properties of constituent molecules, temperature and
pressure of the system are given, the prediction of phase behavior is
determined by the thermodynamic model used and the values of the adjusta-
ble model parameters. For example, in the Flory-Huggins model, the
location of the critical point of a ternary system depends on the interaction
parameters between unlike polymer segments (Scott, 1949).
This paper is primarily concerned with the predictability of the
Flory-Huggins (F-H) and the UNIQUAC models for the aqueous two-phase
polymer systems containing dextran and PEG. Further, a new objective
function is proposed for the estimation of thermodynamic model parame-
ters, and a method for the calculation of the binodal curve is discussed. We
also examine the accuracy of predictions when the F-H parameters are
extracted from intrinsic viscosity and plait points reported in literature. In
addition, the effects of molecular weight and the shape of the polymers on
the binodal curve are examined using the IJNIQUAC model.

THERMODYNAMICS OF POLYMER SOLUTIONS

In the analysis here we consider each of the polymers to be of a single

weight; the case of polydispersivity of the molecular weight can be treated
using the mathematical procedures described elsewhere (Shibata et al.,
1987). The various molecular weights and intrinsic viscosities of dextran and
PEG of interest are shown in Table 1.
The necessary and sufficient condition for equilibrium in a liquid at given
temperature and pressure is that the molar Gibbs energy is a minimum. For
the systems that split into two phases, I and II, the molar Gibbs energy of
mixture is given by

(n’ + #)G = n’G’(n;, M;,. ._> + n”G”(n;‘, n’:,. ..) (1)

where n: and nl’ are the moles of component i in phases I and II,
 247

TABLE 1
Molecular weights and intrinsic viscosities of polymers from Albertsson (1971)

Polymer M, x103 ii?, x lo4 qa (cm3 g-l)

D17 23.0 3.0 16.8

D24 40.5 24.0
D37 83.0 17.9 37.0
D48 180.0 46.0 48.0
D68 280.0 220.0 68.0
PEG4000 3.0 - 3.7
PEG6000 6.0 - 7.5 20.4 ’

a In water at 293 K.
b From Edmond and Ogston (1968).

respectively, and G’ and G” are the molar Gibbs energies of mixtures

corresponding to .I moles of phase I and n rr moles of phase II. Since the
molar Gibbs energy is a minimum, a differential change of compositions at
equilibrium does not produce a change in the molar Gibbs energy provided
the temperature and pressure remain constant; that is

d(G) T,p = 0 (2)

However, eqn. (2) is also satisfied at maximum and saddle points, so it is
impossible to differentiate between these without a further inspection of the
molar Gibbs energy surface. The common equilibrium condition derived
from eqn. (2) is

ApLf=Ap:‘, i=l,2 ,..., n (3)

Equation (3) can be written in terms of the activities a, or the activity
coefficient y,. The interrelation between the excess Gibbs energy Gr- and the
activity coefficients y, is

l-1
3GE
lxTInyi= arli T,p,n,,,
(44

ai = yixi w4
and
ApL,= RT In ai (44
where xi is the mole fraction of component i. Equation (4~) is widely used
for polymer solutions instead of the activities and activity coefficients
because the activities of polymers are so small.
Under the assumption that each solvent molecule or segment of a poly-
mer chain occupies only one lattice site, and that polymers are distributed
248

uniformly throughout the system, the combinatorial entropy change of

mixing can be estimated. The excess Gibbs energy can then be evaluated by
introducing van Laar type interaction terms for the enthalpic contribution.
In this way, the F-H model for the Gibbs energy of mixing for an
n-component system is (Tompa, 1956)

(5)
where #i, ai and x ij refer to the volume fraction, number of moles of i, and
the interaction parameters between i and j components, respectively, and
mj is the ratio of the molar volume to a reference volume, here taken to be
that of the solvent, component 3 (thus m3 = 1). Since the number of sites
occupied by a polymer molecule is generally assumed to be proportional to
its volume, the site fractions become volume fractions when the polymer
molecules are divided into segments equal in volume to a solvent molecule.
Since the combinatorial part of eqn. (5) is approximate, the xii terms, in
fact, account for both entropic and enthalpic contributions (Flory, 1971).
The chemical potential of each component is obtained by differentiation
of the excess Gibbs energy of mixing with respect to the number of moles of
that species, ni, as given by Tompa (1956)
AFL- +j
& =ln It/i + I- mjC *22 + miCXij+, - miC C Xjk+jJ/k (6)
J J _i j k>j

From this equation it is not possible to obtain analytical expression for the
binodal curve, though the spinodal can be obtained analytically (Scott,
1949).
Using the local composition concept, Abrams and Prausnitz (1975) de-
rived the UNIQUAC model for the excess Gibbs energy. Using weight
fractions rather than mole fractions (because the mole fractions of polymers
are very small) the UNIQUAC model is (Oishi and Prausnitz, 1978)

where
 249

and

Here aj denotes the activity of j component, Y,’ and qJ’ are the volume
and surface area parameters of the j constituent per unit mass, $J and 0,l
are volume and surface area fractions and x, is the weight fraction of
species j. Also, li and r$ are given by

