A Method for Prediction of UNIFAC

Group Interaction Parameters

Hugo Edson Gonza`lez, Jens Abildskov, and Raqul Gani
CAPEC, Dept. of Chemical Engineering, Technical University of Denmark, DK-2800 Lyngby, Denmark

Pascal Rousseaux and Brice Le Bert

Processium, 43 Bvd du 11 Novembre 1918, 69622 Villeurbanne Cedex, France

DOI 10.1002/aic.11171
Published online May 1, 2007 in Wiley InterScience (www.interscience.wiley.com).

Group-contribution-based property estimation methods are suitable for obtaining

quick evaluations of phase equilibrium under different conditions of temperature, pres-
sure and composition. One of the best known and successful group-contribution (GC)
based methods for prediction of liquid-phase-activity coefcients in mixtures of organic
compounds is the UNIFAC method. One of the principal limitation of the UNIFAC
method and all other GC based methods is that groups or group-interaction parame-
ters needed for a specic property estimation problem may not be available. The
method presented, is to be called the CI-UNIFAC (Connectivity Index UNIFAC)
method, to predict the missing UNIFAC group-interaction parameters for the calcula-
tion of vapor-liquid equilibrium (VLE). The method is described as the CI-UNIFAC for
predicting missing group interaction parameters, as well as re-estimating known group
interaction parameters, using a set of atom connectivities and their regressed CI inter-
action parameters. The performance of the CI-UNIFAC method with experimental data
is compared, with a reference UNIFAC method, as well as cases where the CI-UNI-
FAC method is used only for the missing UNIFAC group interaction parameters.
2007 American Institute of Chemical Engineers AIChE J, 53: 16201632, 2007
Keywords: UNIFAC, vapor-liquid equilibrium (VLE), group-contribution methods, con-
nectivity index (CI), interaction parameters, prediction

Introduction ous process alternatives. Model based solution approaches

The production of chemical products usually involves where the property model is embedded within the process
processing steps, such as extraction, distillation, absorption model offer a quick evaluation of process alternatives if the
and crystallization where phase equilibria play a major role. property model is able to calculate the needed properties
In many chemical plants, phase equilibrium based separation efciently and reliably. Therefore, it is practical to use truly
processes contribute a large percentage of the total cost, predictive models of large application range, and acceptable
especially if the separation steps involve distillation opera- levels of accuracy and uncertainties.
tions. Incentives to improve process efciency in terms of Group-contribution-based property estimation methods are
energy consumption may require the examination of numer- very suitable for preliminary screening of process alterna-
tives, because they are able to provide quick and reliable
estimates at low-computational costs. In reverse property pre-
Correspondence concerning this article should be addressed to R. Gani at diction problems, such as the design of molecules/solvents,
[email protected].
group-contribution-based methods are ideally suited because
2007 American Institute of Chemical Engineers they offer a means of generating molecular structures with

1620 June 2007 Vol. 53, No. 6 AIChE Journal

desired properties, as well as a predictive method for estima- ogy, therefore, employs a parallel method for calculating the
tion of properties (Gani et al.1). liquid-phase-activity coefcients, based on connectivity indi-
Among the GC based methods for prediction of liquid- ces and their interactions, which can be used to predict the
phase-activity coefcients for mixtures, the UNIFAC missing group interaction parameters, as well as predicting
method2 is the best known and most successful. It has been molecular interaction parameters. The article describes the
used in many areas, for example; developed connectivity index-based method (to be called the
(1) To calculate vapor-liquid equilibria (VLE)2 CI-UNIFAC method), together with a description of the pa-
(2) For calculating liquid-liquid equilibria,3 rameter regression step, and an analysis of the results in dif-
(3) For calculating solid-liquid equilibria,4 ferent application scenarios. Note, however, that the current
(4) For determining activities in polymer solutions,5 version is limited only to the UNIFAC groups composed of
(5) For determining vapor pressures of pure components,6 carbon, hydrogen, oxygen and/or nitrogen atoms and the
(6) For determining ash points of solvent mixtures,7 molecular groups, water and methanol. Note also that the
(7) For determining solubilities of gases,8 objective here is not to improve the accuracy of the reference
(8) For estimation of excess enthalpies.9 UNIFAC model, but to extend its application range.
The UNIFAC method has several well-known limita-
tions.10 In this article, however, we are only concerned with
the problem of the missing groups or group-interaction pa- Theoretical Background
rameters needed for the calculation of liquid-phase-activity Connectivity indices
coefcients related to vapor-liquid equilibrium. The status
of groups and group-interaction parameters can be seen The connectivity indices nw are formalisms dened via
from the published UNIFAC group-interaction parameter graph theoretical concepts intended to describe topological
matrix10,11it can be observed that there are still many characteristics of molecular structures (Kier and Hall,12,13
blank spaces (indicating missing parameters) in the matrix. Trinajstic,14 Randic15). The graph theoretical treatment of the
If one wants to use the UNIFAC method to predict the molecule starts by the construction of the hydrogen-sup-
phase equilibrium of a mixture for which interaction parame- pressed graph of the molecular structure. For example, the
ters are not available, or for which the compounds in the differences between the molecular structure and its corre-
mixture cannot be described using the functional groups of sponding hydrogen-suppressed graph for acetic acid are illus-
the selected UNIFAC method, the usual procedure would be trated in Figure 1. On the lefthand side of Figure 1, the rep-
to collect the necessary experimental data, and then to resentation of acetic acid with two groups is illustrated, while
regress the missing (new) interaction parameters by matching the corresponding molecule and group representation in
the collected data. This, however, can be time consuming, terms of hydrogen suppressed graph is shown on the right-
expensive, and sometimes even infeasible as it may not be hand side of Figure 1.
possible to perform the necessary experiments. Also, since As shown on the righthand side of Figure 1, the nonhydro-
most of the available data may have already been used to gen atoms become vertices 1, 2, 3, 4 in the graph, while the
estimate the UNIFAC group-interaction parameters, new pa- bonds become edges a, b and c. The omission of hydrogen
rameters cannot be regressed without additional experimental and double bonds in the graph is compensated by the manner
data. Gmehling and coworkers have been doing this by add- in which the atomic (valence) index dn for each vertex is
ing new experimental data.11 dened.
Another way to address this problem is to create the miss- The atomic index dn for each atom/vertex is dened as
ing groups (when needed), and to predict the missing group- follows
interaction parameters using the available data. Connectivity

indices (Kier and Hall12), which use information on the Zn
NH
d

n
 (1)
atoms and their connections, can be used to create new Z
Zn
1
groups as shown by Gani et al.1 for pure component property
prediction. In this way, just as groups are used to represent where
molecules that are not included in the regression step, atom Zn is the number of valence electron in the atom, NH is the
connectivities are used to represent groups that are also not number of hydrogen attached to the atom, Z is the atomic
included in the regression step. number.
In this article, a methodology is proposed to predict the Table 1 lists the values of atomic indices for various
missing UNIFAC group-interaction parameters for the predic- atoms/vertices. The dened atomic index comprises of not
tion of vapor-liquid equilibrium (VLE), by using atom con- only information about the nature of the atom associated to
nectivity indices, and their interaction parameters regressed
by using only the currently available experimental data.
According to this methodology, when a group-interaction pa-
rameter is not available, using the (atom) connectivity indices
that represent the corresponding groups and the regressed CI-
interaction parameters, the missing group-interaction parame-
ter is calculated (that is predicted) for the reference UNIFAC
method (in this article, the UNIFAC method together with its
groups and parameter tables reported by Hansen et al.10 is
selected as the reference UNIFAC method). The methodol- Figure 1. Molecular structure to graph.

