Fluid Phase Equilibria, 3 (1979) 47-82 47

@ Elsevier Scientific Publishing Company, Amsterdam - Printed in The Netherlands

Review series

LIQUID-LIQUID EQUILIBRIUM DATA: THEIR RETRIEVAL,

CORRELATION AND PREDICTION

PART II: CORRELATION

JENS M. S(dRENSEN l, THOMAS MAGNUSSEN 2, PETER RASMUSSEN 1

and AAGE FREDENSLUND 1*
1 lnstituttet for Kemiteknik, Danmarks Tekniske Hbjskole, DK-2800 Lyngby (Denmark)
2 DECHEMA, Postfach 970146, D-6000 Frankfurt (Main) 97 (BRD)
(Received October 20th, 1978)

ABSTRACT
Serensen, J.M., Magnussen, T., Rasmussen, P. and Fredenslund, A., 1979. Liquid-liquid
equilibrium data: their retrieval, correlation and prediction. Part II: correlation. Fluid
Phase Equilibria, 3: 47-82.

The correlation of liquid-liquid equilibrium data using models for the liquid phase
activity coefficients - for example NRTL and UNIQUAC - is reviewed. Different numer-
ical procedures used in the computation of liquid-liquid equilibrium compositions on the
basis of these models are evaluated. Methods for obtaining parameters from liquid-liquid
equilibrium data are described, and a method leading to improved representation of tie
lines using relatively few parameters is recommended. The NRTL and UNIQUAC equa-
tions are compared with respect to their ability of representing binary and ternary liquid-
liquid equilibrium data. The UNIQUAC equation appears to be more convenient to use
and to correlate the data slightly better than does NRTL with the same number of param-
eters.

INTRODUCTION

In Part I of this series (Sorensen et al., 1979) we described the different

types of liquid-liquid equilibrium phase behavior in binary and ternary mix-
tures of interest in the petrochemical industry and related applications. The
mixtures taken into consideration are those containing organic compounds
and water at temperatures between 0 and 150°C and pressures up to 20 atm.
Mixtures containing strong electrolytes, polymers, or components with nor-
mal boiling points below 0°C were excluded from consideration. In Part II
we describe the correlation of liquid-liquid equilibrium data for these mix-
tures; we both review work in the area reported in the literature between
1965 and 1978 and present recent results of our own.
By correlation we mean representation (or reproduction) of experimental

* To whom correspondence should be addressed.

48

liquid-liquid equilibrium data using analytical models. We shall deal mainly

with binary and ternary data, but many of our conclusions apply also to the
general, multi-component case.
The computation of liquid-liquid equilibria using group-contribution
models and temperature-extrapolation beyond a limited range are examples
of prediction. That will be covered in Part III. However, Part II does to a
limited extent describe the computation of ternary tie lines from binary
data only.
A design engineer may pose questions regarding liquid-liquid equilibria
at two different levels. He may ask: given a certain mixture, do I have one
or do I have two liquid phases? Or he may ask: if there are two liquid phases
in equilibrium, what is the composition in each of the phases?
In this article we deal mainly with the latter question. Of course, any
method giving a satisfactory answer to the quantitative question will answer
the qualitative one correctly.
Traditionally, two different categories of models have been used in corre-
lating liquid-liquid equilibrium data: (1) Models which empirically describe
component distribution ratios (pi) as functions of composition. The Hand
(1930) method is a model of this type. Because of their empirical nature,
models of this type can only be used for storage and interpolation purposes.
(2) Models which are founded in the thermodynamics of fluid-phase equilib-
ria. Most of these models describe the mixture Gibbs energy as a function
of composition and (perhaps) temperature. These models may, at least in
principle, be used for both interpolation and extrapolation purposes.
This work considers only models of the latter type.

CONDITIONS OF LIQUID-LIQUID EQUILIBRIA

Consider nT moles of liquid mixture with N components at fixed tempera-

ture and pressure. The necessary and sufficient condition of equilibrium is
that the Gibbs energy of mixing n,AG for the mixture is minimum.
If the mixture achieves the lowest Gibbs energy by splitting into two liquid
phases I and II we may write
n,AG = n.‘AG’(ni nf, . ..) + n’lAG”(n~r, n!, . ..) (1)
where nf and nil are the number of moles of component i in phase I and II,
respectively, and AG’ and AG” are the molar Gibbs energies of mixing corre-
sponding to n’ moles of phase I and nrl moles of phase II, respectively.
Since the molar Gibbs energy of mixing AG is minimum, a differential
change of’composition occurring at equilibrium at fixed pressure and tempera-
ture will not produce any change in AG and hence:

d(AG)P,T = 0 (2)
This criterion is a necessary, but not sufficient condition of equilibrium
between phases I and II; eqn. (2) does not help us in distinguishing between
 49

a maximum, an inflection point, and a minimum.

The usual equilibrium condition is derived from eqn. (2):
/J!I = /J-1!’
I i=l,2 ) . .. . .) N (3)
where plr is the chemical potential of component i in phase I or II.
Equation (3) is normally rewritten in terms of the fugacities fi or more
conveniently in terms of the activities Ui
f; = f;’ (4)

where yi is the activity coefficient.

It is characteristic for liquid-liquid equilibria that often the component
fugacities cannot be estimated by independent means. All we know is that
f,! equals fi’. Th is accounts for the fact that computation of liquid-liquid
equilibrium compositions is more difficult than the similar computations for
vapor-liquid equilibria. In the latter case we can often estimate actual values
of the component fugacities from the vapor-phase properties. This observa-
tion leads to yet another significant conclusion. Since in liquid-liquid equi-
libria it is not possible to estimate component fugacities by independent
means, there is no “thermodynamic consistency test” corresponding to that
for vapor-liquid equilibria (Van Ness et al., 1973).
The molar Gibbs energy of mixing for a single phase may be stated in terms
of an ideal and an excess contribution:

AG’ = RTC xf In xi + GE(‘)

i = 1, 2, . . . . N (6)
AG” = RTA xfl ln xf’ + Gnui)

The excess Gibbs energy GE and the activity coefficients yi are interrelated
by the following relationships:

GE = RTC Xi In yi ,
and
i = 1, 2, ,.., N (7)

where ni is the number of moles of component i and nT is the total number

of moles. The differentiation is at constant temperature, pressure, and num-
ber of moles of each component j except i.
Equations (2) and (5) are those which are most often used in liquid-liquid
equilibrium calculations. In order to actually carry out these calculations
50

one must have available:

(i) a model giving GE or, equivalently, yi as functions of composition and
temperature,
(ii) a method for calculating liquid-liquid equilibrium compositions using
the above model.
These points will be covered in subsequent sections.

MODELS FOR LIQUID-LIQUID EQUILIBRIA

The correlation of liquid-liquid equilibria using Gibbs energy functions is

often treated as an “appendix” in articles dealing mainly with vapor-liquid
equilibria. The same few triangular diagrams find repeated use as illustrations
of the capabilities in predicting liquid-liquid equilibria. A systematic study
of the correlation and prediction of liquid-liquid equilibria has to our knowl-
edge not been carried out to date. In this section we list the models, based on
classical or molecular thermodynamics, used for correlating liquid-liquid
equilibria in the literature between 1965 and 1978.
Three different types of models have been used:
(i) models for activity coefficients or the excess Gibbs function (see eqns.
6 and 7),
(ii) equations of state,
(iii) group-contribution methods (to be discussed in Part III).

