Received: 3 July 2022 Revised: 23 September 2022 Accepted: 27 September 2022

DOI: 10.1002/qua.27026

RESEARCH ARTICLE

Machine learning approach for the prediction of surface

tension of binary mixtures containing ionic liquids using
σ-profile descriptors

Widad Benmouloud | Cherif Si-Moussa | Othmane Benkortbi

Biomaterials and Transport Phenomena

Laboratory (LBMPT), Department of Process Abstract
and Environmental Engineering, University of
Ionic liquids (IL) are a new class of liquids considered as green solvents; less toxic, less
Yahia Fares Medea, Medea, Algeria
flammable, and less polluting which retain their liquid state over wide temperature
Correspondence
ranges and are considered alternatives to volatile organic solvents. The surface ten-
Widad Benmouloud, Biomaterials and
Transport Phenomena Laboratory (LBMPT), sion of IL-organic solvent mixtures plays an important role in the design and develop-
Department of Process and Environmental
ment of many industrial processes. This work investigated the capability and
Engineering, University of Yahia Fares Medea,
Medea 26000, Algeria. feasibility of four ANN model topologies (“trainbr, logsig”; “trainbr, tansig”; “trainlm,
Email: [email protected]
logsig”; “trainlm, tansig”), a PSO-SVM model, and an LSSVM model to predict the
Funding information surface tension of binary systems containing IL. For this purpose, 1623 data points
Algerian Ministry of Higher Education and
corresponding to the experimental surface tension values of binary mixtures con-
Scientific Research, Grant/Award Number:
A16N01UN260120220003; University Yahia taining IL were collected from the literature. The surface tension values were
between 18.9 and 72.7 mN m
1. The temperature, the composition in mole fraction
Fares of Medea

of IL (XIL), descriptors based on the sigma profiles, relating to the H-bond donor and
to the H-bond acceptor character, the anion, the cation and the solvent were used as
input variables of the model in order to differentiate the different compounds
involved in the binary systems. A comparison of the experimental and the predicted
values in terms of several statistical metrics showed good agreement, however, the
prediction (trainbr, logsig) was better than the other approaches with an overall aver-
age absolute relative deviation of .8466% and a mean square error of .4952. These
results are very encouraging for future projects modeling other physical and chemical
properties of ILs.

KEYWORDS
artificial neural networks, ionic liquids, least-squares support vector machine, support vector
machine-particle swarm optimization, surface tension, σ-profile descriptor

1 | I N T RO DU CT I O N

In recent years, the situation has evolved considerably in the field of the use of organic molecular solvents in various industrial chemical processes.
Due to their harmfulness (toxicity, flammability, and volatile organic compounds (VOC) emission), stringent regulations aim to limit the use of sol-
vents presenting dangers to human health and the environment [1]. Ionic liquids (ILs) have attracted much attention from the scientific community
over the past two decades due to their wide variety of applications in many fields of chemistry and chemical industry [2]. They are a new class of
liquids, considered as green solvents, that retain their liquid state over wide temperature ranges [1, 3], with high solvation properties, negligible
vapor pressure, and high thermal, chemical, and electrochemical stability [4]. Potential applications of ILs require knowledge of physicochemical

Int J Quantum Chem. 2022;e27026. http://q-chem.org © 2022 Wiley Periodicals LLC. 1 of 22

https://doi.org/10.1002/qua.27026
2 of 22 BENMOULOUD ET AL.

properties such as density, viscosity, melting point, solvent properties, vapor pressure and surface tension for pure ILs and their mixtures with
other solvents [5].
As there are countless combinations of cations and anions that form an IL, the synthesis of all the IL resulting from all these possible combina-
tions is practically impossible. Moreover, measurements of all properties for all synthesized IL and their mixtures are laborious, time consuming
and costly. Therefore, it is necessary to develop reliable models that could be considered as a suitable alternative to experimental measurements
for predicting the properties of IL mixtures under various conditions [6].A key property in various fields such as oil, gas and chemical industries, is
the surface tension of pure and mixed systems [7]. This property plays a particular role in process design by affecting mass and heat transfer at
the interface [8, 9].
Several authors have developed predictive models of the surface tension of pure ILs, but few studies have been carried out on their mixtures.
Huang et al [6]. used in their studies, semi-empirical models and artificial neural network (ANN) models for the prediction of the surface tension
of mixture of ILs. They found that the overall average absolute relative deviation (AARD) of the semi-empirical and ANN models is less than 2%.
The study conducted by Atashrouz et al [3]. for the prediction of physico-chemical properties of IL mixtures such as surface tension has allowed
the development of flexible computational approaches based on support vector machine (SVM), least square support vector machine (LSSVM)
and the group method of data processing type polynomial neural network systems (GMDH-PNN). They found that the LSSVM model is more
robust and reliable in predicting physicochemical properties of IL mixtures with AARD of 1.18% for surface tension.
Hashemkhani et al [10]. used three methods namely SVM, CSA-LSSVM and GA-LSSVM to predict the surface tension of binary mixtures con-
taining ILs using 748 data points. They obtained a better precision in the case of CSA-LSSVM model where the average % AARD is 1.3785. A
modeling method based on ANN trained by Bayesian regulation back propagation training algorithm (trainbr) has been proposed by Soleimani
et al [11]. to predict surface tension of the binary IL mixtures. A comparison with different models such as SVM, GA-SVM, GA-LSSVM, CSA-
LSSVM, and GMDH-PNN was carried out. They concluded that the proposed model was better in terms of accuracy with an average
AARDof .44%.
In a more recent paper, Cardona and Valderrama [12] proposed a modeling approach based on a cubic equation of state and the concept of
geometric similarity for predicting the surface tension of pure substances and mixtures containing organic substances, water and ILs of 90 mixtures
binary (2660 data) and 12 ternary mixtures (467 data) considered in a wide range of temperatures from 278.15 to 348.15 K. For primary estima-
tion, they concluded that the model is accurate in design and process simulation.
The use of alternative and complementary methods to experimentation such as quantitative structure–property/activity relationships (QSPR/
QSAR) became of great interest and is the most widely adopted methods to augment experimental analytical techniques [13]. The development
of QSPR/QSAR mathematical models linking physico-chemical properties and biological activities to a set of molecular descriptors allows to
explain the origin of these activities/properties and to predict them for molecules whose experimental data are not available [14].
Klamt et al [15]. developed a quantum chemical approach (COSMO-RS) for the prediction of thermodynamic properties of pure and mixed
polarity distribution. From the literature, the distribution area of the σ profile (Sσ profile) has been adopted as a quantitative measure representa-
tion of the polar surface screen charge of the molecule on the polarity scale, obtained from the histogram profile function σ given by the
COSMO-RS calculation [14, 16]. The modeling of the properties of ILs is mainly based on the use of equations of state and the application of
machine learning algorithms. These algorithms show that they have various applications in different fields such as the medical [17], the electrical
and electronic engineering [18], the petrochemical [19, 20], chemical engineering [21, 22] and the civil and environmental engineering [23].
Among the alternative methods of computational intelligence, are the ANN, the LSSVM, and the SVM fine tuning with particle swarm optimiza-
tion (SVM-PSO) are robust and accurate predictive methods that have recently been successfully applied for the prediction of various
properties [6].
The aim of this work is to evaluate the ability of different machine learning techniques to model surface tension for correlation and/or predic-
tion purposes. The study intends to estimate the surface tension of binary mixtures (IL-Water and various organic solvents) using three methods:
a predictive ANN model, a LSSVM and a SVM fine tuning with PSO (SVM-PSO). The same set of inputs is considered for both types of models
including temperature (T), composition of the mixture in mole fraction of IL (XIL), two descriptors based on COSMO-RS sigma profiles of anion,
cation and solvent are used as input variables of the model in order to differentiate the different compounds involved in the binary systems.