where

q’.k- hk
Ajk = R (7f)
where qj is the interaction energy between i and j molecules or segments
of polymer chains and z is the coordination number, which is usually set
equal to 10. Since the combinatorial contribution to the excess Gibbs energy
is applicable to mixtures containing very large molecules as well as mole-
cules of similar size, UNIQUAC may be used to represent the properties of
polymer solutions (Abrams and Prausnitz, 1975).
To use the two models discussed above to predict the phase behavior of
polymer solutions, we must determine the physical properties of the pure
components. In the F-H model, the specific volumes of components are
required to calculate the volume fractions, while volume and surface area
parameters, r’ and q’, are needed in the UNIQUAC model. For simplicity,
the polymers of interest are considered to be linear with the same repeat
units. For instance, dextran D37 (En = 8.3 X 104) is assumed to differ from
D48 (M, = 1.8 X 105) only in the number of the repeat units corresponding
to the difference of their molecular weights.
For determination of specific volumes of polymers, density data of
aqueous solutions of D48 and PEG6000 at 293 K were taken from Alberts-
son (1971). Using the Redlich-Kister expansion (Sandler, 1977) for the
correlation of the solution densities, specific volumes of 0.6428 and 0.80982
cm3 g-’ were obtained for dextran and PEG, respectively, when only the
first cross-product term was retained. For comparison, the group contribu-
tion tabulations of Schroeder and of Le Bas (Reid et al., 1977) yielded 0.663
250

TABLE 2
r’ and q’ parameters for components estimated by various correlations for polymers

Component r’ X10’ q’Xl0’

Dextran 2.717 a (2.94 b, 1.96 a (1.99 b,
PEG 3.87 = (3.62 b, 2.57 ’ (3.0 b,
Water 5.11 c 7.78 ’

a and b are based on correlations of Brelvi (1982) and Abrams and Prausnitz (1975),
respectively.
’ From Gmehling et al. (1982). The polymers are assumed linear as discussed in the text.

and 0.8095 cm3 g-l, respectively, for dextran, while 0.6 (Edmond and
Ogston, 1968) and 0.611 cm3 g-’ (Bamberger et al., 1985) were reported in
the literature. For PEG, the group contribution method (Reid et al., 1977)
gives 0.888 cm3 g-’ while a value of 0.837 cm3 g-’ has been reported in the
literature (Edmond and Ogston, 1968). In this study, we used 0.626 and
0.832 cm3 g-’ for the specific volumes of dextran and PEG, irrespective of
molecular weights, which were obtained by correlations’of experimental data
of D48-water and PEG6000-water systems (Albertsson, 1971) assuming
additive mixture densities.
It is much more difficult to estimate the r’ and q’ parameters in the
UNIQUAC model. For PEG and water, those parameters are readily
available from the literature (Gmehling et al., 1982), while the parameters of
dextran had to be determined from the correlation of these parameters with
other physical properties information. A correlation of these parameters for
simple compounds has been proposed by Brelvi (1982) based on the critical
volume of the compound. The critical volume of dextran was estimated by
the group contribution method of Lydersen (Reid et al., 1977). Although the
correlation coefficients for the polymer are not given in his work, our
estimate was made using parameters for similar compounds. We have also
used the correlation method of Abrams and Prausnitz (1975) based on van
der Waals volumes and surface areas with values recommended by Bondi
(1968). The group parameters r’ and q’ so evaluated are normalized to the
-CH,- unit as proposed by Abrams and Prausnitz (1975). The parameters
thus determined are compiled in Table 2.

PARAMETER ESTIMATION PROCEDURE

For ternary systems, the F-H model contains three adjustable parame-
ters, xi,, which characterize the mutual interactions between molecules. In
the UNIQUAC model for ternary systems, six interaction parameters be-
tween molecules or segments are adjustable. In all cases we consider the
 251

interaction parameters to be purely adjustable, and their values are obtained

by fitting the models to experimental data.
The model parameters can be estimated by minimizing an appropriate
objective function based on the errors in the predictions (Sorensen, 1980).
For multiple data sets, an objective function based on the equivalence of
chemical potentials for each component in both phases can be written as the
sum of squared differences in chemical potentials over all data sets

However, this objective function has two disadvantages. The equivalence

condition of chemical potentials is derived from the necessary condition that
the molar Gibbs energy be a minimum at equilibrium, and its surface is
stationary at this point. Therefore, unstable or inflection points which also
satisfy the equivalence condition could be obtained during the minimization
of the objective function. The other disadvantage is that the chemical
potentials are not directly proportional to the compositions. Thus, the
minimization of the chemical potential differences does not directly imply
the minimization of the differences between the predicted and experimental
compositions, which is desired. Therefore the estimated parameters using the
objective function of eqn. (8) may result in a phase envelope which is not
optimum for composition correlations.
Based on the maximum likelihood principle (Anderson et al., 19X3), the
objective function in terms of compositions has been used for the estimation
of the interaction parameters (Varhegyi and Eon, 1977; Sorenson, 1980). In
the maximum-likelihood method the parameters are chosen to make the
experimental observations appear to be most likely when taken as a whole.
Assuming that the experimental error is distributed randomly, the following
objective function has been used

F,(P)=ccc (94
k i j
X22k = %2k (9b)
or

MP)=c rnczC (94

k i j

where xijk is the weight fraction of i component in phase j on the k th tie

line, and zZ~~,is the predicted composition with current parameters in the
minimization procedure.
252