Models for activity coefficients

The different models used for correlating liquid-liquid equilibria are

shown and briefly commented on in Table 1. The equations for the activity
coefficients may be obtained from the expressions for GE/RT by differentia-
tion according to eqn. (7).
The Margules, Van Laar, Redlich-Kister, and Black equations have in com-
mon that they in the past have been useful for correlating liquid-liquid equi-
librium data, often with quite good results. However, extrapolation to con-
centrations beyond the range of the data or the prediction of ternary phase
diagrams from only binary information should not be carried out with these
models - the results are often not even qualitatively correct.
Lo@ composition models (modifications of Wilson’s equation, NRTL,
and UNIQUAC) have proven superior to the older models, both for corre-
lating binary and ternary liquid-liquid equilibria and for predicting ternary
phase diagrams from binary data. The correlation can often be done with an
accuracy sufficient for design purposes, and the predictions are frequently
qualitatively correct. Modifications of Wilson’s equation and NRTL have
received much attention in the literature on liquid-liquid equilibria in recent
years - NRTL because it has proven relatively successful for correlating
liquid-liquid equilibria, and Wilson’s equation because it has only two
adjustable parameters per binary, whereas NRTL has three. Among the many
TABLE 1

Activitv coefficient models for correlatmg liquid-liquid equilibria

Name and original Equation for excess Gibbs function References, 1965-1978 Comments
reference

Margules fwo-suffix Joy and Kyle. 1969 These equations ca” be used for correlating binary and
(Margules. 1695) Newsham and “ahdat. 1971 ternary LLE bul often the results are poor. They
should not be used to Predict ternary LLE from binary
information or in extrapolatio” to concentrations out-
i. , = 1, 2. _.., N side the range of the data.
The Margules and Van Laar equations contain only
“a” Laar Joy and Kyle. 1969
binary parameters.
(Van Lxll. 1910) Ragaini et al.. 1974
The Van Laar and Redlich-Kister equations are
written for ternary mixtures: they can be extended
to the general. multicomponent case.

(ternary ; for blnar” x , = 0)

Redlich-Kistrr Ragaini and Baldini, 1973

=xlxl[~,2+ C,#,-x2)+ D,z(.xl -x2)* +...I
(Redlich and Kister. Ragaini et al., 1974
1948)
(binary)

+ r,xp3[C+ D,U- X3) + D2(x3-x,) + ...I

Kretschmer-Wiebe It 1s assumed that one of the components is linearly self- Nl‘ta et al., 1974 Maunly used for mixtures containing an alcohol. The
(Krctrchmvr and associating. The association esuilibrium constant is corre- Takeuchi et al.. 1975 correlation of binary and ternary systems is reason-
Wmbe. ,954) lated as a funclion of temperature. Binary equations we Sugi et al., 1976 ably good. although difficulties are encountered when
give” by Kretschmer and Wiebe (1954):useful exte”llons Nagata et al.. 1978 the two-phase region is large. The prediction of ternary
to ternary systems are given by Su$i et al. (1976) and LLE from binary data appears questionable.
Nawta et al. (1978).

GE Guffey and Wehe. 1912 Three parameters are needed Per binary. Correlation of
== <Van Law) + C,2(Xl mmx2pxIx2 +
Newsham and Vahdat. 1971 binary and tetnau data reasonably successful. Should
not be used to predict ternary VLE from binary data
+C*3(X~-r3pX*X3 +c3,(x3-x,Px)x, or to ertra~olate beyond ‘he range of the data.

(ternary: for binary x3 = 0)

--- --- - --- - -. - .^ - _. - __ -

52
TABLE 1 (contmued 2)

Name and onpnal Equation for excess Gibbs function References. 1965-1978 Co”I!“e”ts
reference

Modifications of Hiranuma. 1975 Same remarks as for the modification of Nagata et al.
Wilson’s equation Schulte. 1977 Requires additional computation. since the model
1.i = 1. 2. _._,N
(WllsO”. 1964) cannot be wntte” in an explicit form.

Ati now takes into account m&-actions from both “ear-

est and nextnearest neighbows. An expression for Afj is
given. It has the form:

PC is a -0.85
constant.
NRTL I Rena” and Praurnitz, 1966 The most used model for LLE to date. There are three
(Renon. 1966) JOY and Kyle. 1969 parameters per binary: ‘ti, rji. a”d “0. The term OLd
and modifications i,i,l=l,2 ,..., N Rena” et al., 1971 may be fixed accordzng to Rena” and Prausnitz (1968)
Heideman” and Mandhane, 1973 Correlation of binary and ternary LLE is often quanti-
Marina and Tasmos. 1973a tatively correct. Prediction of ternary LLE from
Marina and Tassios. 1973b binary data IS often aualitativelv correct. Can be used
Tji = @,I _ *Tli)/RT; Gj, = eXP(Qj*Tji) Katavama et al., 19’73 in extrapolatm” with respect to composition.
Sji = Piji rji zrg: Gji# Gg Ragaim and Baldmi. 1973
Ragaini et al.. 1974
aji = CY’ir
Ftiinger and Bittrich. 1974
Greiner et al., 1974
Heinrich and Dojcansk,‘. 1375
Mattelin and Verhoeye, 1975
Leach, 1977
PabILes et al.. 1977
Newsham and “abdat. 1977
“arhew. and Eon, 1977
This work. 1979

S,m,lar to NRTL Palmer and Snuth. 1972 TWO wrametwr per b,“arv. Not thoroughly tested.

LEMF As NRTL with a,, = -1. Marina and Tasmos. 1973a and b Only two parameters per binary. ApproximatelY
alllandj same quality of results as NRTL with ti,j fixed accord-
~“g to Rena” and Prausmtz (1968).
~~ ..~
TABLE 1 (continued 3)

Name and original Equation for excess Gibbs function References. 1965-1978 Comments
reference

NRTL Renon and Prausnitz. 1968 Only two parameters per binary (Cd does not cbntmn
(Rem,“, 1966) GE air). Not as good as NRTL for ternary LLE. Cannot
E= NRTL- h~ih[h Gjixjl
and modifications be used for Large miscibility gaps.

(Hei, and Prausnitz, Gji = (Vj/Vj) eXP(lji)

1966)
I, , = 1. 2. . . .. N

UNIQUAC Abrams and Praumitz. 1975 Only two parameters DP.~binary (TJ and Tj;). C~rrela-
GE = Gknb. + ees.
(Abrams and Raus- Anderson and Prausnitz. 1978 tion of binary and ternarv LLE is often quantltativclY
nitz. 1975) This work. 1979 corxct. Prediction of ternary LLE from binary data
and modifications Go often qualitatively correct. Can be used in extrapola-
_comb,/RT= kXr,“%.~ hqx&
XI 2 I bon with respect to composition. in this work we
conclude that UNlQUAC is as good as or better than
NRTL for correlating LLE.

Gf&,iR?. = - z& 9,~~ 1111 A ojr,il

Oi = XiLl,/ f; XjSj (surface drea fraction)

0, = Xi’,, i: %I‘J (“oiume fractron)

rj, = exp,--(ujl - uiJiRT1

Tjl * 7”; L = 10

i, I = 1. 2. . N

“era et al.. 1977 Two parameters per binary. Not thoroughly tested.
May yield similar results as UNIQUAC.
 55

proposed modifications of Wilson’s equation, that due to Hiranuma (1975)

is relatively well documented (Schulte, 1977). Hiranuma’s modification is
difficult to use, since it is not explicit in concentration. All in all, in liquid-
liquid equilibrium calculations NRTL is preferable over the various modifica-
tions of Wilson’s equation. NRTL is thoroughly tested, and the third param-
eter does provide extra flexibility. However, the third parameter cannot
always be treated as a parameter - it is only possible to fix two parameters
from binary, mutual solubility data.
The UNIQUAC model has only two adjustable parameters per binary.
For a few systems, Abrams and Prausnitz (1975) show that LJNIQUAC per-
forms reasonably well, both in predicting ternary diagrams from binary infor-
mation only and in correlating ternary diagrams. Anderson and Prausnitz
(1978) show that UNIQUAC predicts very well ternary diagrams from binary
information when binary vapor-liquid and liquid-liquid equilibrium data
are correlated simultaneously with only a few ternary tie lines.
Later in this article we give a detailed comparison between UNIQUAC and
NRTL in representing liquid-liquid equilibria.