2 | MATERIALS AND METHODS

2.1 | Database

First of all, it should be mentioned that the experimental data used in this work are not exhaustive of all the data published in the literature. They
are limited to binary mixtures containing ILs for which the data of the sigma profiles of anions and cations are available in the published literature.
Thus, the experimental data, collected from numerous works, consists of 1623 surface tension (σ) data points relating to 62 binary mixtures of IL
and molecular solvent at different temperatures and compositions. The cations involved are: imidazolium (Im), pyridinium (Py), ammonium (N) and
BENMOULOUD ET AL. 3 of 22

phosphonium (P), whereas the anions are: tetrafluoroborate (BF4), hexafluorophosphate (PF6), bis(trifluoromethylsulfonyl)imide (BTI), bro-
mide (Br), chloride (CL), acetate (AC), alkyl sulfate (RSO4), dimethyl phosphate (DMP), trifluoromethylsulfonate (TfO), nitrate (NO3) and
dicyanamide (Dca). Molecular components intended to cover common solvents such as water, alcohols, dimethyl sulfoxide, acetonitrile and
tetrahydrofuran. Table 1 shows the source and domains of experimental data of the studied binary mixtures. The global database comprising
the experimental data and the calculated sigma profile descriptors of the cation, the anion and the solvent can be found in the supplementary
information file.

2.2 | COSMO-RS

Molecular descriptors based on σ profiles derive from the COSMO-RS theory considered as a continuous solvation model that combines quantum
chemical theory, dielectric continuum models, surface interactions and statistical thermodynamics [24].
It calculates molecular interactions from the shielding charge (polarization) densities on surface molecular segments. Quantum chemical calcu-
lations provide a discrete surface segment around a single molecule embedded in a virtual conductor [25]. The surface of each segment is charac-
terized by its area and the shielding charge density of the segment, which takes into account the electrostatic interaction of the solute molecule
by its environment and retro-polarization of the solute molecule [26]. The screening charge density is usually given in the probability diagram of
the statistical distribution of the charge density on the surface of a molecule, known as the σ profile which shows the probability of the relative
amount of area with the σ polarity at the surface of the molecule considered as the characteristic properties of the molecule [27].
In other words, the σ profile of a molecule includes the main chemical information necessary to predict possible electrostatic, hydrogen bond-
ing and dispersion interactions of the molecule in a fluid. σ-profiles have been shown to be an effective molecular descriptors for establishing
QSPR models able to predict the physical, chemical and toxicological properties of ILs [13].

2.3 | Models development

This step consists of the selection of the input variables, which are the independent variables of the model. In this respect, the temperature (T),
the mole fraction of IL (XIL), two descriptors based on the sigma profiles, one relating to the donor character of H-bond, the other relating to the
acceptor character of H- bond, for the anion, the cation and the solvent are used as input variables of the model in order to differentiate the dif-
ferent compounds involved in the binary systems given as follows:

σ m ¼ f ðT, XIL , Sσ1c , Sσ2c , Sσ1A , Sσ2A , Sσ1S , Sσ2S Þ ð1Þ

where Sσ1c and Sσ2c are the sigma profiles descriptors of the cation, Sσ1A and Sσ2A of the anion and Sσ1S and Sσ2S of the solvent.

2.3.1 | Artificial neural network models

Artificial neural networks are a computer model used to analyze data. Knowledge is acquired by the network through a learning process, and the
connecting strengths of interneurons called synaptic weights are used to store it [28]. Their most important advantages are: their ability to find
input traceability and their flexibility to test model interpolation, extrapolation and prediction [29]. ANNs, as parallel distributed systems, are gen-
erally composed of an input layer, some hidden layers and an output layer, each neuron is connected to the other neurons of a previous layer
thanks to synaptic weights adaptable. Knowledge is usually stored as a set of connection weights [28]. Multilayer perceptron (MLP) networks and
radial basis function (RBF) networks are two popular ANNs [30, 31].
The network determines the relationship between the variables and stores the values of the weights and biases that give the lowest error
between the calculated and experimental data of the dependent variable through an optimization process using traditional backpropagation algo-
rithms or evolutionary algorithms and genetic algorithms (GAs) [29, 32].

2.3.2 | Least squares support vector machine

The LSSVM which was suggested by Suykens and Vandewalle [33, 34] is considered as an alternative to the supervised SVM learning method pro-
posed by Vapnik in the 1990s used for classification and regression to analyze data and identify patterns [35, 36]. This new version replaces the
convex quadratic programming and inequality constraints of the original SVM by solving a linear set of equations and instead using equality
4 of 22 BENMOULOUD ET AL.

TABLE 1 Source and domains of experimental data of the studied binary mixtures

Solvent LI T (K) XLI N References

Water [C1MIm][DMP] 298.15 [.0000–1.0000] 16 [51]
[C1MIm][MeSO4] [296.8–298.1] [.0000–.2920] 10 [52]
[C2MIm][BF4] [298.15–338.15] [.0000–1.0000] 53 [7, 53]
[C2MIm][DEP] 298.15 [.0000–1.0000] 16 [51]
[C2MIm][DMP] 298.15 [.0000–1.0000] 11 [51]
[C2MIm][EtSO4] 298.15 [.0062–.5791] 12 [54]
[C2MIm][MeSO3] [300.2–303.2] [.0000–.4850] 27 [52]
[C4MIm] [AC] [298.15–338.15] [.8533–1.0000] 49 [12]
[C4MIm] [CL] [298.1–302.1] [.0000–.4800] 46 [12]
[C4MIm] [PF6] 303.15 [.9268–.9969] 9 [12]
[C4MIm][BF4] [298.15–338.15] [.0000–1.0000] 59 [7, 53]
[C4MIm][DBP] 298.15 [.0000–1.0000] 11 [51]
[C4MIm][DMP] 298.15 [.0000–1.0000] 11 [51]
[C4Py] [BF4] [293.15-323.15] [.0000–1.0000] 72 [12]
[C4Py][NO3] 298,15 [.0000–.9903] 15 [55]
[C6MIm] [AC] [298.15–338.15] [.8597–1.0000] 49 [12]
[C6MIm] [CL] [297.2–298.2] [.0000–.143] 10 [12]
[C6MIm][BF4] [298.15–338.15] [0, 3–1.0000] 40 [7, 53]
[C8MIm] [PF6] [298.05–335.05] .7827 5 [12]
[N112(hoe)][Br] 298.15 [.9915–.9981] 3 [56]
[N311(hoe)][Br] 298.15 [.9900–.9979] 3 [56]
[P666(14)][BTI] [298.1–343.3] .0891 6 [57]
[P666(14)][Dca] [298.2–342.8] .4932 6 [57]
Methanol [C1MIm][DMP] 298.15 [.0000–1.0000] 11 [51]
[C1MIm][MeSO4] 298.15 [.0000–1.0000] 9 [56]
[C2MIm] [AC] [278.2–318.15] [.0000–1.0000] 55 [12]
[C2MIm][DEP] 298.15 [.0000–1.0000] 22 [51]
[C2MIm][EtSO4] 298.15 [.0000–1.0000] 11 [12]
[C2MIm][MeSO4] [293.15–298.15] [.0000–1.0000] 21 [12, 58]
[C4MIm][BTI] [283.15-298.15] [.1049–.9577] 24 [12]
[C4MIm][DBP] 298.15 [.0000–1.0000] 11 [51]
[C4MIm][DMP] 298.15 [.0000–1.0000] 11 [51]
[C4Py] [BF4] [293.15–323.15] [.0000–1.0000] 56 [12]
[C8MIm] [BTI] [283.15–298.15] [.0978–.9368] 20 [12]
Ethanol [C1MIm][DMP] 298.15 [.0000–1.0000] 11 [51]
[C2MIm] [AC] [278.15–338.15] [.0000–1.0000] 77 [12]
[C2MIm][C8SO4] 298.15 [.0399–1.0000] 18 [59]
[C2MIm][DEP] 298.15 [.0000–1.0000] 11 [51]
[C2MIm][DMP] 298.15 [.0000–1.0000] 11 [51]
[C2MIm][EtSO4] 298.15 [.0000–1.0000] 11 [12]
[C2MIm][MeSO4] 293.15 [.0000–1.0000] 14 [12]
[C4MIm][BF4] 298.15 [.0986–1.0000] 12 [53]
[C4MIm][BTI] [283.15–313.15] [.0999–.9601] 36 [12]
[C4MIm][DBP] 298,15 [.0000–1.0000] 11 [51]
[C4MIm][DMP] 298.15 [.0000–1.0000] 11 [51]
[C6MIm] [EtSO4] 298.15 [.03–1.0000] 13 [12]
BENMOULOUD ET AL. 5 of 22