AIChE Journal June 2007 Vol. 53, No. 6 Published on behalf of the AIChE DOI 10.1002/aic 1621
 Table 1. Atomic Indices Values for Acetic acid valence connectivity indices have values of 2.355, 0.928 and
Atom 1 2 3 4
0.519, respectively.
A general formula14 for calculating m-th order valence
Zv 4 4 6 6
NH 3 0 1 0
connectivity index is given below
Z 6 6 8 8 X Y

1=2
dv 1 4 5 6 w
n m
dni t 1; T; i 1; M (7)
t i

the vertex, but also about the way it is bonded to its sur- where, the sum is over all type T subgraphs, with M edges
rounding atoms. using the valence dn values. Figure 2 shows the different
The zeroth-order (atomic) connectivity index nw0 is dened type T subgraphs for each order M (with pentane as an exam-
as a summation over the vertices of the hydrogen suppressed ple and Mmax 3). Calculations of valence connectivity
graph (see Figure 1). indices for groups are done exactly the same way as for mol-
8 9 ecules (see also Gani et al.1).
X> 1 >
>
> > Valence connectivity indices of higher-orders than two can
w :pn> i 1; L
n 0
; (2)
be dened to represent topological features of larger frag-
i di
ments. However, for the sake of simplicity and the objective
where L is the number of vertices (atoms) in the graph, and of this work (predict missing group interactions), these are
the din are the valence atomic indices whose values can be not employed in this article. The reason being that that the
obtained from Table 1 for the corresponding atom. missing groups to be created through the CI-UNIFAC
The rst-order valence bond indices bn can be dened for method will represent only a small part of the total molecular
each bond, or path of length one-edge on the graph (as structure (none of the UNIFAC groups10 contain more than
shown in Figure 1), by using the dn corresponding to the pair four different atoms, not counting the molecular groups),
of bonding atoms. That is and, therefore, higher-order terms are not necessary. There-
fore, the number of needed CI-interaction parameters are
bn dni :dnj (3) kept at a manageable level. Note that the UNIFAC groups10
are of the rst-order type, and, hence, should not need higher
dimensional structural parameters.
where i and j are the atoms connected by the bond.
Similarly, the rst-order (bond) valence connectivity index
n 1
w is dened as a summation over the edges of the hydro- Reference UNIFAC method
gen-suppressed graph (see Figure 1)
The UNIFAC method, initially introduced by Fredenslund
8 9 et al.,2 is here presented in a different form that is more
X> 1 >
w
n 1 >
>p >
> i 1; M appropriate for purposes of computation.16
: n;
(4)
i b i The UNIFAC method considers activity coefcient as
composed of two additive parts a combinatorial part to
where M is the number of paths of length one-edge in the account for molecular size and shape differences, and a re-
graph while the rst-order valence bond index bn is given by sidual part to account for molecular interactions. For each
Eq. 3. compound i in a mixture, we have
The second-order valence bond indices en can be dened
for each path of length two-edges on the graph (as shown in ln gi ln gCOM
i ln gRES
i (8)
Figure 1), by using the dn of the corresponding atoms.
That is

en dni :dnj :dnk (5)

where i, j and k are the atoms involved in the path.

Similarly, the second-order valence connectivity index nw2
is dened as a summation over the paths of length two-edges
of the hydrogen-suppressed graph (see Figure 1)
8 9
X> 1 >
>
> >
w :pn > i 1; N
n 2
; (6)
i ei

where N is the number of paths of length two-edges in the

graph while the second-order valence bond index en is given
by Eq. 5.
For acetic acid, whose hydrogen-suppressed graph is
shown in Figure 1, and whose valence atomic index values
can be found in Table 1, the zero-, rst- and second-order Figure 2. Types of subgraphs.

1622 DOI 10.1002/aic Published on behalf of the AIChE June 2007 Vol. 53, No. 6 AIChE Journal
 The combinatorial and residual terms, respectively, are regressed by matching phase equilibrium data (the published
expressed as set of Hansen et al.10 is used in this article).
8 9 To use the reference UNIFAC method for predicting VLE
> Ji Ji > of a mixture, one needs to give the molecular concentrations
ln gCOM
i 1
Ji ln J i
5qi :1
ln ; (9)
Li Li xi for each compound, the group surface and volume parame-
ters Qk and Rk, and the group-interaction parameters amk for
X8 > ski
9
ski > all the groups present in the mixture. If a new group is nec-
ln gi qi 1
ln Li

RES >
:yk
Gki ln > ; (10)
Zk Zk essary to represent part of a molecular structure, it will need
k
to be created (Gani et al.1), and the corresponding group pa-
The quantities Ji and Li are given by the following relations rameters estimated. If a group already exists, but not the
group interaction parameter, then only the necessary group
ri interaction parameters will need to be estimated.
Ji P (11)
j rj xj

qi
Development of the CI-UNIFAC method
Li P (12) In order to be able to estimate the missing UNIFAC
j qj rj
group-interaction parameters for UNIFAC groups containing
C, H, O and/or N atoms, the following procedure is applied:
In addition, the following denitions apply
(1) A relation is derived between the group-interaction pa-
X i rameters amn and:
ri nk Rk (13)  The number of C-atoms in each group,
k
 The number of O-atoms in each group,
X i  The number of N-atoms in each group,
qi nk Qk (14)  The connectivity indices for each group,
k
 CI-interaction parameters.
i
(2) Existing VLE data is used to regress the CI-interaction
G k nk Q k (15) parameters, while the atom numbers and the connectivity
X indices are obtained directly from the group denition.
yk Gki xi (16) (3) Using the regressed CI-interaction parameters and the
i derived relation, the missing group-interaction parameters
X amn are predicted.
sni Gmi tmn (17)
m Relation between group- and CI-interaction parameters
X The CI interactions are considered at 4 different levels:
Zn sni xi (18)
i
 Level 1: interaction between 0th and 0th order connectiv-
ity indices
and  Level 2: interaction between 0th and 1st order connectiv-
ity indices