Equations of state

The activity coefficient of component i in a mixture is related to the mix-

ture F-V-T properties by the relationship:

where e is the pure-component vapor pressure and ni is the number of moles

of component i. An expression for yi can thus be derived from any equation
of state which relates P, V and T. It is, therefore, in principle possible to pre-
dict liquid-liquid equilibria from equations of state.
The Redlich-Kwong (Redlich and Kwong, 1949) equation of state was
used to predict the ternary phase diagram for the system nitroethane-octene-
octane at 0°C from binary data (Lu et al., 1974). Only one parameter per
binary was employed, and the result was surprisingly good, even near the
plait point. A modification of the Redlich-Kwong equation of state was used
by Heidemann (1974) to calculate vaporliquid-liquid equilibria for mix-
tures of water and lower hydrocarbons. The Peng-Robinson equation of
state has been used for similar systems (Peng and Robinson, 1976). The two
latter references are mainly concerned with the high-pressure region.
The computation of liquid-liquid equilibria using equations of state has
to date received only limited attention. However, the results referred to here
indicate that this possibility warrants further exploration.
56

PARAMETERS IN LOCAL COMPOSITION EQUATIONS

For binary mixtures, both the NRTL and UNIQUAC equations contain
parameters pi a and ~si expressed in principle as

712 = exp[-(4, - k)lRTl and 721 = exp[-(u2i - ~11)1~~1

where the parameters Uij in some way characterize the energy of interaction
between molecules i and j; based on physical arguments, the requirement
Uji = Uij

is a reasonable one.
Only differences between the Uij’s appear in the parameters rij. Hence the
value of rij does not change if we add the same constant to each Uij, which
implies that one of the Uij’s may be fixed arbitrarily. Therefore, for a binary
data set we may determine either the parameters

(U12 - U22) and (u21 -u11) (2 parameters)

or

42 and u22 (2 parameters)

having arbitrarily fixed the value of u1 i .
For a ternary mixture, the parameters ris, r3i, 723 and r3s are needed in
addition to r1 2 and r2i. In that case we may fit to a ternary data set either
the parameters

(u12 -u22), (u21 -u11), (%3-u33), (u31 -Ull), (u23 -u33)

and

(“32 - U22) (6 parameters)

or

u12, u22, %3r u33 and u23 (5 parameters)

having arbitrarily fixed the value of uil.
The general, multicomponent case is discussed by Hala (1972).
The NRTL equation contains, included in the parameters Tijr the non-
randomness parameter afj (aij = aj,, see Table 1). The value of Olijmay be
preset, or it may be left as a parameter to be determined. Thus the NRTL
equation,has 5, 6, 8 or 9 adjustable parameters for describing a ternary data
set.

COMPUTATION OF LIQUID-LIQUID EQUILIBRIUM COMPOSITIONS

Let us assume that we have a liquid mixture of given overall composition

and temperature and that the liquid has separated into two equilibrium
phases. How does one calculate the compositions of the two phases using a
 57

given excess Gibbs energy model with given parameters?

A method of doing this must be based on one of the formulations of the
equilibrium criterion, eqns. (l)-( 5). Most authors use a method based on
eqn. (5) (the isoactivity method), and some use eqn. (1) (minimization of
Gibbs energy) or eqn. (2).
In the case of the isoactivity method, the following system of equations is
established:
x;y; = “;I$’ (N equations) (5)
n.=n!
I 1 +n!’I (N equations) (9)

where ni, ni and n:’ are, respectively, the number of moles of component i in
the whole system, phase I and phase II.
In the case of minimization of the Gibbs energy, eq. (9) must be satisfied
together with the criterion of minimum in the Gibbs energy with respect to
the independent compositions. In some cases, eqn. (9) is replaced by both of
eqns. (10) and N - 2 concentration specifications:

It is thus possible to use either of the two equilibrium criteria (minimiza-

tion of Gibbs energy or the isoactivity criterion) for calculating the unknown
compositions. The question now is: Does it matter which one of the two
methods we use? Are the two methods equally good, and do they give the
same results?
The best method to use is the AG-minimization, since it is based on eqn.
(1) which is both a necessary and sufficient criterion for equilibrium between
two immiscible liquid phases. However, AG-minimization is no guarantee
against false solutions represented by local minima in the Gibbs energy sur-
face, unless a very time-consuming search for a global minimum is carried out
in the whole composition space. This is illustrated for a binary mixture in
Fig. 1. Both the lowest tangent and the dot-dash tangents correspond to
liquid-liquid equilibrium compositions representing minima in AG. However,
the dot-dash tangents correspond to local minima yielding erroneously a
liquid phase with composition x O. The true minimum is given by the lowest
tangent corresponding to compositions x1 and XI’.
Figure 2 shows a more problematic binary AG-curve. If we have a feed
with composition zr , the predicted equilibrium compositions will be LX:and
LX:when the Gibbs energy is truly minimized. Thermodynamically speaking,
this is the correct solution. However, the observed compositions are X: and
xi’. This type of behaviour is associated with model weakness, or more pre-
cisely, inadequate parameter estimation techniques. In fact, the given model
with the given parameters does not represent a binary mixture with mutual
solubilities X: and xi’, but rather one with solubilities X: and x:.
The isoactivity criterion gives rise, as does AG-minimization, to false solu-
58

x: _
0 e 4 1 OX; 21 It
”

4 1
Fig. 1. Multiple solutions in binary liquid-liquid equilibrium calculations. - - - True
solution;. -. - . false solution including the local minimum at xy ; - - - - -- false solution
corresponding to maxima.

Fig. 2. Unfortunate representation of binary liquid-liquid equilibrium data. - - - Solu-

tion corresponding to experimental solubilities; . - . - solution corresponding to global
minimum.

tions represented by local minima. However, since the isoactivity criterion is

based on eqn. (5) it has even more problematic cases. Firstly, eqn. (5) has the
trivial solution:
x!1 = jc?I
I i = 1, 2, . . . . N
This solution could lead one to the erroneous conclusion that the mixture is
homogeneous. Secondly, using the necessary condition only may result in
compositions which represent maxima or saddle points in the Gibbs energy
surface (see dotted line in Fig. 1). Maxima or saddle points may be avoided
by a stability analysis, see Appendix A. When this is done, the isoactivity
method becomes nearly identical to minimizing AG.
The minimization of AG may be performed using a direct search procedure.
A faster method is to solve eqn. (2) by e.g. Newton-Raphson iteration and
check whether the solution represents a minimum in AG (Councell et al.,
1970; Varhegyi and Eon, 1977). In this way both the results and the con-
sumption of computer time will be the same as for the isoactivity method,
when the stability criterion in Appendix A is used.

COMPUTATION OF BINODAL CURVES

The previous section was devoted to the computation of a single tie line.
Additional considerations need to be made when a complete binodal curve is
to be computed.
Consider the following task: Given a set of binary (UNIQUAC or NRTL)
interaction parameters, calculate the complete binodal curve for a ternary
system. Obviously this may be accomplished by establishing a series of tie
lines by the method of the previous section. If these tie lines are spaced
throughout the two-phase region, the entire binodal curve is readily drawn.
 59

1' .l” .2 .3 .4 .5 .6 .7 .B .o
‘3
Fig. 3. The benzene(l)-2-propanol(2)-water(3) ternary system represented by UNIQUAC
with parameters obtained from binary data only, Note the discontinuities at A, B and C.
The encircled numbers indicate the number of liquid phases in equilibrium in each region.
PI and P, are plait points.

While the above procedure appears to be straightforward, there is in reality

an additional difficulty which occurs frequently for both UNIQUAC and
NRTL and for both correlation of ternary data sets and their predictions
from binary data. The difficulty arises from unfortunate solutions to eqns.
(2) and (5) as illustrated in Fig. 3, which shows binodal curves calculated for
system 508 of Table 2 using the UNIQUAC model with parameters obtained
from binary data only. The situation shown occurs when the slopes of the tie
lines change radically over a small vertical distance, here near point A, in the
triangular diagram, A is in equilibrium with both B and C, and hence.B is also
in equilibrium with C. We have thus obtained three liquid phases in equilibri-
um - a situation which does occur in practice (see Heidemann and Mandhane,
19’73), but very rarely so. The predicted l-, 2-, and 3-liquid phase regions are
indicated in Fig. 3. The experimental data for the benzene-2-propanol-water
system do not verify this rather exotic behaviour.
In correlating ternary data sets it is important to ensure that the obtained
parameters cannot yield extra two and three phase regions where these do not
exist.

PARAMETER ESTIMATION FROM LIQUID-LIQUID EQUILIBRIUM DATA

In this section we discuss various aspects of determining parameters in the

models shown in Table 1 from liquid-liquid equilibrium data. The parameter
60

estimation is carried out by minimizing an objective function using a minimiza-

tion procedure.
The choice of minimization procedure influences the rate of convergence
and the amount of computer time needed. Excellent minimization procedures
are incorporated in the standard program packages of most major computer
centers, and it is beyond the scope of this work to review these procedures.
The choice of objective function influences what the attained parameters
represent, e.g. whether certain parts of the phase diagram are reproduced
better than others.
At the end of this section we describe the parameter estimation program
used for obtaining the results shown at the end of this work.