TABLE 1 (Continued)

Solvent LI T (K) XLI N References

[C6MIm][BF4] 298.15 [.0980–1.0000] 10 [53]

[C8MIm] [BTI] [283.15–313.15] [.1042–.9784] 39 [12]

[C8MIm][BF4] 298.15 [.1012–1.0000] 10 [53]
Tetrahydrofuran [C2MIm][BTI] [293.15–308.15] [.0000–1.0000] 44 [60]
[C4MIm][BTI] [293.15–308.15 [.0000–1.0000] 40 [61]
Acetonitrile [C2MIm][BTI] [293.15–313.15] [.0000–1.0000] 45 [60]
[C4MIm][BTI] [293.15–313.15] [.0000–1.0000] 45 [61]
[PYR-4,1] [BTI] [288.15–308.15] [.0000–1.0000] 27 [12]
Dimethyl sulfoxide [C2MIm][BTI] [293.15–313.15] [.0000–1.0000] 45 [61]
[C4MIm][BTI] [293.15–313.15] [.0000–1.0000] 50 [61]
1-Propanol [C2MIm] [AC] [288.15–348.15] [.0000–1.0000] 77 [12]
[C4MIm][BTI] [283.15–313.15] [.0000–1.0000] 31 [12, 62]
[C8MIm] [BTI] [283.15–313.15] [.0974–.9461 33 [12]
1-butanol [C4MIm][BTI] [283.15–313.15] [.0000–1.0000] 29 [12, 62]
[C8MIm] [BTI] [298.15–318.15] [.0000–1.0000] 48 [12]
1-Hexanol [C8iQuin]BTI] [298.15–318.15] [.0000–1.0000] 24 [63]

constraints in the LSSVM method, regression error is applied to the optimization settings. In fact, in SVM algorithms, the regression error is mini-
mized in the learning phase whereas in LSSVM methods, it is mathematically defined and solved [36].
The goal minimization principle in the LSSVM method can be expressed as the following cost function (1) subject to the consequent con-
straint (2):

1 1 XN
Minimize : cost function ¼ wT w þ γ e2 ð2Þ
2 2 i¼1 i

Subjected to : yi ¼ wT φðXi Þ þ b þ ei , i ¼ 1, 2, …N ð3Þ

where wT represents the transpose matrix of w, ei refers to regression errors, b is bias, γ denotes the regularization parameter which controls
errors, and the subscript “i” expresses data points for training, and N represents the total number of training points.
Equation (4) is the Lagrangian form, used to solve the LSSVM problem:

1 1 XN XN  
lðw, b, e, aÞ ¼ wT w þ γ e2k
ak wT φðxk Þ þ b þ ek
yk ð4Þ
2 2 i¼1 i¼1

which ai are Lagrangian multipliers.

The LSSVM problem is solved by equating the derivatives of the Lagrangian form to zero:

8 XN
> ∂L
>
> ¼0)w¼ a φðxk Þ
>
> ∂W k¼1 k
>
>
>
> X
N
>
> ∂L
>
< ∂b ¼ 0 ) ak ¼ 0
k¼1 ð5 – 8Þ
>
> ∂L
>
> ¼ 0 ) ak ¼ γek , k ¼ 1, 2, …, N
>
>
> ∂ek
>
>
>
>
> ∂L ¼ 0 ) wT φðx Þ þ b þ e
y ¼ 0, k ¼ 1, 2, …, N
: k k k
∂ak

The solution of the LSSVM problem is possible by solving the above mentioned set of linear equations instead of the quadratic
programming problem. SVM and LSSVM methods are kernel-based approaches. In this study, the RBF kernel function was applied according to
Equation (9):
6 of 22 BENMOULOUD ET AL.

 
kðx, xi Þ ¼ exp
kxi
xk2 =σ 2 ð9Þ

where kxi – xk is the Euclidean distance of the ith input from the center of xc. There are two tuning parameters in LSSVM, that is, γ and σ 2. The
parameters are tuned by minimizing the differences between the predicted values and their corresponding experimental values.

2.3.3 | Support vector machine

Support vector machine as an efficient type of supervised machine learning method, which was developed by Vapnik [37], is used for solving clas-
sification and nonlinear regression problems [38]. SVM has many attributes including good generalization ability to avoid over-fitting based on
regularization, nonlinear classification ability based on kernel trick, and global error minimization based on convex optimization [39]. The regres-
sion version of SVM is called SVR with the central goal of finding the best line of fit in the hyperplane [40]. The prediction or approximation func-
tion used by a basic SVM is [41]:

X
l
f ðxÞ ¼ αi kðx, xi Þ þ b ð10Þ
i¼1

where xi is a feature vector corresponding to a training object, K(x, xi) is a kernel function and αi is some real value. The component of vector α
and the constant b represent the hypothesis and are optimized during the training. K(x, xi) is a kernel function, which value is equal to the inner
product of two vectors x and xi in the feature space Φ(x) and Φ(xi), that is,

K ðx, xi Þ ¼ ΦðxÞ  ΦðxiÞ ð11Þ

For a dataset, only the kernel function and the regularity parameter C should be selected to specify an SVM. Any function that satisfies Mer-
cer's condition could be used as a kernel function. The Gaussian kernel:

 
K ðu, v Þ ¼ exp
γ  ju
vj2 ð12Þ

is the most commonly used in support vector regression.

2.3.4 | Particle swarm optimization

In 1995, PSO was introduced by Kennedy and Eberhart based on a social simulation model known as the stochastic optimization algorithm, inspired
by the social behavior of flocks of birds or schools of fish. PSO is similar to GAs, in terms of initializing the population with random solutions and find-
ing the optimum by updating the generations. However, unlike GA, PSO does not undergo crossover or mutation, as the particles move through the
problem space following the current optimal particles. Since there are only a few parameters to set in PSO, it is easy to implement [42]. The underly-
ing concept is that these particles move over the search area with flexible speed and maintain the best position they have discovered in the search
space. Each particle can revise its velocity vector to explore the best position through its flight expertise and the flight expertise of other particles in
the search space [19]. Mathematically, a swarm of particles is randomly initialized on the search space and moves through the D-dimensional space
to search for new solutions. Let xik and vik respectively be the position and the velocity of the ith particle in the search space at the kith iteration, then
its velocity and the position of this particle at (k + 1) the iteration are updated using the following equations [42]:
   
vi kþ1 ¼ w:v i k þ c1 :r 1 : pi k
xi k þ c2 :r2 : pg k
xi k ð13Þ

xi kþ1 ¼ xi k þ v i kþ1 ð14Þ

where r1 and r2 represent random numbers between 0 and 1, c1 and c2 are constants, pik represents the best ever position of ith particle, and pgk
corresponds to the global best position in the swarm up to kith iteration.