amn  Level 3: interaction between 1st and 1st order connectiv-
tmn exp (19) ity indices
T
 Level 4: interaction between 0th and 2nd order connec-
Subscript i identies species, and j is a dummy index run- tivity indices
ning over all species. Subscript k identies subgroups, sub- Note that the interaction between 1st and 2nd, 2nd and
script n identies main groups, and m is a dummy index 2nd orders are neglected because of the relatively small size
running over all main groups. The quantity nk(i) is the number and structure of the UNIFAC groups (compared to the mole-
of subgroups of type k in a molecule of species i. Values of cule they represent). For two groups represented by m and n,
the parameters Rk and Qk are usually calculated by the rules where m is less than n, the following two relations apply
of Bondi,17 while the group-interaction parameters amn are (counting all interactions at all levels):

a mn bC
C ACC mn 0 bC
O A mn 0 bC
N A mn 0 bO
C A mn 0 bO
O A mn 0 bO
N A mn 0 bN
C A mn 0 bN
O A mn 0 bN
N A mn 0
CO CN OC OO ON NC NO NN
|{z}
0th
order interactions
cC
C ACC mn 1 cC
O A mn 1 cC
N A mn 1 cO
C A mn 1 cO
O A mn 1 cO
N A mn 1 cN
C A mn 1 cN
O A mn 1 cN
N A mn 1
CO CN OC OO ON NC NO NN
|{z}
1st
order interactions
dC
C ACC mn 2 d C
O A mn 2 d C
N A mn 2 d O
C A mn 2 d O
O A mn 2 d O
N A mn 2 d N
C A mn 2 d N
O A mn 2 d N
N A mn 2
CO CN OC OO ON NC NO NN
|{z}
2nd
order interactions
eC
C ACC mn 3 eC
O A mn 3 eC
N A mn 3 eO
C A mn 3 eO
O A mn 3 eO
N A mn 3 eN
C A mn 3 eN
O A mn 3 eN
N A mn 3
CO CN OC OO ON NC NO NN
20
|{z}
3rd
order interactions

AIChE Journal June 2007 Vol. 53, No. 6 Published on behalf of the AIChE DOI 10.1002/aic 1623
 a nm bC
C ACC nm 0 bC
O A nm 0 bC
N A nm 0 bO
C A nm 0 bO
O A nm 0 bO
N A nm 0 bN
C A nm 0 bN
O A nm 0 bN
N A nm 0
CO CN OC OO ON NC NO NN
|{z}
0th
order interactions
cC
C ACC nm 1 cC
O A nm 1 cC
N A nm 1 cO
C A nm 1 cO
O A nm 1 cO
N A nm 1 cN
C A nm 1 cN
O A mn 1 cN
N A nm 1
CO CN OC OO ON NC NO NN
|
{z}
1st
order interactions
d C
C ACC nm 2 d C
O A nm 2 d C
N A nm 2 d O
C A nm 2 d O
O A nm 2 d O
N A nm 2 d N
C A nm 2 d N
O A nm 2 d N
N A nm 2
CO CN OC OO ON NC NO NN
|
{z}
2nd
order interactions
eC
C ACC nm 3 eC
O A nm 3 eC
N A nm 3 eO
C A nm 3 eO
O A nm 3 eO
N A nm 3 eN
C A nm 3 eN
O A nm 3 eN
N A nm 3
CO CN OC OO ON NC NO NN
21
|
{z}
3rd
order interactions

In Eqs. 2021, bX
Y, cX
Y, dX
Y, eX
Y, bX
Y , cX
Y , dX
Y , 1st, and 2nd-order connectivity indices. As the group ACH is
eX
Y are the interaction parameters (whose values are used to represent aromatic rings, its 0th and 1st-order connec-
obtained through regression of VLE data) between atoms of tivity indices are calculated assuming that each aromatic car-
type X, and atoms of type Y (where X and Y represent C, O bon is attached to two similar groups, and that these three
and N atoms on a H-atom free basis); (AXY mn)i are the coef- atoms/vertices are included in the same group. This is not
cients to be used with the corresponding CI-interactions, that the case for the second-order connectivity index. Note also
is, at the order i, between atoms of type X, and atoms of that for this rst version of the CI-UNIFAC method, the
type Y, in the interaction between group m and group n, and groups COO and CON, which are available in Hansen
which are dened by the following equations for each order et al.10 have not been considered, but will be added in the
next version of the parameter tables.
  m n 0 n n To cover the 18 UNIFAC (main) groups containing only
nX wn
nY w0m
AXY
mn nw0 nw0
(22) C, H, O and N atoms, which we call functional groups,
0
n m that is, excluding the molecular groups, water and metha-
nol, represented in Table 2, 32 40 CI-interaction parame-
  m n n n ters are needed, instead of 18*17 306 group-interaction pa-
nX w1n
nY w0m
AXY
mn nw1 nw0
(23) rameters in the reference UNIFAC method (including water
1
n m and methanol interaction parameters, this number is 382).
Note that 32 parameters are needed for groups with C, H and
  mn nn
nX w1n
nY w1m O atoms (or for groups with a maximum of three atoms),
AXY
mn nw1 nw1
(24) while an additional 40 parameters are needed when groups
2
n m with N atoms are added (or for groups with a maximum of
four atoms). This means that for future additions of atoms, if
  m n n n
nX w2n
nY w0m they do not have more than four atoms (UNIFAC groups do
AXY
mn nw2 nw0
(25) not have more than four atoms, except for molecular groups),
3
n m
only 40 additional parameters will be needed for each atom.
However, for the addition of S atom groups, the UNIFAC
where n(m)
X is the number of atoms of type X in the group m; group table10 has only the (main) groups containing C, H
n i
w (m) is the i-th order valence connectivity index for the
group m.
Note that Eqs. 2025 can, in principle, be written for any
Table 2. Groups, Stoichiometry and Connectivities
sets of groups containing three atoms (not counting the H
atom). A check of the UNIFAC groups,10 show that none of Index Group nC nO nN w0 w1 w2
the 50 UNIFAC (main) groups contain more than three 1 CH2 1 0 0 0.70711 / /
atoms (not counting the H atom and molecular groups). 2 CC 2 0 0 1.000 0.250 /
Therefore, addition of S and halide atoms will not increase 3 ACH 1 0 0 0.57735 0.66667 0.19245
4 ACCH2 2 0 0 1.20711 0.35355 /
the size of Eqs. 2021, as we would still consider only a 5 OH 0 1 0 0.44721 / /
maximum of three atoms (besides the H atom). Based on the 6 CH3OH 1 1 0 1.44721 0.44721 /
available UNIFAC groups,10 for S atoms, only the interac- 7 H2 O 0 1 0 0.500 / /
tions between C and S need to be considered, while for Cl 8 ACOH 1 1 0 0.94721 0.22361 /
9 CH2CO 2 1 0 1.61535 0.55768 0.14434
and F atoms, only the interactions between C, Cl and F need
10 COH 1 1 0 0.98560 0.23570 /
to be considered. Therefore, the number of CI-interaction pa- 11 CCOO 2 2 0 1.81650 0.65825 0.28746
rameters remains at a manageable level. 12 HCOO 1 2 0 1.39385 0.47140 0.09622
Note that the actual number of CI-interaction parameters 13 CH2O 1 1 0 1.11535 0.28867 /
needed for any missing UNIFAC group interaction parame- 14 CNH2 1 0 1 1.07735 0.28867 /
15 CNH 1 0 1 1.000 0.250 /
ters are quite small since not all the terms in Eqs. 2021 are 16 (C)3N 3 0 1 1.94721 0.72361 0.33541
needed (as highlighted through Eqs. 3031 in case study 1). 17 ACNH2 1 0 1 1.07735 0.28867 /
Table 2 lists the UNIFAC groups10 containing only C, H, 19 CCN 2 0 1 1.44721 0.47361 0.11180
O and N atoms and molecular groups water and methanol, 20 COOH 1 2 0 1.35546 0.42773 0.09129
26 CNO2 1 2 1 1.72474 0.53746 0.23471
along with their index numbers, atom stoichiometry and 0th,