The objective function with one fixed composition chosen equal to the
experimental data has been criticized because in true least-squares or maxi-
mum likelihood approximations the fixed composition must also be consid-
ered to contain error (Varhegyi and Eon, 1977). In fact, the choice of a
different fixed composition and/or a different phase leads to results of
different accuracy (Simonetty et al., 1982). In using the objective function of
eqn. (SC) tie lines were developed, and then an interpolation procedure was
used between the closest tie lines bracketing experimental data (Varhegyi
and Eon, 1977; Sorensen, 1980). However, if the calculated and predicted tie
lines are very different, the interpolation of the predicted tie lines may be
inaccurate and can cause problems in calculating the residue of the objective
function in terms of compositions.
To overcome the violation of the maximum likelihood principle and avoid
the difficulty in using eqn. (SC) mentioned above, we have developed a new
objective function by replacing eqn. (9b) with

xpk= a,; (94

where the superscript F indicates the overall composition of the tie line. The
procedure used is illustrated in Fig. 1. First, with the experimental tie line as
the initial guess, one predicts the tie lines with current interaction parame-
ters. A straight line (line e) is then drawn passing through the origin (point
o) and the overall experimental composition (point h) to find points f and g.
The intersections points f and g are used to obtain two predicted tie lines
(lines a and c) immediately above and below the experimental tie line (line

- Predicted tie lines

- Experimental tie line

dextran (wt%)
Fig. 1. Schematic diagram for tie line estimation method used with the new objective function
(eqn. (13b) with eqn. (9d) in text).
 253

d). The predicted tie line (line b) with the experimental overall composition
is then obtained by linear interpolation between these two predicted tie
lines.
For the calculation above, we need to calculate the tie lines corresponding
to phase equilibrium. For two-phase, three-component systems, we have

Apt=Apf’ i=l,2,3 wd

(lob)
i i

From the equivalence condition for chemical potentials of each component

in each phase (eqn. (lOa)) and material balances (eqn. (lob)), five equations
are obtained. In principle, the above set of equations is solvable by fixing
one of the six unknown compositions. However, satisfactory convergence is
not easily obtained because the chemical potentiais are nonlinear functions
of composition, and the inherent inaccuracy of the models (Guffey and
Wehe, 1972; Marina and Tassios, 1973; Hsu and Prausnitz, 1974).
Alternatively, an objective function equivalent to eqn. (10a) can be used
(Guffey and Wehe, 1972; Marina and Tassos, 1973; Hsu and Prausnitz,
1974)

The advantage of this objective function is that if one is approaching a

homogeneous solution, the denominator of eqn. (11) becomes very small so
that the residue of the function is increased to avoid the trivial solution. By
minimization of eqn. (11) for compositions, a tie line is obtained at a
prefixed composition. Using the tie lines determined with current parame-
ters in eqn. (9a) gives the residue of the objective function in terms of
compositions.
To calculate the objective function of eqn. (9a), S0rensen (1980) con-
strutted the binodal curve using the iterative method developed by Michel-
sen (1980). The predicted tie line to be compared to an experimental tie line
was then found by interpolation between tie lines a little above and a little
below the experimental tie line. The calculation method of S0rensen (1980)
is, however, valid only in the case that every predicted binodal curve can
envelope all the experimental tie lines during the minimization procedure,
which is not always possible since the critical point of the phase diagram
shifts with a variation of the parameters. If the phase envelope predicted
from a set of parameters excludes any experimental tie lines, these tie lines
can not be found from interpolation, and the minimization procedure may
fail. This is especially troublesome here since many of the experimental tie
254

lines for the polymer solutions considered in this study are measured near
the plait point.
Since eqn. (lOa) is derived from the eqn. (2), the parameters satisfying this
condition may not yield a global minimum of the molar Gibbs energy, which
is the necessary and sufficient condition for LLE. The best way to attain the
global minimum is, although time consuming, by using a direct search
method over the whole parameter space.
Two problems can be encountered when eqns. (10) are used to predict
ternary LLE. First, the trivial solution of two identical phases may be found,
so that a homogeneous phase is predicted. Second, the necessary condition
of eqn. (10a) may result in compositions at maxima or saddle points.
Fortunately, the maxima and saddle points in the Gibbs energy can be
avoided by performing the stability test (Sorensen, 1980; van Dongen and
Doherty, 1983). As stated in the previous section, the exact calculation of
the tie lines is very time consuming and must be performed at every step
during the minimization procedure when an objective function in terms of
compositions is used. To facilitate this tedious calculation, Sorensen et al.
(1979) used an approximate linear transformation method. We have also
used this method to generate tie lines, and compare it here with the exact
calculation of tie lines described above.
If a set of parameters predicts a phase envelope close enough to the
experimental one, a function dependent on the predicted compositions can
be expanded by Taylor’s series expansion as follows

I;;(x) =E;;(s) + Ccjj(xj-aj) 024

Cij = 2 with xZZk= 222k 024

J
By truncating terms second order and higher, we have a set of linear
equations in the predicted compositions. Since the function I;l,(a) denotes
the equilibrium condition for the predicted tie line, it is zero so every term
except Rj in eqn. (12a) can be evaluated using experimental data and a
model. Using the material balance conditions for both phases, the predicted
tie line is then approximated for a fixed composition.
As has been pointed out previously (Sorensen, 1980; Yu and Arnold,
1985; Ruiz and Gomis, 1986), the surface of the molar Gibbs energy
becomes more complicated as the number of parameters, and their values,
increase. Therefore, it is desirable to obtain a set of parameters of smallest
 255

value which can correlate the measured phase envelope. This may be
accomplished by the use of a penalty function. However, the penalty term
should not significantly affect the residues of the objective functions. For
this reason penalty terms of the form Q,ciP,’ and Q,C,pf as shown below
were added to eqns. (8) and (9a), respectively, where Q, and Q, were given
with small values so as not to alter the residues of the functions appreciably