Objective functions

There are two different, main strategies for obtaining parameters I (rl, I-~,
TV, . ..) from liquid-liquid equilibrium data at constant temperature and pres-
I* xi*, . ..). Ex p ressed in terms of the least-squares principle
sure (x:, XT,, .. .. x1,
they are:
(1) Minimization of activity differences according to eqn. (5):
k i

c c Wik [x:ky:k (x:k; r ) - xii 7:; (xji ; r)] * (11)

i = 1, 2, . . . . N (components)
k = 1, 2, . . .. M (tie lines) .:
Wik is the weight associated with component i and tie line k; xfh isthe’experi-
mental mole fraction of component i in phase I at tie line k. The calculated
activity coefficients depend on the experimental mole fractions and the
parameters.
(2) Minimization of the distances between experimental mole fractions
Xijk and calculated mole fI?KtiOIW iijk :

k i j

c Cc Wijk[xijk -k?ijk(r)12 (12)

i = 1, 2, .. .. N (components)
j = I, II (phases)
k = 1, 2, . . . . M (tie lines)
Wijk iS the weight associated with component i in phase j at tie line k. Xijk iS
the mole fraction of component i in phase j at tie line k. The calculated mole
fractions depend on the parameters.
For binary data at fixed temperature and pressure, there are two indepen-
dent measurements: xi and x1I1. This enables the determination of a maximum
of two parameters from one binary tie line. It is not possible to determine all
 61

of the parameters in the NRTL and Black equations from binary mutual
solubility data alone.
A ternary liquid-liquid equilibrium data set gives for a number of tie lines
the concentration of each component in each phase, i.e. xi, xa, xi’ and x:, at
constant temperature and pressure. If a data set contains M tie lines, a maxi-
mum of 3 M parameters r can be determined from the data set.

Objective functions in terms of activities

Almost all authors describing liquid-liquid equilibrium calculations use

objective functions in terms of activities. The objective function indicated by
eqn. (11) with Wih = 1 is minimized in most references:

F(r) =Ci; [x&.& -x:;$:,]” (13)

However, other objective functions based on the isoactivity criterion have
been used. The most common is

F(r) =C 2 [ln(xf,yf,) - ln(x:Z,yii)12 (14)

Objective function (14) has the property that it weights small and large
activities equally, since it operates on the ratio of rather than the difference
between activities. An objective function of this type will in the following
be termed relative, whereas objective function (13) is absolute. Another rela-
tive objective function is:

F(r) =g & [(afk - ag)/(a& + af:)12 (1W

aik = YikXik

Objective functions (14) and (14a) result in nearly identical parameter

sets r _ A sometimes used (F%inger and Bittrich, 1974) variant of objective
functions based on activities is:
k i

F(r) =cc [(h afk - h af:)2xjkxft, + (h a:k - h a:1k)2] (15)

i=2,3
where component 1 is a component which is miscible with both components
2 and 3.
Objective function (15) emphasizes the region in the vicinity of the plait
point.
Objective functions in terms of mole fractions
All objective functions based on the isoactivity criterion suffer from the
serious disadvantage that the minimization of activity objective functions ’
62

does not guarantee minimization of (weighted) distances between the experi-

mental and calculated mole fractions. Most often this is what is wanted.
Objective functions stated in terms of concentrations express directly the
desired goal: to represent as accurately as possible the experimental tie lines.
Renon et al. (1971), Varhegyi and Eon (1977), and Anderson and Prausnitz
(1978) use objective functions of type (12):

F(r) =55 t: Wijk[Xijk -2,,k]2 (16)

i = 1, 2, . . . . N; j = I, 1,I; k = 1, 2, . . .. M
Computationally, objective function (16) is more complicated than the
activity objective functions. It contains the computed mole fractions ?ijh
which, for a given current parameter set r , must be predicted for each tie
line k using either the isoactivity method or minimization of Gibbs energy.
In using objective function (16) for ternary systems one must for each
current parameter estimate r = r’, match each experimental tie line k with a
predicted one. The predicted tie line may be chosen (a) by setting one con-
centration in one phase equal to the corresponding experimental value (Renon
et al., 1971) or (b) by minimizing:

(Varhegyi and Eon, 1977).

Using criterion (a) and (b) respectively, objective function (16) can be
more precisely stated as

F(f ) =A5 C Wijk [Xijk -2ijkJ2

Wa)
,.
X22k = X22k; i = 1, 2, 3; j = I, II, k = 1, 2, . .. . M

F(r) =25 mini: i; Wijk[xijlr -3Tijk12 (16b)

i = 1, 2, . . . . N; j = I, II, k = 1, 2, . . .. M
It is computationally more complicated to use eqn. (16b) than (16a). Ob-
jective function (16b) has the advantage that we do not need to match a pre-
dicted tie line precisely to one experimental mole fraction, which - perhaps -
is erroneous.
Objective functions (16), (16a) and (16b) are absolute. The corresponding
relative objective functions may be obtained by replacing with [xijk - $ijk12

[ln Xijk - h zijk12 or [(Xijk -fijk)/(Xijk + 3sijh)12.

Activity objective functions deriving from eqn. (11) suffer from the dis-
advantage that they usually do not directly correspond to the desired objec-
tive. On the other hand, objective functions in terms of mole fractions, eqn.
 63

(16), require a much larger computational effort. The time-consuming step is

the matching of predicted tie lines with experimental ones, for example via
eqn. (17).
It is possible, by a linear transformation, to express eqn. (17) analytically
in terms of activities. The result is an objective function which computation-
ally is only slightly more complicated than the activity objective function
and yet, within the imposed linearization, corresponds to the desired objec-
tive: to minimize F(t ) as expressed by eqn. (16). The procedure is indicated
for ternary systems in Appendix B. It may be extended to the general multi-
component case (Michelsen, 1978).

Parameter estimation method used in this work

A computer program was developed for correlating the data of the liquid--
liquid equilibrium data bank described in Part I of this series. The objective
is to represent phase equilibrium compositions as well as possible using models
such as NRTL and UNIQUAC, and hence mole fraction objective functions
related to eqn. (16) are used.
The procedure is similar, but not identical, to ,tbat of Varhegyi and Eon
(1977). The two procedures differ especially in the following points:
(a) Near the minimum of the objective function, we use the Levenberg-
Marquardt method rather than a direct search routine; this reduces the com-
puter time.
(b) A qualified initial parameter guess is generated by the program.
Consider as an example the estimation of five parameters in the UNIQUAC
equation: ua2, u33r u12, u13 and ua3 (ull is fixed arbitrarily). The experi-
mental tie lines are for example those indicated by the solid lines on Fig. 4,
although the procedure is not limited to type one systems. The parameter
estimation takes place in four steps:
(1) Construction of initial parameter guess;
(2) adjustment of u22, u12, and ~23;
(3) adjustment of all five parameters using eqn. (16a);
(4) final estimation of all five parameters using eqn. (16b).

1. Initial parameter guess

The lowest tie line is projected to the 1-3 base line as shown in Fig. 4.
The approximated mutual solubilities of binary l-3 is used to estimate u3a
and u13 (ull is arbitrarily fixed). Next, u22, u12 and ~23 zre all equated to
cull+ u33 +u13)/3.

2. Adjustment of u22r u12 and u23

The initial parameter estimate is used to generate a series of tie lines, which
may perhaps look like the dashed lines on Fig. 5. The isoactivity method is
used for generating the tie lines. This step incorporates using the check for
convexity described in Appendix A and a check for absence of binodal dis-
64

Fig. 4. Construction of initial parameter estimate. The full lines are the experimental tie
lines.

Fig. 5. Network of computed tie lines (dashed lines) and experimental tie lines (full lines).

continuities (see “Computation of binodal curves”). Each of the experimental

tie lines k is matched, by interpolation, to the calculated one for which
.
x22k = x22ka The resulting values for ?Zijk are inserted into eqn. (16a). The
objective function is evaluated, and the Nelder-Mead search procedure is used
to generate improved values for the parameters uz2, u12 and u2s. For each
new set of parameters a new series of tie lines must be generated.

3. Adjustment of uz2, uss, u12, u13 and uz3

This resembles step 2. Here all five parameters are evaluated, and since we
are near the optimum, the Nelder-Mead iteration procedure is replaced by
the Levenberg-Marquardt procedure (An IMSL library routine).