2.4 | Statistical metrics

To assess the efficiency and the accuracy of the ANN models, several statistical parameters were used, namely, AARD percentage (AARD%), mean
squared error (MSE), mean relative squared error (MRSE), Q2 measurements are briefly presented by three indicators (QF12, QF22, and QF32)
BENMOULOUD ET AL. 7 of 22

and the concordance correlation coefficient (Q2ccc), determination coefficient (R2), correlation coefficient (R), accuracy factor (Af), the bias factor
(Bf) and Akaike's information criterion (AIC).
Where N is the number of data points, Np is the number of parameters in the model and SSE is the sum of the squared error, yiexp is the value
of experimental data sets at the sampling point i, yical is the ith value of the corresponding predicted sampling point i, y exp , and ycal are the average
of the experimental and predicted data, k & k' are the slopes of the corresponding regression lines, r 02 is the square of the correlation coefficient
between the observed value and the predicted value of compounds without an intercept, r 0 02 has the same meaning as r 02, except that it uses the
axes reversed. The mathematical equations of the above mentioned parameters, given by [11, 43–45], are as follows:

1X N  2
MSE ¼ y exp
yi cal ð15Þ
N i¼1 i

N  exp 
1X cal 
yi
yi   100
AARDð%Þ ¼ ð16Þ
N i¼1  yi exp 

N  exp 2
1X yi
yi cal
MRSE ¼ ð17Þ
N i¼1 yi exp

OUT 
nP 2
yi
b
yi=i
QF1 2 ¼ 1
ni¼1 ð18Þ
POUT
ðyi
yTR Þ2
i¼1

OUT 
nP 2
yi
b
yi=i
i¼1
QF2 2 ¼ 1
nOUT ð19Þ
P
ðyi
yOUT Þ2
i¼1

OUT 
nP 2
yi
b
yi=i =nOUT
2 i¼1
QF3 ¼ 1
ð20Þ
P
nTR
ðyi
yTR Þ2 =nTR
i¼1

nP
OUT  
2: ðyi
yEXP Þ: b
yi=i
yCAL
i¼1
Q2 ccc ¼ nOUT OUT 
ð21Þ
P nP 2
ðyi
yEXP Þ þ 2
b
yi=i
yCAL þ nOUT :ðyEXP
yCAL Þ2
i¼1 i¼1

N 
P 2
yi exp
yi cal
R2 ¼ 1
i¼1
N 
ð22Þ
P 2
yi exp
y exp
i¼1

PN    
yi exp
y exp  yi cal
ycal
i¼1
R ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð23Þ
P N  2 PN  2
yi exp
y exp  yi cal
ycal
i¼1 i¼1

P exp cal
y yi
k¼ Pi 2
ð24Þ
ðyi cal Þ

P exp cal
y yi
k0 ¼ P i ð25Þ
ðyi exp Þ2

P  cal 2
yi
kyi cal
Ro 2 ¼ 1
ð26Þ
P  cak 2
yi
yi cal
8 of 22 BENMOULOUD ET AL.

P  exp 2
yi
k0 yi exp
Ro 02 ¼ 1
P  2 ð27Þ
yi exp
y exp

R2
R0 2
m¼ ≤ :1 ð28Þ
R2

R2
Ro 02
n¼ ≤ :1 ð29Þ
R2

 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 
 ffi
Rm 2 ¼ R2  1
R2
Ro 2  ≥ :5 ð30Þ

1X N
y exp ¼ y exp ð31Þ
N i¼1 i

1X N
ycal ¼ y cal ð32Þ
N i¼1 i

0  1
P N  ycali 

B i¼1  yi C
log exp
B C
@ N A
Af ¼ 10 ð33Þ

0P ycal
1
N i
log exp
@ i¼1
N
y
i A
Bf ¼ 10 ð34Þ

 
SSE 2Np ðNp þ 1Þ
AIC ¼ Nln þ 2Np þ ð35Þ
N N
ðNp þ 1Þ

N 
X 2
pred
SSE ¼ i
yi
yobs ð36Þ
i¼0

The prediction of the model is ideal if AARD%, MSE, MRSE, R2, R, Af, and Bf are found to be very close to 0, 0, 0, 1, 1, 1, and 1 respectively. A
small AIC value indicates the better fitted model.
These metrics were used as acceptability criteria, as shown in Table 2.

3 | RESULTS AND DISCUSSION

Experimental data of surface tension in binary mixtures are published at different temperatures and mixture compositions. The cations and anions
constituting the ILs considered in this study are listed in Table 3.

TABLE 2 Acceptability criteria (AC) of a model

Parameters m, n Q2 R2 k, k' Bf, Af

AC values <.1 >.5 >.6 .85 < < 1.15 .9–1.05
BENMOULOUD ET AL. 9 of 22

3.1 | Data from sigma profile descriptors

The first step in the development of the model consists of defining appropriate molecular descriptors from the σ profiles of the cations, anions
and the solvents of the binary mixtures studied. Usually, the σ profile of a solute or solvent is divided into four sections; each section is defined by
an interval of σ (e/nm2). In this work, the sigma profile of IL solvents, anions and cations is divided into two sections. The σ profile descriptor, Sσ ,
previously mentioned, represents the area under the curve of the σ profile as shown in Figure 1. Thus, S σ1 , S σ2 represent the hydrogen bond
donor and acceptor character respectively.
In this work, a MATLAB program was used to calculate the areas Sσi (i = 1,…,6) of the considered cations, anions and solvents after digitizing
the σ profile curves obtained from the reference [46] and the digitized σ profiles for the water and organic compounds. The solvents obtained
from the bases VT-2005 and VT-2006, respectively [27, 47].

3.2 | Dividing data into training set, test set and validation set

The training dataset consists of 70% randomly chosen data using the divisor function of MATLAB 2018. The remaining 30% was used as a test
set for LSSVM and SVM-PSO. In the case of ANN modeling, these 30% were divided into two sets of 15% for testing and 15% for validation. The
purpose of this division is to avoid over-fitting during training.

TABLE 3 List of cations and anions

Cation name Abbreviations Anion name Abbreviations

1,3-dimethylimidazolium [C1MIm] Dimethylphosphate [DMP]
1-Ethyl-3-methylimidazolium [C2MIm] Methylsulfate [MeSO4]
1-Butyl-3-methylimidazolium [C4MIm] Tetrafluoroborate [BF4]
1-Hexyl-3-methylimidazolium [C6MIm] Bis(trifluoromethylsulfonyl)imide [BTI]
1-octyl-3-methylimidazolium [C8MIm] Octyl sulfate [C8SO4]
1-butylpyridinium [C4Py] Diethyl phosphate [DEP]
N-octyILsoquinolinium [C8iQuin] Ethyl sulfate [EtSO4]
Ethyl(2-hydroxyethyl)dimethylammonium [N112(hoe)] Methane sulfonate [MeSO3]
Propyl(2-hydroxyethyl)dimethylammonium [N311(hoe)] Methylsulfate [MeSO4]
trihexyltetradecylphosphonium [P666(14)] Dibutyl phosphate [DBP]
Nitrate [NO3]
Bromide [Br]
Dicyanamide [Dca]
1-butyl-1-methylpyrrolidinium [PYR-4,1] Acetate [AC]
Chloride [CL]
Hexafluorophosphate [PF6]

FIGURE 1 Molecular descriptor based on σ profiles.

10 of 22 BENMOULOUD ET AL.

3.2.1 | ANN modeling

Based on many previous studies, the best algorithms were used in this study such as: the Levenberg-Marqhardt backpropagation learning algo-
rithm (MATLAB function trainlm) and the Bayisian normalization algorithm (MATLAB function trainbr). Concerning the activation function, the
logsig and tansig functions for the hidden layers and the purelin function for the output layer were tested.
A program on MATLAB 2018 was developed, in which a modification of the number of neurons in the hidden layer, the learning algorithm
and the function of activating the hidden layer was made. The program was performed more than 20 times for the same structure to obtain the
model that gives the best MSE. The performance of model was evaluated using statistical metrics mentioned above and graphical tools.

3.2.2 | LSSVM modeling

LSSVM modeling was performed using the LS-SVMlab toolbox. To do this, an appropriate algorithm implemented in LS-SVMlab detects and scales
continuous, categorical and binary variables. A step called tuning consists in determining the LSSVM parameters (regularization constant γ and
kernel parameter σ2) by minimizing a selected performance measure. The constants α and b of the LSSVM model were determined from the learn-
ing step which uses the regularization constant γ and the kernel parameter σ2 adjusted in the first step. The last step consisted in testing the gen-
eralization capacity of the model on all the data reserved for the test. The optimized parameters were 3548.43 and 1.6510 for γ and σ 2 ,
respectively.