F,(P) = ccc
k i j
0w
Thus, the model parameters are determined by minimizing an objective
function which satisfies the equivalence condition of the chemical potentials
for the same component in both phases, and the maximum likelihood
principle. Since the two thermodynamic models we consider differ only in
the numbers of the adjustable parameters, essentially the same minimization
procedure was used. The objective functions based on the equivalence
condition and the maximum likelihood principle are indicated by F,(p) and
F,(p), respectively.
With initial guesses as discussed below, all parameters were estimated by
fitting all the experimental tie lines using minimization of the objective
function F,(p). Once parameters which result in a reasonable binodal curve
are obtained, the objective function was changed to E;,(p) to improve the
predictions. In order to calculate Fx(p), the calculated tie lines were com-
pared with experimental data as explained above. The Nelder-Mead proce-
dure (Nelder and Mead, 1964) was used for minimization of an objective
function with two stop criteria. The first was to compare the standard error
in the residues of the objective functions over a simplex in the form
J, n with a prefixed small value, where n is the number of the
points and yj and ji are a local and average residues over the simplex. Since
we are interested in parameters which result in values of the objective
function close to zero, another stop criterion was comparison of the residue
of the objective function with a small prefixed value.

RESULTS AND DISCUSSION

To estimate the interaction parameters, LLE data of the polymer solu-

tions were taken from Albertsson (1971). In spite of polydispersivity in the
molecular weights of the polymers, the dextran-PEG-water systems were
assumed to be strictly ternary. Furthermore, branching and cross linking of
256

TABLE 3
The estimated Flory parameters for aqueous polymer systems of various dextrans and
PEG6000 using experimental data of Albertsson (1971). Q, and Q, were 1.0 x lo-‘* and
1.0 X lo-‘, respectively

Dextran Temp. (K) Dx-PEG Dx-Water PEG-Water Residue b

D17 293 0.0468 30 0.53500 0.50924 9.55x10-*
D17 273 0.041768 0.52348 0.46998 1.83x10-t
D24 293 0.028491 0.50852 0.44054 2.39x10-l
D37 293 0.027960 0.48799 0.41960 2.06x10-’
D37 273 0.043657 0.51741 0.50160 1.88x lo-’
D48 = 293 0.027406 0.51591 0.48685 2.47x10-l

a Standard error to stop the estimation procedure is 2.3 x 10-4.

b Calculated by eqn. (14) with eqn. (9d).

the polymers were neglected in computing polymer parameters. Thus, the

volume and area per unit mass of each polymer was taken to be independent
of the molecular weight.
The interaction parameters of the F-H model obtained by minimizing
eqns. (13) with eqn. (9d) are listed in Table 3. An initial guess of 0.5 was
used for all parameters. The interaction parameters between unlike polymer
segments were found to be small compared to those between polymer
segments and water. Also, as the difference in the molecular weights of the
polymers increases, the interaction parameter between the polymers was
found to decrease. From Fig. 2 and Table 3 we can see, by comparing the
calculated results with experimental data, that a lower interaction parameter
between the polymers predicts a plait point that is farther away from the
polymer-polymer axis. At 293 K, the difference between the interaction
parameters for each polymer and water also increased with increasing
molecular dissimilarity of the polymers. Also, the larger the difference of the
molecular weights of the polymers the more asymmetric the tie lines and
binodal curves. The location of the plait point, however, is mainly de-
termined by the interaction parameters between the unlike polymers.
Typically, interaction parameters are determined independently for each
binary pair to predict equilibrium in a ternary system. For example, the
interaction parameters between polymers and solvent can be obtained from
osmotic pressure data (Flory, 1971) or vapor pressure data (Booth and
Devoy, 1971) in binary systems. Also, the interaction between the polymers
can be measured by a variation of reversed phase liquid chromatography
(Horvath et al., 1976) or affinity chromatography and gel permeation
chromatography (Kodakura et al., 1981).
Edmond and Ogston (1968) obtained x23 = 0.436 from intrinsic viscosity
of PEG-water solutions using the analysis of Flory (1971) where the sub-
 (A)

1
I I I I I I I I I I I I I I I I I I I I
2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 22.5 25.0 5.0 7.5 10.0 12.5 15.0 17.5 20.0 22.5 25.0 27.5

WEIGHT PERCENT OF DEXTRAN TIE-LINE LENGTH(TLL)

Fig. 2. Comparison of calculated and experimental binodal curves and distribution coefficients of dextran for different
dextrar-PEG6OOC-water systems at 293 K using the F-H model: (2.A) and (2.B) for D17, (2.C) and (2.D) for D24, and (2.E) and (2.F) for
D37. Estimated parameters are listed in Table 3.

- - . ._ - _ -
258

NWLXBCI JO LN313IJJ303 NOIUXtI2lJSItl

._
;4” 33d JO LN331I3d ,LH313M
 259

NvxLx3a .LN313IU303

0 P t n
r

33d d0 LN3X3d ,LHE)I3M

260

TABLE 4
The estimated Flory parameters for aqueous polymer systems of various dextrans and
PEG6000 with xz3 fixed at 0.436 (Edmond and Ogston, 1968). Q, and Q, were as given in
Table 3

Dextran Temp. (K) Dx-PEG Dx-Water PEG-Water Residue a

D17 293 0.040107 0.50020 0.43600 9.36x lo-’
D24 293 0.027694 0.50718 0.43600 2.39 x 10-l
D37 293 0.030125 0.49449 0.43600 2.06 x 10-l
D48 293 0.046780 0.47932 0.43600

a Calculated by eqn. (14) with eqn. (9d).