4. Final parameter estimation

All five parameters are further adjusted using the Levenberg-Marquardt
iteration procedure. The difference between this step and step 3 is that now
we use objective function (16b). That is, for each experimental tie line k we
find, by linear interpolation among the generated series of calculated tie lines,
that calculated tie line for which

is minimum. The obtained values of are inserted into eqn. (16b).

?ijk

The experience with the above procedure has been favorable in the sense
that convergence has been attained in all tested cases and that the obtained
parameters furnish reasonable binodal curves. Examples of this will be shown
in the following section. The computer time required is normally between
15 and 30 CPU sec. per ternary diagram on the IBM 370/165. The computer
program will be described in greater detail in (Wrensen, 1979).
 65

RESULTS OF THIS WORK

Table 2 shows the systems used in the following investigations regarding

the correlation of liquid-liquid equilibrium data:
(1) The ability of a local composition model (here UNIQUAC) in represent-
ing ternary data sets as a function of the number of binary interaction
parameters which are allowed to vary.
(2) The influence of objective function weighting factors on the prediction of
solute distribution ratios.
(3) A study of the UNIQUAC- and NRTL-equations for correlating liquid-
liquid equilibria:
(3a) Correlation of ternary data sets.
(3b) Prediction of ternary liquid-liquid equilibria from binary data.
(3~) Reproduction of binary T-x diagrams.
The experimental tie lines corresponding to each of the data sets of Table 2
are shown in Fig. 6. Note that relatively many of the data sets exhibit unusu-
ally complex tie line behavior (see S9rensen et al. (1979), Fig. 3).

Number of binary interaction parameters (UNIQUAC)

As already discussed, we may fit a total of either 5 or 6 binary interaction

parameters of the UNIQUAC equation to a ternary liquid-liquid equilibrium
data set. For a type one system, one may choose to prefix two of these param-
eters from the mutual solubility data of the (partly) immiscible binary; this
leaves 3 or 4 parameters to be adjusted. For a type two system, one may
choose to prefix four parameters from binary mutual solubility data, which
leaves 1 or 2 parameters to be determined.
Table 3 shows the residuals, F, one obtains with the different number of
parameters :

min 5 h (Xijk -~ijk)2/6 M (18)

i = 1, 2, 3; j = I, II; k = 1, 2, a.., M

The residual F corresponds to the objective function (16b) with Wijn = 1,

all i, j, k.
As one might expect, allowing 6 parameters to vary gives the smallest
residual, and 3 (or 1) gives the largest. This may lead one to conclude that
one should fit 6 (or perhaps 5) parameters to a ternary data set. However,
objective function (16b) with Wijk = 1 has the property that even large rela-
tive errors on small concentrations are overlooked. This means, for example,
that for nearly immiscible binaries, the mutual solubilities may not even be
reproduced correctly within a factor of 100. By fitting 4 parameters for type
one systems (and 2 for type two) the immiscible binary mutual solubilities
can be reproduced correctly at the expense of increasing the residual (18) on
66

TABLE 2
Ternary data sets used in the investigation (Temperature 25’C)

Data set Type of Components (l)--(2)--(3) References

number system

25 Ethane, 1,1,2-trichloro-2- Treybal, R.E., et al., Ind. Eng.

Propanone-water Chem., 38 (1946) 817
37 Ether, diisopropyl-2- Frere, F.J., Ind. Eng. Chem.,
Propanol-Water 41 (1949) 2365.
265 Water-l-Propanol-Heptane Vorobeva, A.I., et al., Russ. J.
Phys. Chem. (Engl. Transl.),
41 (1967) 602
288 1 Cyclohexane-Benzene- Henty, C.J. et al., J. Appl. Chem.,
Furfural 14 (1964) 148
440 1 Phenol-2-Propanol-Water Rusanov, A.I., Vestn. Leningr.
Univ. Ser. Mat. Fiz. Khim.,
14(4) (1959) 132
464 1 Methane, trichloro-2- Reinders, W., et al., Rec. Trav.
Propanone-Water Chim., 66 (1947) 573
508 1 Benzene-2-Propanol-Water Nikurashina, N.I. et al., J. Gen.
Chem. USSR. (Engl. Transl.),
43 (1973) 2093.
509 1 Water-a-Propanol-Cyclo- do.
hexane
546 1 Furfural-2-Propanol-Water Krupatkin, IL. et al., J. Gen.
Chem. USSR. (Engl. Transl.),
40 (1970) 10
556 l-Butanol-Methanol-Water Prochazka, 0. et al., Coll. Czech.
Chem. Commun., 40 (1975) 781
559 Cyclohexane-Cyclopen- Takeuchi, S. et al., J. Chem. Eng.
tane-Methanol Jpn., 8 (1975) 248
569 Methanol-Acetic acid, Sugi, H. et al., J. Chem. Eng.
methyl ester-Cyclohexane Jpn., 9 (1976) 12
676 1-Butanol, 3-methyl- Coull, J. et al., J. Phys. Chem.,
l-Propanol-Water 39 (1935) 967.
289 Cyclohexane-Pentane, Henty, C.J. et al., J. Appl. Chem.,
2,2,4-trimethyl-Furfural 14 (1964) 148
551 Furfural-l-Propanol, Krupatkin, I.L. et al., J. Appl.
2-methyl-Water Chem. USSR. (Engl. Transl.),
42 (1969) 836
649 2 Benzene-l-Propanol, 2- Alberty, R.A. et al., J. Phys.
methyl-Water Chem., 49 (1945) 4
660 2 Benzene-l-Butanol-Water Washburn, E.R. et al., J. Phys.
Chem., 48 (1944) 241
 67

the average by only 50%. In addition, the savings in computer time are sub-
stantial.
Figure 6 shows the binodal curves calculated when 4 (or 2) parameters are
determined from the ternary data set and 2 (or 4) are prefixed from binary
solubility data. The tie lines shown are the experimental ones. It is seen that
in most cases the binodal curves are very near the end points of the tie lines
and - as expected - the boundary conditions given by the binary mutual
solubilities are satisfied. In some cases (e.g. Fig. 6, data set 37) there is an
apparent discrepancy between the binary l-3 and the ternary data sets.

Objective function weighting factors

Objective functions in terms of activities, e.g. eqn. (14a), rather than com-
positions, e.g. eqn. (16b), do not correspond to the desired objective. This is
shown in Table 4. The F-residual given by eqn. (18), when the activity objec-
tive function is used in the data correlation, is much larger than the corre-
sponding F-residual using the composition objective function. Conversely,
the activity residual given by

is not very sensitive to whether objective functions (14a) or (16b) are used in
the data correlation. This is fortunate and supports the arguments in favor of
using composition objective functions.
Large relative errors in small concentrations are neglected using objective
function (16b) with Wijk = 1.
In type one systems we have already taken the small solvent (components
1 and 3) concentrations into account by reproducing correctly the solvent
concentrations on the base line. Now we need not pay too much more atten-
tion to small solvent concentrations measured in the ternary diagram; this is
an advantage, since these concentrations often are determined with large rela-
tive experimental errors. For practical purposes such as design of extraction
processes, small solute (component 2) concentrations should also be given
special consideration. Table 5 shows the error in the solute distribution ratio 0
obtained when objective function (16b) with W’ij, = 1 is used. In Table 5,

k = 1, 2, . . .. M (tie lines)

Pk = x21kh22k and P^k= kZlkis22k

The column W = 0 corresponds to objective function (16b) with Wi,, = 1. It

is seen that a low F-residual as given by eqn. (18) does not guarantee a good
68
 2 2

1 3 1 3
269 551
A A
1 3
649

Fig. 6. Experimental tie lines and calculated binodal curves for the ternary data sets of this study. The binodal curves are calculated
using UNIQUAC with 4 (type one) or 2 (type two) parameters. The parameters corresponding to the immiscible binaries are predeter-
mined from binary data only.
g
70

TABLE 3
Residuals (eqn. (18)) for different number of parameters (UNIQUAC)

Data set Type of Number of parameters

number system
3 4 5 6

25 1 0.60 0.59 0.48 0.46

37 1 1.19 1.10 1.07 0.49
265 1 5.59 2.28 1.99 1.24
288 1 0.18 0.18 0.11 0.10
440 1 2.00 1.24 1.17 0.77
464 1 0.92 0.91 0.78 0.77
508 1 1.90 0.91 1.19 0.74
509 1 1.08 1.01 1.01 1.00
546 1 1.61 1.59 0.75 0.52
556 1 1.10 0.75 1.09 0.39
559 1 0.53 0.52 0.45 0.44
569 1 0.48 0.30 0.41 0.27
576 1 0.46 0.46 0.38 0.23