3.2.3 | SVM-PSO modeling

The development of the SVM-PSO model is carried out by the fitrsvm function of MATLAB environment. This function has several calculation
options relating to the choice of the cross-validation method, the kernel function (gaussian, rbf, and polynomial), the optimal values of the con-
stant C (boxConstraint), epsilon and the parameter of the kernel function (KernelScale) of the model were optimized with PSO algorithm (100 iter-
ations). In this study, we have chosen the Cross-validation method « holdout » which was set at .3 which means 70% of the data is used for
learning and 30% for validation.
The optimal values are 200, 1 and .006 for c, k and ep respectively with gaussian as Kernel function.
Table 4 summarizes respectively the calculation of the different statistical metrics for the four topologies of the ANN model (“trainbr, logsig”;
“trainbr, tansig”; “trainlm, logsig”; “trainlm, tansig”), for the SVM-PSO model and for the LSSVM model.
Table 5 summarizes respectively the calculation of the different errors for the four topologies of the ANN model, for the SVM-PSO model
and for the LSSVM model.
In the light of these results, all the calculated statistical metrics verify the conditions of acceptability. However, the results for the model opti-
mized by using the trainbr training algorithm and the logsig activation function showed the best significant results: a global correlation coefficient
R of .9979, the coefficient of determination R2 of .9958 and the accuracy factor Af of 1.0085. In addition, the error parameters such as: the MSE
(.4952) and the AARD% (.8466). This makes it possible to prove the efficiency and robustness of the model developed, which has more precise
correlation performance and much better generalization and interpolation capabilities. From Table 6, the smallest value obtained of AIC repre-
sents the best fitting model for the six models studied.
Graphical evaluation was performed using scatterplots (for training, testing, validation, and whole data set), % relative error (ER%) plot, and %
relative error (ER%) distribution plot. Figure 2 shows the scatter plot of the calculated surface tension as a function of the experimental surface
tension for the four topologies of the ANN model, the SVM-PSO model and the LSSVM model for the global ensemble. Figure 3 represents the
scatter plot of the calculated surface tension as a function of the experimental surface tension for the optimized model trainbr and logsig. This dia-
gram clearly shows the concordance between the calculated values and the experimental values by a tight dispersion of the points on the first
bisector. However, the existence of a number of outliers was noticed.
In order to quantitatively visualize the distribution of relative errors, two types of diagrams were adopted. The first shows the relative error
as a function of the surface tension represented in Figure 4 and the second shows the distribution of this error represented in Figure 5. This
shows that the majority of the points are located around the zero line. Quantitatively, the relative error for the few outliers did not exceed 20%.
To confirm the generalization capacity of the developed model, we carried out interpolation and extrapolation calculations for the composi-
tion and the temperature. Figure 6 shows the isotherms of the variation in surface tension as a function of the composition of the [C2MIm][BTI]-
Dimethyl sulfoxide mixture for the established model, the interpolated points follow the shape of the calculated points. In the same way, Figure 7
represent the interpolation and extrapolation calculation of the temperature. The interpolation (305.15 K) and extrapolation (290.15 and
315.15 K) isotherms approximately follow the shape of the calculated points and have the same shapes of the adjacent experimental isotherms.
BENMOULOUD ET AL. 11 of 22

TABLE 4 Statistical metrics of the four ANN model topologies, the SVM-PSO model, and the LSSVM model

(a): Trainbr, logsig (b): Trainbr, tansig

Statistical metric Train Test Global Statistical metric Train Test Global
Q2F1 .9955 .9964 .9958 Q2F1 .9951 .9962 .9955
Q2F2 .9956 .9964 .9958 Q2F2 .9952 .9962 .9955
Q2F3 .9955 .9965 .9958 Q2F3 .9951 .9958 .9953
Q2CCC .9978 .9982 .9979 Q2CCC .9976 .9981 .9977
2 2
R .9955 .9964 .9958 R .9951 .9962 .9955
R .9978 .9982 .9979 R .9976 .9982 .9977
K 1.0000 .9994 .9999 K 1.0000 1.0030 1.0008
k' .9996 1.0003 .9998 k' .9996 .9967 .9988
2 2
R0 1.0000 1.0000 1.0000 R0 1.0000 .9999 1.0000
R0'2 1.0000 1.0000 1.0000 R0'2 1.0000 .9999 1.0000
M
.0045
.0036
.0043 M
.0049
.0037
.0046
N
.0045
.0036
.0043 N
.0049
.0037
.0046
Rm .9291 .9365 .9309 Rm .9258 .9358 .9284
Af 1.0080 1.0098 1.0085 Af 1.0083 1.0116 1.0091
Bf 1.0001 1.0003 1.0002 Bf 1.0001 .9983 .9996

(c): Trainlm, logsig (d): Trainlm, tansig

Statistical metric Train Test Global Statistical metric Train Test Global
Q2F1 .9946 .9907 .9934 Q2F1 .9953 .9941 .9949
Q2F2 .9947 .9906 .9934 Q2F2 .9953 .9941 .9949
Q2F3 .9946 .9907 .9935 Q2F3 .9953 .9938 .9949
Q2CCC .9953 .9973 .9967 Q2CCC .9977 .9971 .9975
R2 .9946 .9906 .9934 R2 .9953 .9941 .9949
R .9973 .9953 .9967 R .9977 .9971 .9975
K .9999 1.0002 1.0000 K 1.0001 1.0013 1.0005
k' .9997 .9990 .9995 k' .9995 .9982 .9991
R02 1.0000 1.0000 1.0000 R 02 1.0000 1.0000 1.0000
R0'2 1.0000 1.0000 1.0000 R0'2 1.0000 1.0000 1.0000
M
.0054
.0095
.0066 M
.0047
.0059
.0051
N
.0054
.0095
.0066 N
.0047
.0059
.0051
Rm .9217 .8945 .9129 Rm .9272 .9182 .9242
Af 1.0101 1.0150 1.0116 Af 1.0091 1.0128 1.0102
Bf 1.0003 .9990 .9999 Bf .9999 .9981 .9993

(e): LSSVM modeling (f): SVM-PSO modeling

Statistical metric Train Test Global Statistical metric Train Test Global
Q2F1 .9906 .9408 .9754 Q2F1 .9721 .9376 .9614
Q2F2 .9907 .9406 .9754 Q2F2 .9722 .9374 .9613
Q2F3 .9394 .9906 .9752 Q2F3 .9721 .9343 .9607
Q2CCC .9953 .9694 .9875 Q2CCC .9855 .9677 .9800
R2 .9906 .9406 .9754 R2 .9721 .9374 .9613
R .9953 .9699 .9876 R .9863 .9687 .9808
K 1.0002 1.0011 1.0005 K 1.0064 1.0081 1.0069
k' .9991 .9942 .9976 k' .9915 .9870 .9901
R02 1.0000 1.0000 1.0000 R 02 .9995 .9991 .9994
'2
R0 1.0000 .9996 .9999 R0'2 .9991 .9979 .9988

(Continues)
12 of 22 BENMOULOUD ET AL.

TABLE 4 (Continued)

(e): LSSVM modeling (f): SVM-PSO modeling

Statistical metric Train Test Global Statistical metric Train Test Global
M
.0095
.0631
.0252 M
.0282
.0658
.0396
N
.0095
.0627
.0251 N
.0278
.0645
.0389
Rm .8947 .7114 .8225 Rm .8112 .7046 .7739
Af 1.0082 1.0219 1.0123 Af 1.0084 1.0217 1.0124
Bf 1.0003 1.0004 1.0003 Bf .9967 .9928 .9955

TABLE 5 Different errors of the four ANN model topologies, the SVM-PSO model and the LSSVM model

Trainbr, logsig Trainbr, tansig

Errors Train Test Global Errors Train Test Global

AARD% .7992 .9812 .8466 AARD% .8256 1.1526 .9108
MRSE .7235 .6441 .7037 MRSE .7416 .6917 .7289
MSE .5235 .4149 .4952 MSE .5499 .4785 .5313

Trainlm, logsig Trainlm, tansig

Errors Train Test Global Errors Train Test Global

AARD% 1.0111 1.4779 1.1511 AARD% .9102 1.2568 1.0141
MRSE .7936 1.0421 .8756 MRSE .7346 .8414 .7682
MSE .6298 1.0860 .7667 MSE .5396 .7079 .5901

LSSVM modeling SVM-PSO

Errors Train Test Global Errors Train Test Global

AARD% .8150 2.1653 1.2201 AARD% .8050 2.0399 1.1755
MRSE 1.0430 2.6514 1.6942 MRSE 1.7923 2.7482 2.1247
MSE 1.0878 7.0301 2.8705 MSE 3.2122 7.5524 4.5143

Note: Bold values signific the best results.