scripts 1, 2 and 3 refer to dextran, PEG and water, respectively. We

estimated x13 and xl1 with this fixed value of xz3 by fitting experimental
data.
The results are listed in Table 4 with values of the corresponding residue
calculated from

with x,; = 22. Except for the D48 system, the interaction parameters
computed with the fixed xz3 are close to those obtained when all three
parameters are estimated simultaneously.
In Table 5 are shown the interaction parameters of the F-H model
determined from other literature data along with xz3 obtained as described
above. The parameter xl3 was calculated from the second virial coefficient

TABLE 5
The Flory parameters determined for aqueous polymer systems of various dextrans and
PEG6000 at 293 K

Dextran Dx-PEG a Dx-Water b PEG- Water ’ Residue d

D17 0.047150 0.4720 0.436 1.16x10-’
D24 0.046415 0.4761 0.436 2.77x10-l
D37 0.041675 0.4798 0.436 2.25 x10-l
D48 0.043808 0.4849 0.436 3.66x10-l
D68 0.045260 0.4867 0.436

a Calculated from eqn. (15) using plait points data (Albertsson, 1971).
b Calculated from second virial coefficients (Edmond and Ogston, 1968) obtained from the
plait point data (Albertsson, 1971).
’ From Edmond and Ogston (1968).
d Calculated by eqn. (14) with eqn. (9d).
 261

using osmotic pressure information (Edmond and Ogston, 1968) and the xi2
was calculated from plait point data using the equation below derived by
Scott (1949)

(15)

This equation is approximately correct here since xl3 = x13. As might be

expected, the binodal curves predicted with these parameters are further
from the experimental data than when the parameters calculated here from
experimental data are used.
The interaction parameters between each polymer and water found by
fitting experimental equilibrium data were close to 0.5, irrespective of the
molecular weight. Interestingly, smaller values of these interaction parame-
ters are calculated from osmotic pressure data.
According to Scott (1949), phase separation occurs more readily the
higher the polymer molecular weight, independent of the interactions be-
tween the polymers and solvent. Table 5 sh.ows the decrease of xl2 with
increase of molecular weight of dextran. Thus we conclude that the entropy
of mixing of the higher molecular-weight dextran is partially compensated
by smaller unfavorable interactions between the polymers on demixing.
The estimated phase envelopes and distribution coefficients are plotted in
Fig. 2. The shift of the binodal curves with change of the molecular weight
of dextran is in the same direction as the experimental data. The difference
between the predicted and experimental compositions of dextran in the
dextran-rich phase is much larger than in the PEG-rich phase. Since relative
differences of the two compositions are used in the objective function rather
than absolute differences, the lower compositions of dextran in PEG-rich
phase is more tightly fitted. The distribution. factor of dextran, defined as
the ratio of dextran weight fraction in the bottom to the top phases, is
shown in Fig. 2 as a function of tie line length (TLL). With increasing tie
line length the difference between predicted and experimental distribution
factors of dextran becomes larger. Three results for the distribution coeffi-
cients of dextran are compared for the D37--PEG6000 system at 293 K in
Fig. 3. Here we see that the interaction parameters obtained from the
literature lead to predictions of the correct behavior, and the two and three
parameter estimations result in an almost identical behavior.
The UNIQUAC model is, in principle, applicable to polymer solutions,
though it has not been applied to the LLE of ternary polymer solutions. The
experimental data for the ternary polymer solutions of tAtis study were also
fitted to estimate the UNIQUAC parameters. As outlined before, the
parameters which resulted from the minimization of the objective function
 l Experiments
0 Three-parameter estimation
. Two-parometsr estrrdion
P Parameters from other sources

1
I I I I I I I I
6 8 10 12 14 16 18 20

TIE-LINELENGTH(TLL)
Fig. 3. Comparison of distribution coefficients of dextran calculated with different interaction
parameters for D37-PEG6000-water system at 293 K. Results mentioned in the legends are
given in Table 6.

of eqn. (13a) were used for the initial guess of the next adjustment by
minimization of the objective function of eqn. (13b) with the constraint of
eqn. (9b) or eqn. (9d). To calculate the equilibrium compositions with the
given model parameters, both the linear transformation method and the
minimization of eqn. (11) in compositions were tested. In Table 6 are listed
the interaction parameters determined by different objective functions and
calculation methods for the D17-PEG6000 system at 293 IS. Results A and
B were obtained by minimizing eqns. (13) with eqn. (9b), though the linear
transformation method was used to obtain result B. The results C and D
were obtained with the tie lines calculated by minimizing eqn. (1Oa) for
 263

TABLE 6
UNIQUAC parameters determined for the D17-PEG6000-water system at 293 K using
various objective functions and constraints. The initial guesses for A,, and A,, were 115.62
and - 95.701 with subscripts 1, 2 and 3 for dextran, PEG. and water. r’s and q’s of dextran
and PEG were 0.02717 and 0.0387, and 0.0196 and 0.025’7, respectively. Q, and Q, were set
equal to l.OX1O-8

A 12 a 43 A21 A 23 A31 A32 Residue b

A 865.55 287.67 - 214.22 1475.5 - 311.32 -8.1594 1.31x10-’
B 815.28 366.67 257.23 40.181 - 1061.9 - 178.24 1.83 x10-l
C 812.71 355.69 270.95 28.781 - 1057.5 - 201.63 7.44x 1o-2
D 305.24 - 140.15 68.143 - 203.87 287.87 - 597.68 5.97x1o-2

A: Estimated by minimizing eqn. (13a).