1624 DOI 10.1002/aic Published on behalf of the AIChE June 2007 Vol. 53, No. 6 AIChE Journal
 Table 3. Functional Groups CI-Interaction Parameters tions of this work are exclusively, based only on isothermal
CI-interaction CI-interaction
data. Isothermal data has been used, because in this case, it
Parameter (i-j) Value Parameter (j-i) Value is not necessary to specify the vapor pressure or its depend-
bC
C 977.87 bC
C
145.02
ence upon temperature.18 The plan has been to use other data
cC
C
108.11 cC
C
281.54 for purposes of validation. The VLE data sets were extracted
dC
C 104.61 dC
C 321.40 from the CAPEC Database19 and Processiums online data-
eC
C
109.42 eC
C 261.28 base e-thermo1,20 and were checked for thermodynamic con-
bC
O
1134.83 bC
O 401.21 sistency using the test of Van Ness.21
cC
O
17.44 cC
O 358.76
dC
O 45.25 dC
O
69.53 Due to the differences in the units and the magnitude of
bC
N 20.04 bC
N
52.78 pressure values in the isothermal data sets, the following
cC
N 9.94 cC
N
117.17 objective function was chosen, which is expressed as the av-
dC
N 47.79 dC
N
212.21 erage relative pressure quadratic deviation
eC
N
6.54 eC
N 20.01
N 8 92
1X
eC
O 61.83 eC
O 66.83

11028.85 >
> Pexp
Pcalc >
>
bO
C bO
C 63.32 S : ; (26)
cO
C
3913.23 cO
C
34.00 N i1 Pexp
dO
C
633.11 dO
C 195.52
eO
C 2128.63 eO
C
30.87
bO
O 11347.17 bO
O
515.78
where N is the number of experimental data points used in
cO
O 4250.92 cO
O
472.59 the estimation.
dO
O 329.24 dO
O 1026.39 For systems not needing association terms, the equilibrium
eO
O
2150.65 eO
O
398.08 pressure was calculated as follows
bO
N 0.00 bO
N
1.121
cO
N 10.82 cO
N 2.27
X
dO
N 47.79 dO
N
3.85
Pcalc i POYi
xi gi Psat (27)
eO
N
2.80 eO
N 18.62 i
bN
C 45.41 bN
C 194.75
cN
C
32.91 cN
C 450.73 where i is an index running over all species in the mixture,
dN
C 333.83 dN
C
251.60 and POYi is the Poynting correction.
eN
C 108.41 eN
C
283.58 For systems needing an association term, chemical equilib-
bN
O 1.83 bN
O
395.38 rium were solved and bubble-dew point calculations were
cN
O 11.81 cN
O
91.91
dN
O 6.27 dN
O 27.47 performed with association constants, using the method of
eN
O 30.51 eN
O
283.66 Hayden and OConnell.22
bN
N 43.47 bN
N
4.63 The optimization algorithm used for the data tting was
cN
N
38.85 cN
N
69.02 the Levenberg-Marquardt23,24 technique, which is a local
dN
N 327.89 dN
N
65.09
optimization method. It is, therefore, strongly dependent on
eN
N 83.23 eN
N 20.00
the initial guesses for the parameters. In order to ensure that
the best parameter values are obtained, the following scheme,
involving sequential and simultaneous parameter estimation,
and S atoms, meaning that only interactions between C-S using the results from the previous steps as initial guesses for
and S-S connectivities would be needed (that is 24 additional each addition of a group/data set to the system, was
parameters). If groups with C, H, Cl and F atoms are consid- employed. Also, to ensure that the parameters did not move
ered, 64 additional CI-interaction parameters would be too far away from the previous step, and, thus, guarantee that
needed. the previously investigated groups/data sets would still be
A set of 32 CI-interaction parameters is used to calculate represented with comparable accuracy, parameter regulariza-
methanol interactions (only for C, H and O atoms), and
another 32 CI-interaction parameters for water interactions
(only for C, H and O atoms). Therefore, a total of 72 32 Table 4. Methanol CI-Interaction Parameters
32 136 connectivity index parameters are needed to CI-interaction CI-interaction
estimate all of the UNIFAC group-interaction parameters for Parameter (i-j) Value Parameter (j-i) Value
groups containing only C, H, O and N atoms plus water and bC
C 417.10 bC
C 14.01
methanol interactions (with groups containing C, H and O cC
C
22.34 cC
C
236.84
atom groups). This means that less than 50% parameters are dC
C
54.19 dC
C
250.59
needed for the CI-UNIFAC method to cover the interaction eC
C
3.51 eC
C 0
bC
O 396.91 bC
O
94.79
parameters for the 20 groups listed in Table 2. cC
O
29.41 cC
O
236.84
dC
O
94.63 dC
O
250.59
eC
O
14.42 eC
O 0
Parameter Optimization bO
C
66.42 bO
C 86.40
Values for the set of CI-interaction parameters (see Eqs. cO
C 61.12 cO
C
97.14
dO
C
54.19 dO
C
35.20
2021) have been determined by regression using a database eO
C
3.51 eO
C 0
involving 265 binary data sets (4,408 data-points) represent- bO
O
74.84 bO
O 0.14
ing vapor-liquid equilibrium for the selected 18 functional cO
O 93.49 cO
O
97.14
groups, 36 data sets (636 data-points) for methanol, and 40 dO
O
94.63 dO
O
35.20
eO
O
14.42 eO
O 0
data sets (515 data-points) for water. The parameter estima-

