https://doi.org/10.36375/prepare_u.iiche.

a366

Generalized NRTL model for predicting vapor-liquid equilibrium data

from activity coefficient of binary component systems: using molecular
descriptors
ANNISHH BEHHARA†1, DANUSH SAI RUDRAPATTI BADRINARAYANAN†2, IMRAN
RAHMAN2*, KARTHIKEYAN M2
1
Department of Chemical Engineering, Birla Institute of Technology and Science, Pilani, K K Birla
Goa Campus, Goa, India
2
Chemical Engineering and Process Development, CSIR – National Chemical Laboratory, Pune,
Maharashtra, India
*Corresponding author’s mail id: [email protected]
†
Equally contributed

Abstract:
Vapor-Liquid Equilibrium data is crucial for separation processes like distillation, extraction and manufacturing.
Obtaining this data experimentally for desired conditions and systems is time-consuming and expensive.
Therefore there is a necessity for an a priori generalized model which predicts this data based on the molecular
descriptor information (of the desired system) given as an input to the model. This model is based on the Non-
Random Two-Liquid (NRTL) model to predict binary interaction (NRTL) parameters. These predicted parameters
are, in turn, used to calculate the activity coefficient, which is used to calculate the vapor-phase composition from
the liquid-phase composition of the system. In this study, the molecular descriptors for individual components of
the 28 binary systems were generated. The arithmetic mean of the molecular descriptors of the corresponding
components was used as the molecular descriptor set for that binary system. The molecular descriptors based on
properties relevant to vapor-liquid equilibrium were selected and used as independent variables to build the model
using an Artificial Neural Network (ANN) in python. Better predictions were obtained with the coefficient of
determination greater than 0.85 for each NRTL parameter. Once the liquid-phase composition is known, the model
can predict the vapor-phase composition at the desired pressure and temperature.
Keywords: Non-Random Two Liquid (NRTL), Vapor-Liquid Equilibrium, Artificial Neural Networks,
Molecular Descriptors, Binary Component System

Introduction: UNIQUAC, and UNIFAC. In contrast, non-ideality

in the vapor phase is described using equations of
Information about thermodynamic properties state models such as the Van der Waals Equation,
and phase equilibria of mixtures are essential for Redlich Kwong Equation, Redlich Kwong Soave
designing industrial processes like distillation, Equation, and Peng Robinson Equation at higher
extraction, separation, designing new cost-effective pressures2.
processes, synthesis of new materials, and
technological advancements. Experimenting to Ravindranath3 developed a structure-based
determine properties takes time, money, energy and generalized model for predicting pure-fluid vapor
effort1. As a result, theoretical and empirical pressures, saturated phase densities, and binary
thermodynamic property models are required, with mixture vapor-liquid equilibrium. Linear and non-
experimental data used to predict properties. linear (back-propagation network) NRTL and
However, these models only succeed when UNIQUAC models were used. Gebreyohannes4
insufficient or no experimental data is available, so worked on a generalized predictive model for Non-
developing a generalized property model to predict Random Two-Liquid binary interaction parameters.
properties is critical.
The molecular descriptor values of the binary
The fugacity and activity coefficient must be systems were taken as the absolute difference
considered since, in reality, most mixtures behave between the corresponding molecular descriptor
non-ideally. However, the method for calculating values of the component molecules in the systems.
phase equilibrium in systems that are non-ideal in Furthermore, in this study, a modified NRTL model
the liquid phase only is based on activity coefficient was taken, in which only one parameter was used,
models such as Margules, Van Laar, Wilson, NRTL, which was used to prevent different prediction