TABLE 6 AIC values of the four ANN model topologies, the SVM-PSO model and the LSSVM model

AAN model

Model Trainbr, logsig Trainbr, tansig Trainlm, logsig Trainlm, tansig LSSVM model SVM-PSO model
AIC
1178.675
1058.967
435.523
88.721 1808.694 2578.408

Note: Bold values signific the best results.

3.3 | Methods for quantifying variable importance in ANNs

To determine the relative importance of the input variables, different statistical methods were applied.[ [48, 49]]In the present work, Garson's
method was used to split the hidden output connection weights into components associated with each input neuron using absolute connection
weight values. as well as the relevance factor r, which is in the range of
1 to +1 and is given by the following equation [22]:

P n   
Xk,i
Xk Y i
Y
i¼1
r k ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi ð37Þ
P n  2 P n  2
X k,i
Xk Yi
Y
i¼1 i
BENMOULOUD ET AL. 13 of 22

F I G U R E 2 Surface tension scatter plot computed as a function of the experimental surface tension for the global basis by the three ANN
model topologies, the SVM-PSO model, and the LSSVM model.

where Xk,i is the ith importing parameter, Yi is the ith exporting value, Xk is the average value of the kth input, Y is the average value of exporting
parameter, and n is the number of sets.
14 of 22 BENMOULOUD ET AL.

F I G U R E 3 Scatter plot of the surface tension calculate as a function of the experimental surface tension: (A) Global data set; (B) Training set;
(C) Test set by the best model of the ANN « trainbr, logsig ».

It can be seen from Figure 8 that for the established model, all input variables have almost the same Garson's method importance with the
exception of temperature. This can be explained by the low variability of the temperature data, that is, for some binary mixtures the experimental
data are available only at one or two temperatures. However, the sigma profile descriptor related to the H-bond acceptor character of the solvent,
Sσ2S, has the highest relative importance (20.26%) in Figure 9, the surface tension showed a straight dependency on the descriptors named:
S_σ2A, S_σ1S, and S_σ2S and an opposite dependency to the rest of the descriptors. The most relevant descriptors are the S_σ2S, and tempera-
ture with a relevance factor of +.021 and –.0235, respectively. Contrariety, S_σ1A and S_σ2A were the least relevant, with a relevance factor of
–.0052 and +.0024, respectively.

3.4 | Outlier diagnostics

The scatter plot showed a number of points far from the first bisector which can be attributed to the probable existence of outliers. This allows to
perform a diagnosis of potentially suspicious data according to the method described in the following two references [31, 50].
BENMOULOUD ET AL. 15 of 22

FIGURE 4 The relative error (%) of the surface tension by the ANN model “trainbr, logsig”.

FIGURE 5 Distribution diagrams of RE (%) of the ANN model “trainbr, logsig”.

F I G U R E 6 The interpolation and extrapolation calculations for the composition for the isotherms of the variation of surface tension as a
function of the composition of the mixture [C2MIm][BTI]-Dimethyl sulfoxide for the “trainbr, logsig” ANN model.

Therefore, it is essentially necessary to find rigorous methods to detect outliers in order to remove inaccurate experimental data and improve
model accuracy. In this study, an outlier detection method based on the Williams plot was used. This graph explains the relationship between the
Hat indices represented by Equation (38) and the residuals defined as differences between the experimental data and the corresponding esti-
mated values (R):
16 of 22 BENMOULOUD ET AL.

F I G U R E 7 The interpolation and extrapolation calculations for the temperature for the isotherms of the variation of surface tension as a
function of the composition of the mixture [C2MIm][BTI]-Dimethyl sulfoxide for the “trainbr, logsig” ANN model.

FIGURE 8 The relative importance of the variables by the Garson method of the ANN model “trainbr, logsig”.

FIGURE 9 The results of the relevance factor performed on the ANN model « trainbr, logsig ».

 
1
H ¼ X Xt X Xt ð38Þ

where X refers to the m  n matrix (m and n represent the number of samples and the parameters [input variables] of the model, respectively).
The values of Hat are obtained from the main diagonal of the matrix of H.
BENMOULOUD ET AL. 17 of 22

Figure 7 shows William's plot based on the results of the ANN model. In this chart, the critical leverage (H*) (threshold) is usually set to the
value given by the following equation:

3ðn þ 1Þ
H ¼ ð39Þ
m

FIGURE 10 Diagnosis of potentially suspicious data and domain of model applicability.

TABLE 7 Comparison between this work and the work of Huang et al [6] in terms of AARD%

This work Huang et al [6]. model

Systems {Ionique Number of Average I Average I Average

liquids + solvents} points %ΔσI I%ΔσImax I%ΔσSMI %ΔσImax I%ΔσΑΝΝI I%ΔσImax
[C4MIm][BTI] + 1-butanol 13 2.13690 14.2458 2.39918 4.83586 2.99069 5.56928
[C4MIm][BTI] + 1-propanol 11 .94144 2.64873 1.35199 3.04797 1.23155 2.18982
[C2MIm][BTI] + acetonitrile 45 .58221 2.03760 1.10309 3.61209 .35043 1.63301
[C4MIm][BTI] + acetonitrile 45 .28893 1.22930 1.32803 4.15549 .4058 .99813
[C2MIm][BTI] + Dimethyl 45 .48975 2.16922 1.59518 4.79112 .13389 .49756
sulfoxide
[C4MIm][BTI] + Dimethyl 50 .90014 19.16957 2.29169 8.29236 .15671 .53247
sulfoxide
[C2MIm][C8SO4] + ethanol 18 1.01942 2.10372 1.84492 4.20381 .44156 1.42752
[C4MIm][BF4] + ethanol 12 .86002 2.50953 .70218 1.78356 1.0015 6.5699
[C6MIm][BF4] + ethanol 10 3.20687 2.82941 1.14056 1.50424 .74439 3.40772
[C8MIm][BF4] + ethanol 10 .76441 2.28718 .82575 1.84567 1.17232 2.60078
[C1MIm][MeSO4] + methanol 9 .89797 19.1695 1.04938 9.30710 .56188 3.46873
[C2MIm][MeSO4] + methanol 7 .69485 1.75784 .61541 2.47781 .67376 3.26845
[C2MIm][BTI] + tetrahydrofuran 44 .93363 3.58668 .46162 1.46741 .46749 1.4991
[C4MIm][BTI] + tetrahydrofuran 40 1.22362 4.48224 .58329 2.60047 .63473 11.623
[C1MIm][MeSO4] + water 10 .86816 3.80506 .95716 2.41303 .33636 1.47093
[C2MIm][BF4] + water 9 .63134 1.20231 .18729 1.07727 .19068 .50178
[C2MIm][EtSO4] + water 12 .57436 2.21932 1.59173 3.93446 .60338 1.77915
[C2MIm][MeSO3] + water 27 1.73691 3.70604 3.89171 6.62818 .33508 1.02659
[C4MIm][BF4] + water 15 .73874 1.46092 .55320 1.96164 .4154 1.07962
[C4Py][NO3] + water 15 1.04161 6.56224 3.71712 9.3071 1.6599 11.623
[C6MIm][BF4] + water 8 .33916 .57460 .19374 .38724 .20127 .43478
[P666(14)][BTI] + water 6 .16640 .34975 1.41799 2.47305 .11722 .22418
[P666(14)][Dca] + water 6 .16995 .38829 1.24882 2.36076 .33602 .54937
18 of 22 BENMOULOUD ET AL.

where m is the number of samples, and n is the number of input variables of the model.
The normalized residuals are calculated from the data of the experimental surface tension and that calculated by the model.

 exp 
σ i
σ cal
ðR_NormÞi ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
i
ffi i ¼ 1, …m ð40Þ
Varðσ exp
σ cal Þ

The normalized residual of three is normally considered a “threshold” value for accepting data points within a range of ±3 SDs from the
mean (to cover 99% of the normally distributed data). If the majority of the data points fall within the ranges of 0 ≤ Hat ≤ H* and

3 ≤ R_Norm ≤ 3, this indicates that the model development and its predictions are performed in the applicability domain, leading to a statis-
tically valid model. Thus, we can affirm that there are “Good High Leverage” points in the domain of 0 ≤ Hat ≤ H* and
3 ≤ R_Norm ≤ 3. The
points located in the range of R_Norm<
3 or R_Norm>3 (whether larger or smaller than the H* value) are referred to as model outliers or
“Bad High Leverage” points.
Figure 10 shows that the majority of data points are lying inside the applicability domains, except for the existence of 2.82% of suspected
data. This confirms the observations made during the interpretations of the scatter plots and the relative error plots.