AIChE Journal June 2007 Vol. 53, No. 6 Published on behalf of the AIChE DOI 10.1002/aic 1625
 Table 5. Water CI-Interaction Parameters where N is the number of data points in the whole set of
CI-interaction CI-interaction
data.
Parameter (i-j) Value Parameter (j-i) Value From the results in Table 6, it can be noted that the aver-
bC
C 0 bC
C 0
age absolute relative deviation, AARD(%), is slightly lower
cC
C 0 cC
C 0 for the CI-UNIFAC method when compared for the func-
dC
C 0 dC
C 0 tional groups(2.54% vs. 2.57%), methanol (2.56% vs.
eC
C 0 eC
C 0 3.98%), and water (3.34% vs. 3.71%), even though less
bC
O
845.17 bC
O
2205.53 parameters (136 instead of 308 in the reference UNIFAC
cC
O 0 cC
O 0
dC
O 0 dC
O 0 table10) are used to estimate VLE for the corresponding func-
eC
O 0 eC
O 0 tional groups, methanol and water.
bO
C
268.83 bO
C 123.53 Figures 36 present the comparison between the devia-
cO
C 706.28 cO
C
1071.60 tion from experimental data with the original UNIFAC
dO
C 0 dO
C 0
eO
C 16.30 eO
C 1.17
method, and the CI-UNIFAC method, for each of the data
bO
O 963.82 bO
O 909.49 sets used in the parameter estimation process, and for each
cO
O 71.54 cO
O 5821.92 of the CI-interaction parameters sets (functional groups,
dO
O 0 dO
O 0 methanol and water). The deviation plotted in these gures
eO
O 3.92 eO
O 0 is the AARD (%) for the corresponding data set, dened by
Eq. 29, but where N is the number of data points in the
tion was performed by adding a second term in the objective considered data set only. From these gures, one can see
function. that the range of the deviation compared to the experimen-
tal data using the CI-interaction parameters to calculate
N 8 92
1X >
> Pexp
Pcalc >
>
X
2 group-interaction parameters is within the range of the devi-
S : ; wreg AIPj
AIPIG
j (28) ation of the reference UNIFAC method. The deviations are
N i1 Pexp j
even better in most cases, especially for the systems that
where N is the number of experimental data points, AIPj is are poorly represented with reference UNIFAC method. It
the current value of the CI-interaction parameter j, AIPIG indicates that the regressed CI-interaction parameters from
j its
corresponding initial guess, and wreg a weighting value used Tables 35 can be used to calculate the missing group-inter-
to increase or decrease the inuence of regularization in the action parameters, and will make VLE predictions with sim-
optimization. ilar accuracy as the original UNIFAC method. Note that
since the CI-UNIFAC method is proposed only as an exten-
sion of the reference UNIFAC method when the needed
Model Analysis and Prediction group interactions are missing, it is not reasonable to expect
Results and model analysis this method to correct the errors of the reference model.
The regressed CI-interaction parameters needed to predict The KT-UNIFAC25 has been developed to correct for some
a group-interaction parameter, are given for the functional of the inaccuracies of the original UNIFAC method. How-
ever, based on the parameters (Tables 35) and the corre-
groups in Table 3, for methanol in Table 4, and for water in
Table 5. sponding correlation results (Figures 36), it is reasonable
Table 6 presents the average absolute relative deviation to consider that the CI-UNIFAC method can be used for
estimating missing group-interactions with acceptable accu-
(AARD%; calculated using Eq. 29 for the reference UNIFAC
method, and the CI-UNIFAC method over the entire data-set racy. This certainly would be true if only one pair of miss-
used for the estimation of the CI-interaction parameters pre- ing UNIFAC group interactions need to be generated
through the CI-UNIFAC method.
sented in Tables 35, for the groups with C, H, O atoms and
molecular groups, water and methanol. For correlation results The interested reader can obtain a copy of the CI-interac-
involving also groups with N atoms, see Figure 4. tion parameter tables, as well as the lled reference UNIFAC
parameter table from the corresponding author. Also, more
N  
1X  Pexp
Pcalc  details on the numerical values related to Figures 36 can be
AARD%   100 (29) obtained from the corresponding author.
N i1  Pexp

Table 6. Average Absolute Relative Deviation (AARD) in % Over the Whole Set of Data
C, H & O Atoms C, H, O, & N Atoms
Reference UNIFAC10 CI-UNIFAC Method Reference UNIFAC10 CI-UNIFAC Method
Functional groups 2.57 2.54 5.18 3.66
Figure 3 Figure 4
Methanol 3.98 2.56
Figure 5
Water 3.71 3.34
Figure 6

Comparison of the values for functional groups, methanol and water using the reference UNIFAC method,10 and the CI-UNIFAC method.

1626 DOI 10.1002/aic Published on behalf of the AIChE June 2007 Vol. 53, No. 6 AIChE Journal
 group-interaction parameter akl, and the rst one (bC
C,
bC
O,. . .) for the group-interaction parameter alk.
The application of the CI-UNIFAC method is illustrated
below through three case studies:
(1) Involves a system where an available group-interaction
parameter is considered as missing, and calculated using the
CI-interaction parameters,
(2) Involves a system where there is a missing group-inter-
action parameter, which is calculated using the CI-interaction
parameters,
(3) Involves a system that is not well represented by the
original UNIFAC groups, in which a new group is created,
Figure 3. Functional groups: AARD (%) for each indi- and its group-interaction parameters are calculated using the
vidual data set. CI-interaction parameters.
Comparison between the reference UNIFAC method,10 and For these examples, and whenever possible (cases (1) and
the CI-UNIFAC (new) method (results sorted by ascending (3)), a comparison of the deviation from experimental data is
deviation using the new method for groups with C, H & O
atoms). [Color gure can be viewed in the online issue, made between the original UNIFAC method, and the CI-
which is available at www.interscience.wiley.com.] UNIFAC method. Extensive tests have been made for the
CI-UNIFAC method. However, when data is missing, it is
difcult to provide any comparisons. Also, the three case
VLE Prediction using the new method studies follow the general trend of all the tests made for the
The use of the CI-interaction parameters to calculate a CI-UNIFAC method.
missing group-interaction parameter requires the following: Case study 1: Known Interaction Considered as Missing
(1) Use the water CI-interaction parameters (Table 5) to and Re-Estimated. In this case study, the CI-UNIFAC
calculate any group-interaction parameter involving water, method is be used to re-estimate the interaction parameter
(2) Use the methanol CI-interaction parameters (Table 4) between the CH2 and the OH main-groups. Since these two
to calculate any group-interaction parameter involving metha- groups have limited connectivity (only one nonhydrogen
nol, excluding methanol-water, atom in each group), only the zeroth-order connectivity indi-
(3) Use the functional groups CI-interaction parameters ces are used to calculate the group interactions. Also, since
(Table 3) to calculate any group-interaction, excluding the the CH2 group has an index value of 1, and the OH group
water and methanol interactions. has an index value of 5, the bC
O CI-interaction parameter
To estimate the group-interaction parameter between two will be used to calculate the aCH2
OH interaction, while the
groups k and l: bO
C atom interaction parameter will be used to calculate
 If k has a lower group index than l, it will use the rst the aOH
CH2 interaction. All the CI-interaction parameters
set of CI-interaction parameters (bC
C, bC
O,. . .) for the used in the calculations are taken from Table 3. Details of
group-interaction parameter akl, and the second one (bC
C , the calculations are given below
bC
O ,. . .) for the group-interaction parameter akl,
CH2 OH OH CH
 If k has a higher group index than l, it will use the sec- nC w0
nO w0 2
ond set of CI-interaction parameters (bC
C , bC
O ,. . .) for the aCH2
OH bC
O OH CH
(30)
w0 w0 2

Figure 4. Functional groups: AARD (%) for each individual data set.
Comparison between the reference UNIFAC method,10 and the CI-UNIFAC (new) method (Results sorted by ascending deviation using
the new method for groups with C, H, O and N atoms for the 58 additional datasets). [Color gure can be viewed in the online issue,
which is available at www.interscience.wiley.com.]