Data set Type of Number of parameters

number system
1 2 5 6

289 2 0.23 0.16 0.09 0.09

551 2 2.44 1.29 1.21 1.12
649 2 0.23 0.27 0.22 0.13
660 2 0.49 0.39 0.27 0.24

Average 1.24 0.82 0.75 0.53

TABLE 4
Comparison of residuals obtained using activity and concentration objective functions

Data set Activity objective function Concentration objective function

number (eqn. (14a)) (eqn. (16b))

Activity Concentration Activity Concentration

residual residual residual residual
(eqn. (18a)) (eqn. (18)) (eqn. (18a)) (eqn. (18))

37 10.07 3.78 16.76 1.10

288 1.09 0.36 1.95 0.18
508 25.75 14.50 39.86 0.91
546 2.85 1.65 10.72 1.59
559 1.02 2.21 5.52 0.52
676 7.10 4.77 18.22 0.46

Average 7.98 4.55 15.51 0.79

 71

TABLE 5
Residuals for different objective functions when four binary UNIQUAC parameters are
fitted to ternary type one systems. The residuals F are given by eqn. (18)

Data set Resid- Objective function

number ual
w=o W=lO w= 100 w = 1000 Relative eqn.
form (16a)
of (16b)

25 0.59 00.60 0.74 0.94 1.02 0.60

2.68 2.51 1.97 1.46 4.55 2.58
1.10 1.12 1.23 1.38 1.64 1.18
37 G 21.63 17.07 10.59 9.33 25.14 10.48
2.28 2.58 4.93 5.25 2.50 3.11
265 &.I 33.47 30.98 16.88 15.86 47.31 28.30
288 F 0.18 0.18 0.19 0.24 0.21 0.19
A@ 2.60 2.57 1.88 1.67 2.04 4.09
464 F 0.91 0.94 1.33 1.65 1.71 0.96
4 8.52 7.18 5.26 5.02 8.00 6.00
509 F 1.01 1.35 2.67 2.96 1.84 1.09
4 42.59 33.64 28.70 28.24 32.23 46.53
556 F 0.75 0.89 1.07 1.17 2.86 1.05
A0 41.88 17.88 5.37 3.40 13.79 11.89
559 F 0.52 0.53 0.68 1.25 0.60 0.59
AP 5.44 4.36 1.61 0.92 3.87 2.85
569 F 0.30 0.31 0.43 0.52 0.49 0.37
W 14.82 10.29 3.82 2.12 2.42 17.21
616 F 0.46 0.50 0.62 1.11 0.78 0.55
43 27.01 12.77 8.27 5.84 21.38 8.01
Average F 0.81 0.90 1.39 1.65 1.37 0.97
values fW 20.06 13.93 8.44 7.39 16.07 13.79

fit to the distribution ratio p. This is illustrated in Fig. 7, where the calculated
tie line is quite close to the experimental one. Upon closer examination, how-
ever, the predicted value of /_Iis erroneous by more than a factor two. We
have tried to remedy this unfortunate situation in three different ways:
(a) By using the relative objective function corresponding to (16b) with
Wijh = 1.
(b) By using objective function (16a) which fixes the solute concentration
at one end, 32sah= Xs2h.
(c) By weighting the solute (component 2) concentrations highly. In eqn.
(16b) we use

1 + lfijk -XijkI. w 2
(20)
Xijk
72

-_ --_
A
1 3
Fig. 7. Example of tie line which is obtained using an absolute objective function, eqn.
(16b), and which exhibits large error in the solute distribution ration - Experi-
mental; - - - . - - calculated.

fori=2; j = I, II; h = 1, 2, . . .. M
Wijh = 1

for i = 1, 3; j = I, II, k = 1, 2, . . . . M
W is an arbitrary, positive number.
Method (a) weights small and large concentrations equally and gives on the
average a smaller Ap than objective function (16b) with Wijk = 1. The dis-
advantage with this method is that small solvent concentrations are weighted
as highly as small solute concentrations, even though the experimental scatter
on the small solvent concentrations is often relatively large. In some instances
this even results in a larger Ap than before.
Method (b) gives a significant improvement in A/3 because the small solute
concentrations in solvent 3 are kept on their experimental values. However,
the. objective function will still not emphasize small solute concentrations
in solvent 1.
Method (c) weights the solute concentrations in objective function (16b)
by one plus a multiple of the relative deviation of the solute concentrations.
Table 5 shows that for W = 100, a significant decrease in A/I is obtained with-
out significant increase in the F-residual.
For type two systems, it has to be decided whether component 1 or 2 or
perhaps both should be selected as solutes. Then method (c) may also be
applied to these systems.

Correlation of ternary systems using UNIQUAC and NRTL

Table 6 shows F-residuals given by eqn. (18) when 4 NRTL or UNIQUAC

parameters are fitted to type one systems and 2 to type two. For this study,
we have in all cases used objective function (16b) with Wijh = 1. Figure 6
shows the binodal curves calculated using UNIQUAC.
In terms of overall performance, the UNIQUAC equations is found to
 73

TABLE 6
Residuals (eqn. (18)) from the UNIQUAC- and NRTL-equations
-__. ~ _._
Data set Correlation Prediction
number -..____. -----
UNIQUAC NRTL UNIQUAC NRTL

(Y = 0.1 o( = 0.2 o! = 0.3 CY= 0.2

~~______ ___~ __.~__
25 0.59 0.58 0.79 1.46 4.32 2.94
37 1.10 0.78 0.91 1.47 4.05 4.52
265 2.28 4.71 2.45 7.58
288 0.18 0.18 0.19 0.21
440 1.24 1.08 0.93 0.98
464 0.91 0.87 1.12 1.79 6.06 2.05
508 0.91 4.72 0.55 1.06 * l

509 1.01 1.19 2.25 3.49

546 1.59 0.46 0.27 0.29
556 0.75 0.61 0.63 0.57
559 0.52 0.57 0.57 0.57
569 0.30 0.45 0.44 0.41 2.80 2.75
676 0.46 0.54 0.27 0.26
289 0.16 0.30 0.23 0.21 0.31 0.33
551 1.29 1.70 3.32 4.55
649 0.27 1.98 0.42 *
0.64 3.16
0.65 l
660 0.39 1.32 0.57 1.91
Average 0.82 1.30 0.95 1.76

l The back-calculated binodal curve exhibits discontinuities as shown in Fig. 3.

yield slightly better results than NRTL with (Y= 0.2. NRTL with (Y= 0.1 and
CY= 0.3 results in somewhat poorer correlation. (NRTL with cx = -1, i.e. the
LEMF equation, was also found to yield poorer results than NRTL with
QI= 0.2).
If (Y in the NRTL equation were fitted to each individual ternary diagram,
the correlation would, obviously, be somewhat improved. This is not shown
here.

Prediction of ternary diagrams from binary information

We now find 5 or 6 UNIQUAC or NRTL (e = 0.2) parameters from experi-

mental vapor-liquid and liquid-liquid equilibrium data for the three binaries
and see how well we can predict the ternary liquid-liquid equilibrium dia-
gram.
When 6 binary parameters are used, there is no doubt about the procedure:
each binary data set furnishes two parameters. When 5 parameters are used,
two different routes are possible for a type one system:
(A) Choose arbitrarily ull. Find u33 and u13 from the immiscible binary.
74

Find uz2 and u12 from vapor-liquid equilibrium data for the l-2 binary.
Find u2s from vapor-liquid equilibrium data for the 2-3 binary. (Method A).
(B) Choose arbitrarily u1 1. Find us3 and u13. Find u22 and ~23. Find u12.
(Method B).
In general, methods A and B give different results. Furthermore, using
5 parameters is on the average found to yield poorer results than using
6 parameters. For these reasons, the predictions shown in Table 6 were
carried out using 6 parameters. The vapor-liquid equilibrium data used to
establish interaction parameters for the miscible binaries were obtained from
the literature and were in most cases less than 30°C removed from the tem-
perature of the ternary diagram.
The results indicated for NRTL and UNIQUAC are too few to make a con-
clusive comparison with respect to predicting ternary liquid-liquid equilibri-
um diagrams from binary data. However, it may be seen that the residual

1A 3
464 569

1 3 1 3
289 649

Fig. 8. Experimental tie lines and predicted binodal curves for several ternary data sets.
The predictions are carried out using UNIQUAC with binary interaction parameters ob-
tained from binary LLE and VLE data.
for prediction is 2-10 times that for correlation. The binodal curves predicted
using UNIQUAC are shown in Figure 8.
Blanks in the two columns of Table 6 indicate that it was not possible to
find the necessary vapor-liquid equilibrium data, and an * indicates that the
parameters found gave rise to breaks in the calculated binodal curve as
described in the section “Calculation of Binodal Curves”.
Prediction of multicomponent liquid-liquid equilibria from information
on binaries will be further covered in Parts III and IV of this series.