1
 https://doi.org/10.36375/prepare_u.iiche.a366

models for different binary interaction parameters molecules. Because of this reason, aij and aji are not
(for example, different models for aij and aji), as that equal to each other. Similarly, bij, cij and bji, cji and are
could give different values for the same parameter not equal.
depending on the order of components taken. In the
present work, the molecular descriptor values of the
binary systems were taken as the arithmetic mean of Materials and methods:
the corresponding molecular descriptor values of the A set of 28 binary mixtures containing Acetone,
component molecules in the systems. A standard Benzene, Chloroform, Ethanol, Methanol, Toluene,
NRTL model was used in the present work. Water and p-Xylene are considered. The SMILES
Different predicting models were made for different format files of molecular properties for each
binary interaction parameters by taking a particular component in the systems were obtained from the
order of components based on their boiling points. ChEMBL online database6. The SMILES format
The “i” component was considered to have the lower files were used as input for the paDEL software7,
boiling point, and the “j” component was considered which was used to generate the molecular descriptors
to have the highest boiling point between the data for each component in the systems. 1444
system's two components. molecular descriptors were obtained in total. The
arithmetic mean of the corresponding molecular
The local composition method parameters are descriptor values of the components of the binary
temperature-dependent, which makes them systems was taken as their molecular descriptor
applicable to any temperature. In contrast, the values. This was done to make the model applicable
parameters apply only to a given temperature range to systems of the same component.
in simple empirical methods. The local composition A set of relevant molecular descriptors (those
representing chemical bonds and functional groups)
method's multi-component behaviour can be
were shortlisted for this work based on their
obtained from binary data. In contrast, simple
properties. These molecular descriptors’ data were
empirical methods are restricted to binary systems. preprocessed by discarding those descriptors whose
The local composition methods consider that local value was zero (0) for all the binary systems and
compositions are different from bulk composition, those that had ‘NaN’ values. A final set of 162
which is the case when the interaction among similar molecular descriptors was obtained, and these
molecules is different from between molecules of descriptors were the independent variables, and the
one component and the other. In contrast, the simple NRTL parameters (aij, bij, cij, aji, bji and cji) were the
empirical methods do not differentiate between local dependent or target variables in the respective
and bulk compositions5. The present work obtained models.
the activity coefficient using a non-random two- While making the 28 binary systems (from
liquid (NRTL) local composition model. Acetone, Benzene, Chloroform, Ethanol, Methanol,
Toluene, Water and p-Xylene), “i” component was
considered to have the lower boiling point and “j”
∑𝑗 𝑥𝑗 𝜏𝑗𝑖 𝐺𝑗𝑖 𝑥𝑗 𝐺𝑖𝑗 ∑𝑚 𝑥𝑚 𝜏𝑚𝑗 𝐺𝑚𝑗 component was considers to have the highest boiling
ln 𝛾𝑖 = + ∑ (𝜏𝑖𝑗 − )
∑𝑘 𝑥𝑘 𝐺𝑘𝑖 ∑𝑘 𝑥𝑘 𝐺𝑘𝑗 ∑𝑘 𝑥𝑘 𝐺𝑘𝑗 point between the two components of a system.
𝑗
Other ways might be to take the reverse of it or use
𝑏𝑖𝑗
Where, 𝜏𝑖𝑗 = 𝑎𝑖𝑗 + and 𝐺𝑖𝑗 = 𝑒 −𝑐𝑖𝑗 𝜏𝑖𝑗 any other property (like the molecular weight of the
𝑇(𝐾)
molecule, etc.) to assign “i” and “j”. The purpose of
classifying “i” and “j” based on a specific property,
The activity coefficient of component ‘i’ is
and not randomly chosen, is to have uniformity in the
represented by γi. The binary interaction parameters
model and to get more ordered results. The binary
for the local domain are aij, bij, and cij, with the
interaction parameters database for the 28 binary
molecule of component ‘j’ as the center surrounded
compound systems was taken from Appendix B8. An
by molecules of component ‘i’. Gij and τij represent
Artificial Neural Network (ANN) was trained in
the adjustable parameters for the local domain, with
Python (Jupyter Notebook - Anaconda platform) for
molecules of component ‘j’ in the center and
the NRTL parameters aij, bij, aji, bji, cij and cji to
molecules of component ‘i’ surrounding it. The
determine the relationship between these parameters
component's mole fraction in the liquid phase is
and the appropriate molecular descriptor set.
represented by 'x.'
The prediction model for the parameters aij, bij and
cij correspond to the local domain where the “j” Results and Discussions:
component molecule is at the center, surrounded by
the “i” component molecules and the prediction The Train dataset consists of 16 binary systems. The
model for the parameter aji, bji and cji corresponds to Validation dataset consists of 6 binary systems, and
the local domain where the “i” component molecule the Test dataset consists of 6 binary systems.
is at the center, surrounded by the “j” component