AIChE Journal June 2007 Vol. 53, No. 6 Published on behalf of the AIChE DOI 10.1002/aic 1627
 Table 7. Group Interaction Matrix for CH2 and OH
with the Reference UNIFAC Method10 and the
CI-UNIFAC Method
Reference
UNIFAC10 CI-UNIFAC Method
Groups CH2 OH CH2 OH
CH2 0 986.5 0 932.16
OH 156.4 0 52.04 0

Figure 5. Methanol: AARD (%) for each individual data As one can see from Figure 7, the results from the original
set. UNIFAC method and the CI-UNIFAC method are very com-
Comparison between the reference UNIFAC method10 and parable for predicting the vapor-liquid equilibrium of the 2-
the CI-UNIFAC (new) method (Results sorted by ascending butanol/n-hexane system. The new method is even slightly
deviation using the new method for groups with C, H and better near the azeotrope composition. Note, however, that a
O atoms). [Color gure can be viewed in the online issue,
which is available at www.interscience.wiley.com.] different set of UNIFAC group parameters11 may give an
improved UNIFAC correlation.
Case study 2: Estimation of a Missing Group-Interaction
OH CH2 CH OH Parameter. In this case study, the CI-UNIFAC method is
nO w0
nC 2 w0
aOH
CH2 bO
C CH OH
(31) used to estimate the interaction parameter between the
w0 2 w0 HCOO (formate), and the H2O (water) groups, which is a
missing interaction in the reference UNIFAC method10 used
Applying Eqs. 3031, and the connectivity index values for in this article. Since the H2O group has limited connectivity
the groups from Table 2, we get, for aCH2
OH interaction (only one nonhydrogen atom), only its zeroth-order connec-
tivity index is used to calculate the group interactions, while
1  0:44721
1  0:70711
aCH2
OH
1134:18  the HCOO group has extended connectivity (three nonhydro-
0:44721  0:70711 gen atoms), and its zeroth-, rst- and second-order connectiv-
aCH2
OH 932:16 ity indices will be used. Also, since the H2O group has an
index value of seven, and the HCOO group has an index
and for aOH
CH2 interaction value of 12, the bO
C, cO
C, eO
C, bO
O, cO
O and eO
O
CI-interaction parameters will be used to calculate the
1  0:70711
1  0:44721 aH2O
HCOO interaction, while the bC
O and bO
O atom inter-
aOH
CH2 63:32 
0:70711  0:44721 action parameters will be used to calculate the aHCOO
H2O
aOH
CH2 52:04 interaction. Note that all the CI-interaction parameters used
in the calculations need to be extracted from Table 5, since
Table 7 presents the different values of the group-interaction water is involved in the interactions. In this case, Eqs. 2021
parameters between the CH2, and the OH groups with the take the form (for purposes of illustration, only the rst
reference UNIFAC method, and the CI-UNIFAC method. expression for the calculations and the nal numerical results
The results of the two methods for predicting the vapor-
liquid equilibrium are compared with an experimental data
set26 (2-butanol/n-hexane at 333.15 K), which was not used
during the parameter estimation process. Results of this com-
parison are illustrated in Figure 7.

Figure 7. Comparison of the original UNIFAC method

Figure 6. Water: AARD (%) for each individual data set. (dashed line), and the CI-UNIFAC method
Comparison between the reference UNIFAC method10 and (full line) toward a 2-butanol/hexane experi-
the CI-UNIFAC (new) method (Results sorted by ascending mental data set26 (n) at 333.15 K.
deviation using the new method for groups with C, H and
O atoms). [Color gure can be viewed in the online issue, [Color gure can be viewed in the online issue, which is
which is available at www.interscience.wiley.com.] available at www.interscience.wiley.com.]

1628 DOI 10.1002/aic Published on behalf of the AIChE June 2007 Vol. 53, No. 6 AIChE Journal
 Table 8. Group Interaction Matrix for CH2, H2O and tions are not attempting to correct. That is, using additional
HCOO with the Reference UNIFAC10 Method and the data, the water-CH2, water-HCOO and CH2-HCOO interac-
CI-UNIFAC Method tions could be ne tuned or re-estimated together with all the
Reference UNIFAC10 CI-UNIFAC Method other group interactions. The missing UNIFAC group param-
eters are available in Wittig et al.11 which also gives an
Groups CH2 H2O HCOO CH2 H2 O HCOO
improved prediction of the VLE.
CH2 0 1318 507 0 1318 507 Case Study 3: Creation of Group and Estimation of its
H2O 300 0 0 300 0
256.75
HCOO 329.3 0 0 329.3 2314.75 0
Group-Interaction Parameters. In case study 3, the new
method is used to create a group and its group-interaction pa-
rameters. In the original UNIFAC method, 1,2-ethanediol is
described as the combination of two CH2 groups and two
are given) OH groups. In this example, 1,2-ethanediol (ethylene glycol)
will be considered as a group by itself (noted ETHA). We
H O HCOO HCOO H2 O will use the new method for estimating the vapor-liquid equi-
nO 2 w0
nC w0
aH2 O
HCOO bO
C HCOO H2 O librium between 1,2-ethanediol (now considered as a group
w0 w0 by itself) and methanol.
H2 O HCOO HCOO H2 O
n w
nC w0 For calculating the corresponding group-interaction param-
CO
C O 1 HCOO H eters, one needs rst to calculate its connectivity indices.
2 O
w1 w0 According to the method described before, its zeroth-, rst-
H2 O HCOO HCOO H2 O
n w2
nC w0 and second-order connectivity indices are respectively 2.31,
eO
C O HCOO H2 O 1.13 and 0.45. Also, according to the rules of Bondi,17 the
w2 w0
corresponding surface Qk and volume Rk parameters are
H2 O HCOO HCOO H2 O
n w0
nO w0 equal to 3.02 and 2.92, respectively.
bO
O O HCOO H2 O
w0 w0 Since methanol only has zeroth- and rst-order connectiv-
H2 O HCOO HCOO H2 O
ity indices, the bC
C, cC
C, dC
C, eC
C, bC
O, cC
O, dC
O,
n w1
nO w0 eC
O, bO
C, cO
C, dO
C, eO
C, bO
O, cO
O, dO
O and eO
O
CO
O O HCOO H2 O
w1 w0 CI-interaction parameters will be used to calculate the
H2 O HCOO HCOO H2 O
aCH3OH
ETHA interaction, while the bC
C , cC
C , dC
C , bC
O ,
n w2
nO w0 cC
O , dC
O , bO
C , cO
C , dO
C , bO
O , cO
O and dO
O
eO
O O HCOO H2 O
w2 w0 CI-interaction parameters will be used to calculate the
aETHA
CH3OH interaction. Note that all the CI-interaction pa-
aH2 O
HCOO
256:75
rameters used in the calculations need to be extracted from
Table 4, since methanol is involved in the interactions. Using
HCOO H2 O H O HCOO
nC w0
nO 2 w0 Eqs 21 to 25 like in other case studies, the aCH3OH
ETHA and
aHCOO
H2 O bC
O H O HCOO aETHA
CH3OH are calculated and found to be
232.8 and
w0 2 w0
1184.6, respectively.
HCOO H2 O H O HCOO
n w0
nO 2 w0 The values of the group-interaction parameters (needed for
bO
O O H O HCOO
w0 2 w0 the binary system) for the original UNIFAC method are com-
aHCOO
H2 O 2314:75