B: The tie lines during the estimations were obtained by linear transformation,
C: The objective function of eqn. (13b) with eqn. (9b).
D: The objective function of eqn. (13b) with eqn. (9d).
a Units in K.
b Calculated by eqn. (14) with eqn. (9d).

compositions and minimization of the objective function of eqn. (13b) with

constraints of eqn. (9b) and eqn. (9d), respectively.
The use of the linear transformation procedure led to the least accurate
predictions which indicates that the linear transformation cannot predict the
correct compositions with given parameters unless the model is accurate and
the current parameters are close to the optimal ones. Unfortunately, neither
the predictability of the model nor the opt.imal parameters are known
beforehand. The objective function of eqn. (13a) gives qualitatively correct
phase behavior, but quantitatively is no better than the objective function of
eqn. (13b) no matter which constraint is used. The residues in Table 6 were
calculated using eqn. (14) with eqn. (9d); the binodals and the distribution
factors are plotted and compared in Fig. 4.
From different initial guesses for A,, and A,,, different sets of the
interaction parameters can be obtained with the objective functions of eqns.
(13) and constraint of eqn. (9b); these are listed in Table 7. The results
indicate the important role of the initial guess, and the possibility of the
multiple parameter sets resulting from the minimization procedure with the
UNIQUAC model. To reduce the arbitrariness, the initial guesses for the A,,
and A,, were calculated for a hypothetical dextran-PEG mixture obtained
by projecting an experimental tie line of the .largest length onto the poly-
mer-polymer axis. With U,, fixed arbitrarily, here at 100.0, U2s, U,, and U,,
were all equated to (U,, + U,, + U,,)/3 (Sorensen et al., 1979).
The UNIQUAC parameters and resulting residues thus obtained for the
various polymer systems are given in Table 8. Although the UNIQUAC
264

16

(A)
-E -
- -
- Result
Result
A
B
in
in
Table
Table
6
6
- Result C in Table 6
- Result 0 in Table 6

0 I I I I I I I I I I I
2.5 50 7.5 iO.0 12.5 15.0 17.5 20.0 22.5 25.0

weight percent of PEG

1000

(6)

Tie-Line Length(TLL)
Fig. 4. Comparison of binodal curves (4.A) and distribution coefficient of dextran (4.B)
obtained by different parameter estimation methods for D17-PEG6000-water system at 293
K using the UNIQUAC model. Estimated parameters are listed in Table 6.

model can predict the phase behavior fairly well, there is no consistency of
the parameters from system to system, or at different temperatures for the
same system.
It is known that the two UNIQUAC parameters are strongly intercorre-
lated for binary systems (Anderson and Prausnitz, 1978). To investigate the
 265

TABLE 7
UNIQUAC parameters determined for different initial guesses of A,, and A,, of the
D17-PEG6000-water system at 273 K where subscripts 1, 2 and 3 denote dextran, PEG and
water. For the initial guess A, - 100.00 was used for A,, and A,,, while 100.0 and - 100.0
were used for A,, and AZ1 in the initial guess B. r’s and q’s of dextran and PEG were
0.02717 and 0.0387, and 0.0196 and 0.0257, respectively. Residues were calculated by eqn.
(14) with eqn. (9d). Q, and Q, were both 1.0X10-s

A 12 A,3 A21 A 23 A31 A 32 Residue

A 373.10 6021.6 3334.9 9711.2 - 3891.4 - 61.443 1.36x10-’
B 362.02 3682.8 1995.5 294.04 - 1282.8, - 33.566 1.34x10-l

degree of parameter intercorrelation for the D17-PEG6000 system at 293 K,

Ai2 and A,, were varied simultaneously keeping the remaining parameters
constant. The region in which any combination of the parameters can
predict the binodal curve with less than 2.7556 variation of the resulting
residue is plotted in Fig. 5. The results clearly indicate the existence of
multiple parameter sets. Thus the inconsistency of parameters mentioned
above may be the result of multiple parameter solutions.
The predictions of Gustafsson et al. (1986) and the estimated results of
this study using both the F-H and UNIQUAC models are compared for the
D17-PEG6000 system at 293 K in Fig. 6. Clearly, the six parameter
UNIQUAC model best fits the phase behavior of the system.

TABLE 8
UNIQUAC parameters determined for various dextran-PEG-water systems using the new
objective functions. The r’ and q’ parameters are as in Table 7

Systema A,, A,, A 21 A 23 A:,1 A32 Residue

A 305.24 - 140.15 68.143 - 203.81 - 287.87 - 597.68 5.97x1o-2
B 108.79 - 226.18 166.84 5.0415 234.34 - 30.944 5.43 x 1o-2
C 416.20 - 239.67 - 200.18 - 264.71 315.66 442.71 2.63 x 1O-2
D 631.22 - 167.29 - 163.13 - 255.94 - 197.35 - 123.05 1.60x10-i /
E 279.19 169.28 159.18 - 377.44 - 292.40 686.62

a Specifications and the initial guess of Al2 and A,, are given by:

System Temp. (K) PEG Dextran A,, A,:, Qx b NTL ’

A_ 293 6000 D17 115.62 - 95.701 1.0x10-‘0 7
B 273 6000 D17 115.65 - 95.730 1.ox1o-1o 4
C 293 6000 D24 141.76 - 1113.95 l.oxlo-‘2 5
D 293 6000 D37 119.35 - 99.068 1.0x10-s 4
E 293 4000 D48 137.34 - llO.60 1.0xlo-‘2 5

b Q, = 0.0 for A, B, and D but l.OX1O-‘2 for C.

’ Number of experimental tie lines used in the parameter estimations.
266

90

TO
c
.-

n
*-

pj-, 80

I I I I I
zoo 250 300 350 4w 450 10

A(l,Z) in K
Fig. 5. Intercorrelation of the UNIQUAC parameters for D17-PEG6000-water system at
293 K.