Reproduction of binary T-x diagrams

The temperature dependence of the parameters in the NRTL and UNIQUAC

equations is now investigated for a few immiscible binaries. The binary tetra-
hydrofuran(l)-water(2) exhibits a closed phase envelope in the T-X dia-
gram. Figure 9 shows the UNIQUAC-parameters ui2 and u2i as a function of
temperature. The parameters are calculated from the experimental T-x dia-
gram at temperatures varying from the lower to the upper critical solution
temperature.
It is seen that the temperature dependence is almost linear in a large tem-
perature range. Figure 10 shows the back-calculated T-x diagram by using
the parameters on the lines AB and CD (“linear parameters”). Figure 10 also

D K

70 80 90 100 110 120 130 140

Temperature %

Fig. 9. UNIQUAC parameters as functions of temperature for the system tetrahydrofuran-

(1 )-water( 2).
 140. 140.

130. 130.

120. 120.

P P
,o 110. Iy 110.
D
;
;
;
;
;
;100.
EioO.
: :

90. 90.

90. 80.

70. 70.
.o .I .2 .3 A .5 .l .2 .3 .4 .5

Mole fraction Of component 1 Mole fraction of component 1

Fig. 10. T-X diagram for tetrahydrofuram 1 )-water( 2) using UNIQUAC. - experi-
mental solubility ; -- - - solubility calculated with linear parameters; - . - . solubility
calculated with constant parameters.

Fig. 11. T-X diagram for tetrahydrofuran(l)-water(2) using NRTL with (Y = 0.2.
- experimental solubility; - - - solubility calculated with linear parameters;
.-.-. solubility calculated with constant parameters.

Fig. 12. T--x diagram for 1-butanol(l)-water(2). - experimental solubility;

- - - solubility calculated with linear parameters (UNIQUAC); . - . - . solubility calcu-
lated with linear parameters (NRTL, a! = 0.2).
 77

70.

60. I

I
50.
Y 1

z 40.
#/
;-

% 30. !,
E I’
,I
F 20.
I’
11
10. f

0.
-I~
.0004 .0006 98.9 99.0 99.2 99.4 SD.9 99.9

Mole fraction of component 1

Fig. 13. T-x diagram for benzene( 1 )-water( 2). - experimental solubility;
--- solubility calculated with linear parameters (UNIQUAC); . - - . solubility calcu-
lated with linear parameters (NRTL, cx = 0.2). At high mole fractions of component 1, the
experimental and calculated curves coincide.

shows the back-calculated T-x diagram by using the parameters at points B

and D in the whole temperature interval (“constant parameters”).
Linear parameters describe the T-x diagram better than constant param-
eters, but the representation is in no case quantitative.
Figure 11 shows the same results as Fig. 10, but using the NRTL equation
with CI= 0.2. Note that in using parameters which vary linearly with tempera-
ture, one can arbitrarily choose two temperatures where the calculated solub-
ility matches the experimental one.
Figures 12 and 13 show T-x diagrams for two other binaries and predicted
solubilities using UNIQUAC and NRTL (CY= 0.2) with linear parameters.
One may conclude that for quantitative description of binary immiscibil-
ities it is essential to use parameters determined from experimental data at
the relevant temperature.

CONCLUSIONS

The surveyed literature and the present study of three binary and 17
ternary liquid-liquid equilibrium data sets show that complex and intricate
problems of numerical and analytical nature are encountered in correlating
liquid-liquid equilibria. In particular, the solubility-temperature dependence
and the existence of false solutions to the equilibrium criterion create sub-
stantial problems.
The equilibrium condition stated in terms of minimization of Gibbs energy
or the isoactivity criterion coupled with a check for convexity should be used
in calculating liquid-liquid equilibrium compositions from a given model.
Erroneous solutions are avoided, but there is no guarantee against false solu-
tions corresponding to local minima in the Gibbs energy.
In calculating model parameters, objective functions stated in terms of
concentrations rather than activities should be used. We show that by doing
so it is possible to greatly improve the fit to experimental concentrations
without seriously affecting the isoactivity criterion.
The activity coefficient models used today for correlating liquid-liquid
equilibria were mainly developed for vapor-liquid equilibria. Unfortunately,
these models, including NRTL and UNIQUAC, are not fully successful for
liquid-liquid equilibria. In type one ternary systems, for example, it is often
not possible to represent in the entire two phase region both the binodal
curve and the solute equilibrium ratios with sufficient accuracy for extraction
design purposes. When objective functions which emphasize the solute con-
centrations are used in the data correlation, the situation is somewhat im-
proved.
In the correlation of type one ternary data sets, the two (UNIQUAC or
NRTL) parameters representing the immiscible binary were prefixed from
binary data; in type two systems the four parameters representing the two
immiscible binaries were similarly prefixed. Fitting four (type one) or two
(type two) parameters to ternary data sets results in only slightly increased
deviation between experimental and calculated mole fractions as compared
with fitting six parameters. Several advantages associated with using four
(or two) parameters are discussed.
Due to its better built-in temperature dependence and a slightly better
performance in general, we prefer to use UNIQUAC rather than NRTL in
correlating liquid-liquid equilibria. However, models more suitable than
UNIQUAC for this purpose are still needed.
The prediction of multicomponent liquid-liquid equilibria from binary
information only is not given in-depth attention in this article. This subject
will be further covered in Parts III and IV of this series.
The conclusions one may draw regarding the correlation of liquid-liquid
equilibria are very sensitive to the particular data sets chosen for study.
Therefore, experience gained from studying a few - say four or five -
ternary diagrams is insufficient for definite conclusions regarding, for example
the capabilities of a given model. Hopefully, the 20 data sets of this work
fully cover the types of data found in the literature for the components of
interest in this study.

ACKNOWLEDGEMENTS

The authors thank Michael L. Michelsen for many valuable suggestions.

This work was supported by Bundesministerium fiir Forschung und Tech-
nologie and by Statens teknisk-videnskabelige Forskningsrdd.
 79

SUPPLEMENT

A listing of the computer program used for correlating liquid-liquid

equilibrium data is available from the authors upon request.

NOTATION

activity
; fugacity
F objective function or residual
molar Gibbs energy
a”G
GE
molar Gibbs energy of mixing
excess Gibbs energy
number of moles
; pressure
e pure component vapor pressure
R the gas constant
T temperature
7.4 NRTL or UNIQUAC parameter
V molar volume
W objective function weighting factor
X liquid mole fraction
z overall composition, mole fraction

Greek symbols

NRTL non-randomness parameter

solute distribution ratio
7 activity coefficient
12 chemical potential
r NRTL or UNIQUAC parameter