Table 8 lists the group-interaction parameter table involv-

ing the CH2, H2O and HCOO main-groups with the original
UNIFAC method (where the interactions between H2O and
HCOO are missing), and the CI-UNIFAC method (where the
interactions between H2O and HCOO are calculated).
The results of the calculations and the use of the CI-
UNIFAC method for predicting the vapor-liquid equilibrium
is assessed through a set of experimental data27 (ethyl
formate/water at 300 K), which was not used during the CI-
interaction parameter estimation process. Results of this com-
parison are illustrated in Figure 8.
As one can see from Figure 8, the CI-UNIFAC method,
by allowing the user to calculate the interaction between
H2O and HCOO, permits to estimate the vapor-liquid equilib- Figure 8. Assessment of the results given by the CI-
rium of the ethyl formate/water with a reasonable quality. UNIFAC method (full line) for calculating a
Note that the azeotropic point is predicted quite well, even missing UNIFAC group-interaction parameter
though the deviations in pressure from the experimental val- towards a ethyl fromate/water experimental
ues are quite high for ethyl formate mole fractions less than data set27 (n) at 300 K.
0.1. This, however, may be an error from the original
[Color gure can be viewed in the online issue, which is
UNIFAC group-interaction parameters, which the CI-interac- available at www.interscience.wiley.com.]

AIChE Journal June 2007 Vol. 53, No. 6 Published on behalf of the AIChE DOI 10.1002/aic 1629
 Table 9. Group Interaction Matrix for CH2, OH and
CH3OH using the Reference UNIFAC10 Method, and
CH3OH and the new ETHA Group using the
CI-UNIFAC Method
CI-UNIFAC
Reference UNIFAC10 Method
CH2 OH CH3OH CH3OH ETHA
CH2 0 986.5 697.2 CH3OH 0
232.8
OH 156.4 0
137.1 ETHA 1184.6 0
CH3OH 16.51 249.1 0

pared with those predicted by the CI-UNIFAC method in

Table 9.
The performance of the CI-UNIFAC method for the pre-
diction of VLE involving the system 1,2-ethanediol/methanol
at 760 mm Hg, has been compared with a set of experimen-
tal data,28 which was not used during the parameter esti-
mation process. Results of this comparison are illustrated in
Figure 9.
As one can see from Figure 9, the predicted VLE with the
original UNIFAC method (with 1,2-ethanediol described as a
combination of CH2 and OH groups), and the new method Figure 10a. VLE calculations with the CI-method vs.
(with 1,2-ethanediol being a group by itself) are very compa- experimental data29 for butyronitrrile/2-bu-
rable for predicting the VLE for this binary system. The CI- tanol at 298.15 K (D).
UNIFAC method is even slightly better for 1,2-ethanediol
mole fractions between 0 and 0.5.
In Figure 10a, b, the performance of the CI-UNIFAC frequently encountered situation, where some UNIFAC
method is highlighted where all the group interactions have group-interaction parameters are not available. CI-interaction
been estimated through the CI-interaction parameters. Figure parameters that can be used to calculate missing pairs of
10a highlights the VLE for the binary mixture of Butyroni- group-interaction parameters are given for the case of UNI-
trile/2-butanol at 298.15 K,29 while gure 10b highlights FAC functional (main) groups with C, H, O and N atoms,
the VLE for the binary mixture of aniline/ethyl-acetate at
348.15 K.30

Conclusions
The CI-UNIFAC method for estimating missing UNIFAC
group interaction parameters provides a simple means for
calculation of vapor-liquid equilibrium compositions in the

Figure 9. Comparison of the original UNIFAC method

(dashed line), and the CI-UNIFAC method
(full line) toward a 1,2-ethanediol/methanol
experimental data set28 (n) at 760 mm Hg. Figure 10b. VLE calculations with the CI-method vs.
[Color gure can be viewed in the online issue, which is experimental data30 for aniline/ethyl-ace-
available at www.interscience.wiley.com.] tate at 298.15 K (D).