TABLE 8 Comparison between this work and the work of with the Cardona and Valderrama [12] in terms of AARD%

This work « trainbr, logsig » Cardona and Valderrama [12] model

Systems {ionic liquid + solvent} N Average I%ΔσI I%ΔσImax Average I%ΔσI I%ΔσImax
[C2MIm] [AC] + 1-propanol 77 .727311 3.279583 1.73871 22.48719
[C8MIm] [BTI] + 1-propanol 33 .513259 2.078608 6.64114 14.11421
[PYR-4.1] [BTI] + acetonitrile 27 .805194 2.216206 1.708758 4.497877
[C2MIm] [AC] + ethanol 77 .419793 2.068807 7.379528 20.68375
[C2MIm][DEP] + ethanol 11 1.369486 4.561168 11.7884 22.08567
[C2MIm][EtSO4] + ethanol 11 .994978 2.051358 13.25353 22.58346
[C4MIm][BTI] + ethanol 36 .38637 .784236 6.433625 18.33271
[C4MIm][DMP] + ethanol 11 1.343429 2.666525 14.05437 26.31037
[C6MIm] [EtSO4] + ethanol 13 1.091658 2.951107 3.927588 8.77661
[C8MIm] [BTI] + ethanol 39 .44928 1.518148 5.317242 15.36972
[C1MIm][DMP] + methanol 11 .562193 1.096063 6.937792 11.1255
[C2MIm] [AC] + methanol 55 .728288 3.611441 2.137411 9.412672
[C2MIm][EtSO4] + methanol 11 .955441 2.273106 5.599972 10.49703
[C2MIm][MeSO4] + methanol 14 .519645 2.936068 6.746018 15.85075
[C4MIm][BTI] + methanol 24 .194314 .717207 2.579008 6.910141
[C4MIm][DMP] + methanol 11 .63623 1.737604 2.965154 5.431559
[C4Py] [BF4] + methanol 56 .438397 2.181546 7.621787 18.83593
[C8MIm] [BTI] + methanol 20 1.102813 4.23808 2.704397 6.313109
[C1MIm][DMP] + water 16 1.289812 2.932375 1.380303 71.90
[C2MIm][DEP] + water 16 2.236424 12.66962 4.54 18.54
[C2MIm][DMP] + water 11 1.186484 3.990389 4.665865 9.781574
[C4MIm] [AC] + water 49 .385264 1.20122 3.213681 9.603819
[C4MIm] [CL] + water 46 .882057 2.076755 2.79283 7.822873
[C4MIm] [PF6] + water 9 .416549 1.264713 3.662293 7.550244
[C4MIm][DBP] + water 11 .941997 4.423875 5.830528 71.90
[C4Py] [BF4] + water 72 .694389 3.243486 8.413314 24.03831
[C6MIm] [AC] + water 49 .345018 1.013274 3.203462 10.76396
[C6MIm] [CL] + water 10 2.258253 5.479549 17.48413 26.5232
[C6MIm][BF4] + water 8 .339157 .574605 6.808245 10.50152
BENMOULOUD ET AL. 19 of 22

3.5 | Comparison with other models

In this section, a comparison study was carried out to evaluate the model's performance compared to previous papers. To the best knowledge of
the authors, there is no paper with a similar database to the dataset used in this paper. For that a comparison in terms of common systems was
adopted.

F I G U R E 1 1 Comparison between experimental (lines) and calculated surface tension values by our model (red dots) and the Valderrama
model [12].(blue squares) model for some systems.

F I G U R E 1 2 Comparison between experimental (lines) and calculated surface tension values by our model (red dots) and the Huang et al [6].
(green triangle) for his ANN model and (blue stars) for his SM model for some systems.
20 of 22 BENMOULOUD ET AL.

F I G U R E 1 3 Comparison between the predicted based and the experimental data versus the input data: (A) composition of IL,
(B) Temperature, (C) S_σ1C, (D) S_σ2C, (E) S_σ1A, (F) S_σ2A, (G) S_σ1S, (H) S_σ2S
BENMOULOUD ET AL. 21 of 22

A comparison, in term of the error (AARD%), between this work and the work of Huang et al [6]., Cardona and Valderrama.[12]is given in
Tables 7 and 8, respectively and the full comparison is included in the supplementary information file. Figures 11 and 12 represent another com-
parison between the experimental and calculated surface tension of some systems between the main model of this work and the previous men-
tioned papers.
Figure 13 illustrates a comparison between the calculated versus the experimental values of the ANN model (trainbr, logsig) according to the
inputs.

4 | C O N CL U S I O N

A prediction model based on an ANN has been successfully developed to predict the surface tension of binary mixtures. A total of 1623 experi-
mental data points of 62 binary mixtures were collected from various literature resources for use in the ANN model as training, validation and test
data points. The best architecture of the feed-forward network, obtained by a constructive approach, consisted of 26 neurones in the hidden
layers and was trained by the trainbr algorithm when the Logsigmoid (logsig) activation function in the hidden layer was applied. The following
conclusions can be drawn:

1. Significant results were obtained with the proposed ANN method. This fact is supported by the acceptable statistical quality confirmed by var-
ious parameters and the low errors of the ANN model results. The overall % AARD obtained and the MSE of .8466% and .4952%, respectively,
showed a very good capability and feasibility of using the ANN model for the prediction of surface tension of binary mixtures.
2. The established model is of great practical importance, it allows not only to accurately predict the surface tension of binary mixtures, including
ILs, but also promotes this method for other physico-chemical properties of IL mixtures in future studies to overcome limitations in the devel-
opment of industries and technologies based on the IL.

AUTHOR CONTRIBUTIONS
Widad Benmouloud: Methodology; validation; visualization; writing – original draft; writing – review and editing. Cherif Si-Moussa: Supervision;
writing – original draft; writing – review and editing. Othmane Benkortbi: Supervision; writing – original draft; writing – review and editing.

ACKNOWLEDGMENTS
The authors gratefully acknowledge the financial support of the Algerian Ministry of Higher Education and Scientific Research (PRFU Project
A16N01UN260120220003) and the University Yahia Fares of Medea.

DATA AVAI LAB ILITY S TATEMENT

The supplementary data to this article can be found online at (supporting file.pdf).

ORCID
Widad Benmouloud https://orcid.org/0000-0002-9456-8494
Othmane Benkortbi https://orcid.org/0000-0002-1965-7171

RE FE R ENC E S
[1] Q. Zhang, S. Feng, X. Zhang, Y. Wei, J. Mol. Liq. 2021, 328, 115373.
[2] K. J. Wu, C. X. Zhao, C. H. He, Fluid Phase Equilib. 2012, 328, 42.
[3] S. Atashrouz, H. Mirshekar, A. Hemmati-Sarapardeh, M. K. Moraveji, B. Nasernejad, Korean J. Chem. Eng. 2017, 34, 425.
[4] S. Atashrouz, M. Mozaffarian, G. Pazuki, Ind. Eng. Chem. Res. 2015, 54, 8600.
[5] A. Shojaeian, M. Asadizadeh, J. Mol. Liq. 2020, 298, 111976.
[6] Y. Huang, X. Zhang, Y. Zhao, S. Zeng, H. Dong, S. Zhang, Phys. Chem. Chem. Phys. 2015, 17, 26918.
[7] A. Shojaeian, Thermochim. Acta 2019, 673, 119.
[8] R. Sedev, Curr. Opin. Colloid Interface Sci. 2011, 16, 310.
[9] M. Tariq, M. G. Freire, B. Saramago, J. A. P. Coutinho, J. N. C. Lopes, L. P. N. Rebelo, Chem. Soc. Rev. 2012, 41, 829.
[10] M. Hashemkhani, R. Soleimani, H. Fazeli, M. Lee, A. Bahadori, M. Tavalaeian, J. Mol. Liq. 2015, 211, 534.
[11] R. Soleimani, A. H. Saeedi Dehaghani, N. A. Shoushtari, P. Yaghoubi, A. Bahadori, Korean J. Chem. Eng. 2018, 35, 1556.
[12] L. F. Cardona, J. O. Valderrama, Ionics (Kiel) 2020, 26, 6095.
[13] Y. Benguerba, I. M. Alnashef, A. Erto, M. Balsamo, B. Ernst, J. Mol. Struct. 2019, 1184, 357.
[14] Y. Zhao, Y. Huang, X. Zhang, S. Zhang, Phys. Chem. Chem. Phys. 2015, 17, 3761.
[15] F. Eckert, A. Klamt, AIChE J. 2002, 48, 369.
[16] T. Lemaoui, N. E. H. Hammoudi, I. M. Alnashef, M. Balsamo, A. Erto, B. Ernst, Y. Benguerba, J. Mol. Liq. 2020, 309, 113165.
[17] S. Gambhir, S. Kumar, Y. Kumar, New Horiz. Transl. Med. 2017, 4, 1.
22 of 22 BENMOULOUD ET AL.