Superscripts

I, II phases
E excess
,.
estimated value
0 false solution

Subscripts

1, 2, 3 components
i, j components
i phase
k tie line
T total
80

REFERENCES

Abrams, D.S. and Prausnitz, J.M., 1975. AIChE J., 21: 116.
Anderson, T.F. and Prausnitz, J.M., 1978. Ind. Eng. Chem. Process Des. Develop., 17: 56
Black, C., 1959. AIChE J., 5: 249.
Councell, J.F., Lees, E.B. and Spencer, P.J., 1970. DCS Report 9, December 1970.
Fabrics, J.-F., Gustin, J.-L. and Renon, H., 1977. J. Chem. Eng. Data, 22: 303.
Fanger, H. and Bittrich, H.-J., 1974. Chem. Tech., 26: 219.
Greiner, E., Rosier, F., Kepp, W. and Donath, E., 1974. Chem. Tech., 26: 364.
Guffey, C.G. and Wehe, A.H., 1972. AIChE J., 18: 913.
Hala, E., 1972. Ind. Eng. Chem. Process Des. Develop., 11: 638.
Hand, D.B., 1930. J. Phys. Chem., 34: 1961.
Heidemann, R.A., 1974. Hydrocarbon Proc., (Nov. 1974) 167.
Heidemann, R.A. and Mandhane, J.M., 1973. Chem. Eng. Sci., 28: 1213.
Heil, J.F. and Prausnitz, J.M., 1966. AIChE J., 12: 678.
Heinrich, J. and Dojcansky, J., 1975. Collect. Czech. Chem. Commun., 40: 2221.
Hiranuma, M., 1974. Ind. Eng. Chem. Fundam., 13: 219.
Hiranuma, M., 1975. J. Chem. Eng. Japan, 8: 69.
Joy, D.S. and Kyle, B.G., 1969. AIChE J., 15: 298.
Katayama, T., Kato, M. and Yasuda, M., 1973. J. Chem. Eng. Japan, 6: 357.
Kretschmer, C.B. and Wiebe, R., 1954. J. Chem. Phys., 22: 1697.
Leach, M.J., 1977. Chem. Eng., (May 1977) 137.
Lu, B.C.-Y., Yu, P. and Sugic, A.H., 1974. Chem. Eng. Sci., 29: 321.
Margules, M., 1895. Sitzber. Akad. Wiss. Wien. Math. Naturw. Klasse (II), 104: 1243.
Marina, J.M. and Tassios, D.P., 1973a. Ind. Eng. Chem. Process Des. Develop., 12: 67.
Marina, J.M. and Tassios, D.P., 197313. Ind. Eng. Chem. Proc. Des. Develop., 12: 271.
Mattelin, A.C. and Verhoeye, L.A.J., 1975. Chem. Eng. Sci., 30: 193.
Michelsen, M.L., 1978. Private Communication.
Nagata, I., Nagashima, M. and Ogura, M., 1975. Ind. Eng. Chem. Process Des. Develop.,
14: 500.
Nagata, I., Kita, H. and Nakamiya, Y., 1978. Fluid Phase Equilibria, 1: 267.
Newsham, D.M.T. and Vahdat, N., 1977. Chem. Eng. J., 13: 33.
Nitta, T., Takeuchi, S. and Katayama, T., 1974. Chem. Eng. Sci., 29: 2213.
Novak, J.P., Vonka, P., Suska, J., Matous, J. and Pick, J., 1974. Collect. Czech. Chem.
Commun., 39: 3593.
Palmer, D.A. and Smith, B.D., 1972. Ind. Eng. Chem. Process Des. Develop., 11: 114.
Peng, D. and Robinson, D.B., Can. J. Chem. Eng., 54: 595.
Ragaini, V. and Baldini, L., 1973. Ing. Chim. Ital., 9: 1.
Ragaini, V., Santi, R. and Carniti, P., 1974. Chim. I’Ind., 56: 687.
Redlich, 0. and Kwong, J.N.S., 1949. Chem. Rev., 44: 233.
Redlich, 0. and Kister, A.T., 1948. Ind. Eng. Chem., 40: 345.
Renon, H., 1966. Ph.D. Thesis, University of California, Berkeley.
Renon, H. and Prausnitz, J.M., 1968. AIChE J., 14: 135.
Renon, Ii., Asselineau, L., Cohen, G. and Raimbault, C., 1971. Calcul sur Ordinateur des
Equilibres Liquide-Vapeur et Liquide-Liquide, Technip, Paris.
Schulte, H.-W., 1977. Darstellung von Dampf-Fliissig- und Fliissig-Fliissig-Gleichgewichten
durch einen generalisierten Ansatz und eine erweiterte Wilson-Gleichung, Ph.D. Thesis,
University of Dortmund, Germany.
Sugi, H., Nitta, T. and Katayama, T., 1976. J. Chem. Eng. Japan, 9: 12
Sorensen, J.M., 1979. Ph.D. Thesis, The Technical University of Denmark.
S@-ensen, J.M., Magnussen, T., Rasmussen, P. and Fredenslund, Aa., 1979. Fluid Phase
Equilibria, 2: 297.
Takeuchi, S., Nitta, T. and Katayama, T., 1975. J. Chem. Eng. Japan, 8: 248.
 81

Tsuboka, T. and Katayama, T., 1975. J. Chem. Eng. Japan, 8: 404.

Van Laar, J.J., 1910. Z. Physik. Chem., 72: 723.
Van Ness, H.C., 1964. Classical Thermodynamics of Non-Electrolyte Solutions, Pergamon
Press, New York.
Van Ness, H.C., Byer, S.M. and Gibbs, R.E., 1973. AIChE J., 19: 238.
Varhegyi, G. and Eon, C.H., 1977. Ind. Eng. Chem. Fundam., 16: 182.
Vera, J.H., Sayegh, S.G. and Ratcliff, G.A., 1977. Fluid Phase Equilibria, 1: 113.
Wilson, G.M., 1964. J. Am. Chem. Sot., 86: 127.

APPENDIX A. A CHECK FOR MINIMA IN THE GIBBS FUNCTION

Equations (2) and (5) are satisfied not only by minima in the Gibbs func-
tion, but also by maxima and saddle points. However; an equilibrium condi-
tion is characterized by a minimum as indicated in eqn. (1). For a given
model with given parameters, it is possible to verify that a given set of mole
fractions satisfying eqns. (2) or (5) correspond to a minimum.
We require that for a heterogeneous system to be stable, the individual
homogeneous phases must also be stable. The criterion of stability for a
homogeneous phase with N components is that the matrix with the elements

a2G/axiaxj i,j=l,Z , .. ..N-1 (Al)

be positive definite. This corresponds to the Gibbs energy surface being con-
vex at the concentration of interest. That is,

Binary : a2G/a3Cf > 0

Ternary: a2G/axf > 0; a2G/ax; > 0
a2G a2G (-42)
ax; axlax
>o
a2G a2G
axlax a&j

The derivatives of the Gibbs energy are readily evaluated from the model
for the Gibbs energy or from the associated expression for the activity. A use-
ful expression, derived on the basis of equations shown in (Van Ness, 1964,
p. 90), is:

a2G _ a lnai a InaN a ha. I +a haN

(A3)
axiaxi axj ’ axi ax, ax,

where the derivatives on the right-hand side are evaluated as if the mole frac-
tions were independent of one another.
82

APPENDIX B. A LINEAR TRANSFORMATION OF ACTIVITY OBJECTIVE FUNC.

TIONS TO CONCENTRATION OBJECTIVE FUNCTIONS

For a given tie line in a ternary phase diagram, consider two experimental
points

x1= (x11, ~ZlrX31) and XI1 =

(x1,, x229 3c32) 031)

(Xii is the mole fraction of component i in phase j). With a given model and a
given set of parameters I = r’, one may calculate a set of activities correspond-
ing to the experimental mole fractions:

a1 = (a,,, a217 &31) and aI1 = ("12>u22,a32) (B2)

(aij is the activity of component i in phase j). These activities in general do

not satisfy the isoactivity criterion.
We wish now to find two points in the neighbourhood of the experimental
points for which the calculated activities a^satisfy the isoactivity criterion:
4 - -II
a --a (B3)
local linearization gives for both a^’and 6”:

a^= u + C Aji(Xi - ai) i = 1, 2, 3 (B4)

where Aji= aa,/ax,
can be evaluated from the model analytically. Condition
(B3) gives:

(B5)
Furthermore

&xi-a;,=0 and (B6)

Equations (B5) and (B6) constitute five linear equations in six unknowns,
2; and 9:‘. This system does not have a general unique solution; however,
there exists a unique “solution” for which the objective

CC Wij(Xij -aij)' i = 1, 2, 3; j = I, II (B7)

is minimum. Now, we do not calculate the 32ij’s corresponding to this mini-
mum, but we calculate the objective (B7) in every data point and minimize
the sum of these by the Levenberg-Marquardt technique.
This procedure has been found to require only slightly more computer
time than does the direct use of activity objective functions. The only pos-
sible disadvantages of this method as compared with objective functions
stated directly in terms of mole fractions stem from the imposed lineariza-
tion, see eqn. (B4). This method warrants further study.