2
https://doi.org/10.36375/prepare_u.iiche.a366

Initially, when the first few random states (random relationship generalized activity
numbers that vary the composition of train and test coefficient models.
datasets) were used (starting from 0), negative https://shareok.org/handle/11244/14847
values of coefficient of determination (R2, a metric (3) Gow, A. S. (1993). Calculation of Vapor-
to measure the accuracy of the prediction, ranging Liquid Equilibria from Infinite-Dilution
from 0 to 1, where 1 stands for highest accuracy) Excess Enthalpy Data Using the Wilson or
were obtained, which occurred due to insufficient or NRTL Equation. Industrial and
incomplete learning from the train dataset used in Engineering Chemistry Research, 32(12),
that random state. Then, iterations were run up to 3150–3161.
10,000 random states for ANN. The Train and Test https://doi.org/10.1021/IE00024A028/AS
dataset combination that gave the highest coefficient SET/IE00024 028.FP.PNG_V03
of determination value was taken to train the model. (4) Ravindranath, D. (2005). Structure-based
Generalized Models for Pure-fluid
Table 1. Parameters and their Corresponding Coefficient
Saturation Properties and Activity
of Determination and Random States
Coefficients.
Binary Coefficient of Random https://shareok.org/handle/11244/9662
(5) Vetere, A. (2004). The NRTL equation as
Interaction Determination State
a predictive tool for vapor–liquid
Parameter equilibria. Fluid Phase Equilibria, 218(1),
aij 0.989 1670 33–39.
aji 0.880 3074 https://doi.org/10.1016/J.FLUID.2003.10.
bij 0.955 9053 013
bji 0.873 4566 (6) Mendez D, Gaulton A, Bento AP,
cij 0.904 8243 Chambers J, De Veij M, Félix E,
Magariños MP, Mosquera JF, Mutowo P,
cji 0.886 2158
Nowotka M, Gordillo-Marañón M, Hunter
F, Junco L, Mugumbate G, Rodriguez-
Conclusions: Lopez M, Atkinson F, Bosc N, Radoux
CJ, Segura-Cabrera A, Hersey A, Leach
The coefficient of determination obtained with AR, Nucleic Acids Res. 2019; 47(D1):
ANN was greater than 0.85 for all NRTL parameters. D930-D940.
Thus, these models can be used to get reliable https://doi.org/10.1093/nar/gky1075
predictions of the NRTL parameters of binary
systems when their molecular descriptors data is (7) Yap CW (2011). PaDEL-Descriptor: An
given as an input. When the liquid-phase open-source software to calculate
composition and temperature are known, the molecular descriptors and
predicted NRTL parameters can be used to calculate fingerprints. Journal of Computational
the activity coefficients of the binary systems using Chemistry. 32 (7): 1466-1474
the NRTL equation, which can then be used to (8) Beneke, Glasser, Hildebrandt, Peters.
calculate the vapor-phase composition using the (2013) Appendix B In: Understanding
modified Lewis-Randall rule. Distillation Using Column Profile Maps.
John Wiley & Sons, Inc., pp. 345-347.

References:
(1) Del-Mazo-Alvarado, O., & Bonilla-
Petriciolet, A. (2022).Assessment of the
simultaneous regression of liquid-liquid
and vapor-liquid equilibria data of binary
systems using NRTL and artificial neural
networks. Fluid Phase Equilibria, 561,
113537.
https://doi.org/10.1016/J.FLUID.2022.113
537
(2) Gebreyohannes, S., Robinson, R. L.,
Ramsey, J. D., & Hagan, M. (2014).
Quantitative structure-property