[18] M. Geethanjali, S. M. Raja Slochanal, R. Bhavani, Neurocomputing 2008, 71, 904.

[19] M. A. Ahmadi, Z. Chen, Petroleum 2019, 5, 271.
[20] M. Ahmadi, Z. Chen, J. Pet. Explor. Prod. Technol. 2020, 10, 2873.
[21] H. Benimam, C. S. Moussa, M. Hentabli, S. Hanini, M. Laidi, J. Chem. Eng. Data 2020, 65, 3161.
[22] I. Euldji, C. Si-Moussa, M. Hamadache, O. Benkortbi, Mol. Inform. 2022, 2200026, 1.
[23] L. T. Le, H. Nguyen, J. Dou, J. Zhou, Appl. Sci. 2019, 9, 2630.
[24] T. Lemaoui, A. S. Darwish, N. E. H. Hammoudi, F. Abu Hatab, A. Attoui, I. M. Alnashef, Y. Benguerba, Ind. Eng. Chem. Res. 2020, 59, 13343.
[25] J. S. Torrecilla, J. Palomar, J. Lemus, F. Rodríguez, Green Chem. 2010, 12, 123.
[26] M. Diedenhofen, A. Klamt, Fluid Phase Equilib. 2010, 294, 31.
[27] E. Mullins, R. Oldland, Y. A. Liu, S. Wang, S. I. Sandler, C. C. Chen, M. Zwolak, K. C. Seavey, Ind. Eng. Chem. Res. 2006, 45, 4389.
[28] S. A. Kalogirou, Renew. Sustain. Energy Rev. 2000, 5, 373.
[29] H. Benimam, C. Si-Moussa, M. Laidi, S. Hanini, Neural Comput. Appl. 2020, 32, 8635.
[30] Z. Wan, Q. De Wang, J. Liang, Int. J. Quantum Chem. 2021, 121, 1.
[31] A. Baghban, A. H. Mohammadi, M. S. Taleghani, Int. J. Greenh. Gas Control 2017, 58, 19.
[32] S. Abdel-khalek, A. Alhag, M. Ragab, S. M. Abo-Dahab, A. Algarni, H. Ahmad, Int. J. Quantum Chem. 2021, 121, e26446.
[33] Y. Sang, H. Zhang, L. Zuo, 2008 IEEE Int. Conf. Cybern. Intell. Syst. CIS 2008, 2008, 290.
[34] J. A. Suykens, J. Vandewalle, Neural Process. Lett. 1999, 9, 293.
[35] M. N. Kardani, A. Baghban, J. Sasanipour, A. H. Mohammadi, S. Habibzadeh, J. Cleaner Prod. 2018, 203, 601.
[36] S. P. Mousavi, S. Atashrouz, M. Nait Amar, F. Hadavimoghaddam, M. R. Mohammadi, A. Hemmati-Sarapardeh, A. Mohaddespour, J. Mol. Liq. 2021,
342, 116961.
[37] N. H. Farhat, IEEE Expert. Syst. Appl. 1992, 7, 63.
[38] I. Mehraein, S. Riahi, J. Mol. Liq. 2017, 225, 521.
[39] Q. Song, G. Yan, G. Tang, F. Ansari, Mech. Syst. Signal Process. 2021, 146, 107019.
[40] Y. Zhao, X. Zhang, L. Deng, S. Zhang, Comput. Chem. Eng. 2016, 92, 37.
[41] J. Wang, H. Du, H. Liu, X. Yao, Z. Hu, B. Fan, Talanta 2007, 73, 147.
[42] H. Garg, Appl. Math. Comput. 2016, 274, 292.
[43] R. Steele, Understanding and Measuring the Shelf-Life of Food, Woodhead Publishing, 2004.
[44] R. Todeschini, D. Ballabio, F. Grisoni, J. Chem. Inf. Model. 1905, 2016, 56.
[45] O. Falyouna, O. Eljamal, I. Maamoun, A. Tahara, Y. Sugihara, J. Colloid Interface Sci. 2020, 571, 66.
[46] K. Paduszyn ski, Phys. Chem. Chem. Phys. 2017, 19, 11835.
[47] E. Mullins, Y. A. Liu, A. Ghaderi, S. D. Fast, Ind. Eng. Chem. Res. 2008, 47, 1707.
[48] J. D. Olden, M. K. Joy, R. G. Death, Ecol. Modell. 2004, 178, 389.
[49] M. Gevrey, I. Dimopoulos, S. Lek, Ecol. Modell. 2003, 160, 249.
[50] M. Hosseinzadeh, A. Hemmati-Sarapardeh, J. Mol. Liq. 2014, 200, 340.
[51] N. N. Ren, Y. H. Gong, Y. Z. Lu, H. Meng, C. X. Li, J. Chem. Eng. Data 2014, 59, 189.
[52] J. W. Russo, M. M. Hoffmann, J. Chem. Eng. Data 2011, 56, 3703.
[53] E. Rilo, J. Pico, S. García-Garabal, L. M. Varela, O. Cabeza, Fluid Phase Equilib. 2009, 285, 83.
[54] J. S. Torrecilla, T. Rafione, J. García, F. Rodrígue, J. Chem. Eng. Data 2008, 53, 923.
[55] J. Y. Wang, X. J. Zhang, Y. Q. Hu, G. Di Qi, L. Y. Liang, J. Chem. Thermodyn. 2012, 45, 43.
[56] U. Doman ska, A. Pobudkowska, M. Rogalski, J. Colloid Interface Sci. 2008, 322, 342.
[57] H. F. D. Almeida, J. A. Lopes-Da-Silva, M. G. Freire, J. A. P. Coutinho, J. Chem. Thermodyn. 2013, 57, 372.
[58] J. Y. Wang, F. Y. Zhao, Y. M. Liu, X. L. Wang, Y. Q. Hu, Fluid Phase Equilib. 2011, 305, 114.
[59] E. Rilo, M. Domínguez-Pérez, J. Vila, L. M. Varela, O. Cabeza, J. Chem. Thermodyn. 2012, 49, 165.
[60] M. Geppert-Rybczyn ska, J. K. Lehmann, A. Heintz, J. Chem. Eng. Data 2011, 56, 1443.
[61] M. Geppert-Rybczyn ska, J. K. Lehmann, J. Safarov, A. Heintz, J. Chem. Thermodyn. 2013, 62, 104.
[62] A. Wandschneider, J. K. Lehmann, A. Heintz, J. Chem. Eng. Data 2008, 53, 596.
[63] U. Doman ska, M. Zawadzki, A. Lewandrowska, J. Chem. Thermodyn. 2012, 48, 101.

SUPPORTING INFORMATION
Additional supporting information can be found online in the Supporting Information section at the end of this article.

How to cite this article: W. Benmouloud,

C. Si-Moussa, O. Benkortbi, Int. J. Quantum Chem. 2022, e27026. https://doi.org/10.1002/qua.27026