Since the interaction parameters of the UNIQUAC model are based on a

unit area of interacting surface, the interaction parameters should be con-
stant with change of molecular weight of polymers provided the chain is
linear. A set of interaction parameters, which was estimated by fitting the
experimental data of D17-PEG6000 system at 293 K with r’ and 4’ of
0.024218 and 0.01816 for dextran, was used to develop the binodal curve for
systems containing different dextrans. Figure 7 shows the molecular weight
effect on the shift of the binodal curve in both experiment and calculation.
An increase of dextran molecular weight results in a more asymmetric
binodal curve, and shifts the plait point further away from the
polymer-polymer axis in the predictions than is experimentally observed.
To investigate the effect of values of the r’ and q’ parameters on the
predictability of the D37-PEG6000 system at 293 K, values of the shape
parameters were changed alternatively or simultaneously while estimating
the interaction parameters using the objective functions of eqns. (13) with
 267

. Experiments
- Guslafsson’sresults
- - The Flary-Huggins Model
- The UNIQUAC Model

/
/
/
/

I I 1
10 20 30

Tie-Line Length(TLL)

Fig. 6. Comparison of predicted binodal curves for D17-I’EG6000-water system at 293 K.

eqn. (9b). As shown in Table 9, the different shape factors resulted in

different sets of interaction parameters and different residues. This suggests
that in addition to the interaction parameters, the surface areas and volumes
of the components are important properties in the prediction of the phase
behavior of these systems.

CONCLUSIONS

Two thermodynamic models were successfully used to describe the incom-

patibility of the aqueous polymer systems contaming two chemically differ-
ent polymers. To estimate the interaction parameters, a new objective
function in terms of compositions, which is based on the maximum likeli-
268

16 ,

+- D4B,'PEG6000,'water
-O- D37/PEC8000/rater
+I- D24,'PEC6000/rater
h D17/PEG6000/rater

2.5 5.0 7.5 10.0 12.5 15.0 17.5 20.0 22.5 25.0
!I
WEIGHT PERCENT OF DEXTRAN
Fig. 7. Effect of molecular weight of dextran on the predicted binodal curves. The shape
factors r’ and q’ are 0.024218 and 0.01816 for dextran, 0.0387 and 0.0257 for PEG,
respectively. The shape factors for water are given in Table 2. The estimated UNIQUAC
parameters are A,, - 111.6, A,, = - 155.62, A,, = 0.57682, A,, = - 229.29, A,, = - 75.246,
A,, = - 186.27.

hood principle, was introduced and shown to be superior to objective

functions used previously for liquid-liquid equilibrium.
The F-H model with the parameters determined using this new objective
function correlated the liquid-liquid phase behavior in a satisfactory way.
As the dextran molecular weight increased, the estimated interaction param-
eters between the unlike polymers became smaller; the interaction parame-
ters between the unlike polymers determines the 1plait point, while the
asymmetry of the binodal curves is dependent on the interaction parameters
between the polymers and water. It is important to note that the use of
 269

TABLE 9
UNIQUAC parameters determined for the D37-PEGCiOOO-water system at 293 K with
varying shape factors. The objective functions of eqns. (13) with eqn. (9b) were used to
estimate the parameters. The resulting residues were calculated using eqn. (14) with eqn. (9b).
Q, and Q, were 1.0~ lo-’

Result a A,, Ai, A21 A23 A31 A32 Residue

A 52.455 - 189.43 274.81 - 53.174 21.589 -271.64 5.06 x 1O-2
B 32.936 - 273.19 125.65 - 187.18 50.880 - 110.17 9.19 x 1o-2
C 72.234 - 185.94 - 21.017 - 196.01 135.23 - 113.58 1.45 x10-’
D 102.87 - 241 .OO 0.95829 - 214.42 140.55 - 136.79 4.12x 1O-2
E 431.83 - 219.32 - 168.94 - 116.94 239.76 360.08 1.57 x 10-2
F 260.17 - 225.94 - 20.017 - 218.59 129.68 - 11.165 6.02 x 1O-2

a The shape factors for each system are given by:

System r’ X102 q’ x lo2

PEG Dextran PlEG Dextran

A 3.87 2.717 2.57 1.96

B 3.87 2.717 2.57 1.764
C 3.87 2.4453 2.57 1.96
D 3.87 2.4453 2.57 1.764
E 3.483 2.717 2.323 1.96
F 3.87 2.717 2.313 1.96

interaction parameters obtained from binary osmotic pressure and plait

point data resulted in reasonable predictions of liquid-liquid equilibrium in
the ternary system.
We investigated methods of calculating the residue of the objective
function in terms of composition. A smaller residue of the objective function
was obtained by matching the overall compositions rather than by arbi-
trarily fixing one of the equilibrium compositions as has been done in the
past. The linear transformation method was found to be inferior to the
direct minimization method used here. It was also shown that due to the
intercorrelation of parameters, the initial guess of their values affects the
final results. The interaction parameters of the UNIQUAC model are
strongly intercorrelated implying existence of multiple sets of optimal inter-
action parameters. By assuming that the shape factors of dextran based on
unit mass are independent of chain length, the effect of molecular weight on
the phase behavior was predicted in a qualitatively correct manner.
Generally, the six parameter UNIQUAC model yielded a better corre-
lation of the phase behavior of the systems than the three parameter F-H
model. However, since the parameters in the F--H model are better behaved
on varying molecular weight and can be obtained from osmotic pressure and
270

plait point data, the model may be more useful for predicting the phase
behavior of dextran-PEG-water systems.