1630 DOI 10.1002/aic Published on behalf of the AIChE June 2007 Vol. 53, No. 6 AIChE Journal
and, molecular-groups methanol and water. The accuracy of yk dened in Eq. 16
prediction with the CI-UNIFAC method has been found to Zk dened in Eq. 18
be nearly the same as that of the reference UNIFAC method. k
n(i) number of groups k present in molecule i
tmk dened in Eq.19
The performance of the CI-UNIFAC method has been high- wm
n (k)
valence connectivity index of order m in group (or molecule) k
lighted through three representative case studies involving
the calculation of group-interaction parameters and prediction Indexes
of liquid phase activity coefcients, when known group-inter-
calc calculated quantity
actions are considered as missing, there is a missing pair of
COM combinatorial part
group-interactions for the original UNIFAC method, and exp experimental quantity
when a new group is created and its interaction predicted. In IG initial guess
all cases, the predicted vapor-liquid equilibrium has been RES residual part
found to be acceptable and compared well with experimental
data and predictions obtained using the original UNIFAC
method (when possible). While the present range of applic-
ability is limited to components containing only C, H, O and Literature Cited
N atoms, and to vapor-liquid equilibrium, this range can eas- 1. Gani R, Harper P, Hostrup M. Automatic creation of missing groups
ily be expanded to additional atoms and to different types of through connectivity index for pure component property prediction.
phase equilibrium. We believe the scope of the CI-UNIFAC Ind & Eng Chem Res. 2005;44:84368454.
2. Fredenslund A, Gmehling J, Rasmussen P. Vapor-Liquid Equilibria
method is very large as it allows the creation of a very large Using UNIFAC; Amsterdam: Elsevier; 1977.
number of groups and their parameters. To keep the CI-UNI- 3. Magnussen T, Rasmussen P, Fredenslund A. UNIFAC parameter ta-
FAC method manageable, it is recommended not to dene ble for prediction of liquid-liquid equilibria. Ind & Eng Chem Fund.
UNIFAC groups with more than four atoms, out of which 1981;20:331339.
4. Gmehling JG, Anderson TF, Prausnitz JM. Solid-Liquid Equilibria
two should be C and H atoms. Current and future work is Using UNIFAC. Ind & Eng Chem Fundamental. 1978;17:269273.
extending the CI-interactions for groups involving S, fol- 5. Oishi T, Prausnitz JM. Estimation of solvent activities in polymers
lowed by halides. At the same time, a parallel work is esti- using a group contribution method. Ind & Eng Chem Proc Des &
mating the CI-interactions for LLE and evaluating the use of Develop 1978;17:335339.
VLE/LLE parameters for SLE. 6. Jensen T, Fredenslund A, Rasmussen P. Pure Component Vapor
Pressures Using UNIFAC group Contribution. Ind & Eng Chem
Fund. 1981;20:239246.
7. Gmehling J, Rasmussen P. Flash points of ammable liquids using
Notation UNIFAC. Ind & Eng Chem Fund. 1982;21:186188.
8. Sander B, Skjold-Jrgensen S, Rasmussen P. Gas solubility calcula-
akl UNIFAC group interaction parameter between group k tions. I. UNIFAC. Fluid Phase Equilibria. 1983;11:105126.
and group l 9. Dang D, Tasslos DP. Prediction of enthalpies of mixing with a
bX
Y, bX
Y 0th-order CI-interaction parameter between atom X and UNIFAC model. Ind & Eng Chem Proc Des & Develop. 1986;25:
atom Y 2231.
cX
Y, cX
Y 1st-order CI-interaction parameter between atom X and 10. Hansen HK, Rasmussen P, Fredenslund A, Schiller M, Gmehling J.
atom Y Vapor-liquid equilibria by UNIFAC group-contribution. 5. Revision
dX
Y, dX
Y 2nd-order CI-interaction parameter between atom X and and extension. Ind & Eng Chem Res. 1991;30:23522355.
atom Y 11. Wittig R, Lohmann J, Gmehling J. Vapor-liquid equilibria by UNI-
eX
Y, eX
Y 2rd-order CI-interaction parameter between atom X and FAC Group Contribution. 6. Revision and Extension. Ind & Eng
atom Y Chem Res. 2003;42:183188.
Gki dened in Eq. 15 12. Kier LB, Hall HL. Molecular Connectivity in Structure Activity
Ji dened in Eq. 11 Analysis. New York: John Wiley & Sons, Inc.; 1986.
Li dened in Eq. 12 13. Kier LB, Hall HL. Issues in representation of molecular structure
NH number of hydrogen attached to the atom the development of molecular connectivity. J of Mol Graphics &
X number of atoms of type X in group k
n(k) Modelling. 2001;20:418.
P pressure 14. Trinajstic N. Chemical Graph Theory. Ohio: CRC Press; 1983.
P vapor pressure
sat
15. Randic M. Connectivity Index 25 Years After. J of Molecular
Qk surface parameter, for group k Graphics & Modelling. 2001;20:1935.
qi molecular surface parameter, for component i 16. Smith JM, Van Ness HC. Introduction to Chemical Engineering
ri molecular volume parameter, for component i Thermodynamics, 4th ed. New York: Mc Graw Hill; 1996.
Rk volume parameter, for group k 17. Bondi A. Physical Properties of Molecular Crystals, Liquids and
S value of the objective function Glasses; New York: John Wiley & Sons; 1968.
ski dened in Eq. 17 18. Van Ness HC, Byer SM, Gibbs RE, Part I. An appraisal of data
T temperature reduction methods. AIChE J. 1973;19:238244.
xi mole fraction of component i in the liquid phase 19. Nielsen TL, Abildskov J, Harper PM, Papaeconomou I, Gani R. The
wreg weighting factor for regularization CAPEC Database. J of Chem Eng Data. 2001;46:10411044.
Zv number of valence electron in the atom 20. e-thermo1 Online Database, Processium, Villeurbanne, France.
Z atomic number www.processium.com
21. Van Ness HC. Thermodynamics in the treatment of (vapor liquid)
Greek letters equilibria. J Chem Thermodynamics. 1995;27:113134.
22. Hayden JG, OConnell JP. A generalized method for predicting sec-
dn 0th-order (atomic) valence index ond virial coefcients. Ind & Eng Chem Proc Des & Develop.
bn 1st-order (bond; path of length one edge) valence index 1975;14:209216.
en 2nd-order (path of length two edges) valence index 23. Levenberg KA. Method for the solution of certain problems in least
gi activity coefcient for component i squares. Quarterly App Math. 1944;2:164168.
gCOM
i activity coefcient, combinatorial part 24. Marquardt D. An algorithm for least-squares estimation of nonlinear
giRES activity coefcient, residual part parameters. SIAM J of App Math. 1963;11:431441.

AIChE Journal June 2007 Vol. 53, No. 6 Published on behalf of the AIChE DOI 10.1002/aic 1631
25. Kang JW, Gani R, Abildskov J, Cobas J. Estimation of mixture 29. Garriga R, Sanchez F, Perez P, Gracia M. Excess gibbs energies at
properties from rst and second-order group contributions with eight temperatures and excess enthalpies and volumes at 298.15 K
UNIFAC models. Ind & Eng Res. 2002;41:32603273. for butanenitrile plus 2-Butanol. J Chem Eng Data. 1997;42:7883.
26. Maczinska Z. International data series. Selected Data Mixtures. 30. Khurma JR, Muthu O, Munjai S, Smith BD. Totalpressure vapor-
1983;3:199201. liquid equilibrium data for binary systems of aniline with 1-chloro-
27. Linek J, Wichterle I, Marsh KN. Vapor-liquid equilibria for water butane and ethyl-acetate. J Chem Eng Data. 1983;28:108112.
diacetone alcohol, ethyl methanoate water, and ethyl methanoate
phenol. J Chem Eng Data. 1996;41:12191222.
28. Baker TH, Fisher GT, Roth JA. Vapor liquid equilibrium and refrac-
tive indices of the methanol-ethylene glycol system. J Chem Eng Manuscript received Jun. 21, 2006, revision received Dec. 22, 2006, and nal
Data. 1964;1:1112. revision received Mar. 2, 2007.

1632 DOI 10.1002/aic Published on behalf of the AIChE June 2007 Vol. 53, No. 6 AIChE Journal