ACKNOWLEDGMENTS

This work was supported in part by a grant from the Korean Ministry of
Education to C.H.K., and by Grant No. DE-FG02-85ER 13436 from the
United States Department of Energy to the University of Delaware.

LIST OF SYMBOLS

A,j UNIQUAC parameters

‘i
activity
cij Jacobian matrix in eqn. (12a)
F objective function
G Gibbs free energy
cl- constant in eqn. (7d)
Mj molecular weight
*j molar volume ratio of polymer to water
N number of experimental tie lines
n number of moles
P unknown parameters
Pi adjustable parameters
Q constant
4; surface area parameter
ri volume parameter
u UNIQUAC parameter
X weight or mole fraction
X composition vector
Yi residue of a function
z coordination number

Greek letters

activity coefficient
2 surface area fraction
lli chemical potential
‘ji
UNIQUAC parameter
+j volume fraction
Xij Flory-Huggins parameter
J/j volume fraction
 271

Superscripts

C pertaining to the critical point

I pertaining to properties per unit mass
,T
estimated properties
averaged

Subscripts

1 dextran
2 PEG
3 water
e based on equivalence condition
i, j, k component, phase and sequential number of tie lines, respectively
X in terms of compositions

REFERENCES

Abrams, D.S. and Prausnitz, J.M., 1975. AIChE J., 21: 11.6.
Albertsson, P., 1971. Partition of Cell Particles and Macromolecules, 2nd edn. Wiley, New
York.
Allen, G., Gee, G. and Nicholson, J.P., 1960. Polymer, 1: 56.
Anderson, T.F. and Prausnitz, J.M., 1978. Ind. Eng. Chern. Process Des. Dev., 4: 552.
Anderson, T.F., Abrams, D.S. and Grens II, E.A., 1978. AIChE J., 24: 20.
Bamberger, S., Brooks, D.E., Sharp, K.A., van Alstine, J.M. and Webber, T.J., 1985. In: H.
Walter et al. (Editors), Partitioning in Aqueous Two-phase Systems. Academic Press,
London, p. 90.
Bondi, A., 1968. Physical Properties of Molecular Crystals, Liquids and Gasses. Wiley, New
York, p_ 485.
Booth, C. and Devoy, C.J., 1971. Polymer, 12: 309.
Brelvi, S.W., 1982. Ind. Eng. Chem. Process Des. Dev., 21: 367.
Edmond, E. and Ogston, A.G., 1968. Biochem. J., 109: 569.
Flory, P.J., 1971. Principles of Polymer Chemistry. Cornell University Press, Ithaca, p. 510.
Gmehling, J., Rasmussen, P. and Frendenslund, A., 1982. Ind. Eng. Chem. Process Des. Dev.,
21: 118.
Guffey, C.G. and Wehe, A.H., 1972. AIChE J., 18: 913.
Gustafsson, A., Wennerstrom, H. and Tjerneld, F., 1986. Polymer, 27: 1768.
Horvath, C., Melander, W. and Molnar, I., 1976. J. Chrornatogr., 125: 129.
Hsu, C.C. and Prausnitz, J.M., 1974. Macromolecules, 7: .320.
Kodakura, S., Miyamoto, T. and Inagaki, H., 1981. Biopolymers, 20: 1113.
Marina, J.M. and Tassios, D.P., 1973. Ind. Eng. Chem. Process Des. Dev., 12: 271.
Michelsen, M.L., 1980. Fluid Phase Equilibria, 4: 1.
Nelder, J.A. and Mead, R., 1964. Computer J., 7: 308.
Ogston, A.G., 1962. Arch. Biochem. Biophys. Suppl., 1: 3’3.
Oishi, T. and Prausnitz, J.M., 1978. Ind. Eng. Chem. Process Des. Dev., 17: 333.
Reid, R.C., Prausnitz, J.M. and Sherwood, T.K., 1977. The Properties of Gases and Liquids,
3rd edn. McGraw-Hill, New York, p. 57.
272

Robledo-Muniz, J.G., Tseng, H.S., Lloyd, D.R. and Ward, T.C., 1985. Polym. Eng. Sci., 25:
934.
Ruiz, F. and Gomis, V., 1986. Ind. Eng. Chem. Process Des. Dev., 25: 216.
Sandler, S.I., 1977. Chemical and Engineering Thermodynamics. Wiley, New York, p. 359.
Scott, R.L., 1949. J. Chem. Phys., 17: 279.
Shibata, SK., Sandler, S.I. and Behrens, R.A., 1987. Chem. Eng. Sci., in press.
Simonetty, J., Yee, D. and Tassios, D., 1982. Ind. Eng. Chem. Process Des. Dev., 21: 174.
Sorensen, J.M., 1980. Correlation of liquid-liquid equilibrium data. Ph.D. Thesis, Instituttet
for Kemiteknik Denmarks Tekniske Hojskole, Denmark, p. 54.
Sorensen, J.M., Magnussen, T., Rasmussen, P. and Frendenslund, A., 1979. Fluid Phase
Equilibria, 3: 47.
Tompa, H., 1956. Polymer Solutions. Butterworth, London, p. 183.
Van Dongen, D.B. and Doherty, M.F., 1983. Ind. Eng. Chem. Fundam., 22: 472.
Varhegyi G. and Eon, C.H., 1977. Ind. Eng. Chem. Fundam., 16: 182.
Yu, C.Y. and Arnold, D.W., 1985. Ind. Eng. Chem. Process Des. Dev., 24: 494.