QSPR strategy to model and analyze surface tension of binary-liquid mixtures: A large-data-set case

• Novel QSPR strategy is introduced for estimation of surface tension of binary-liquid mixtures via a MLP network and GA–PLS.
• A data set of 8200 points is employed for modeling.
• The GAvs method is applied to select the best descriptors based on RQK statistical criteria.
• k-fold cross-validation algorithm is used for training the suggested MLP networks.

This study introduces a method of quantitative structure–property relationship (QSPR) to evaluate/predict surface tension (ST) of binary-liquid mixtures using a large-data-set over the (254.85–443.25 K) temperature range, and the whole concentration span. To construct and establish an accurate, comprehensive, reliable, and predictive model, the subsequent procedure is followed. Firstly, a set of 8200 data points is collected. The data set belongs to experimental surface tension values attained for 469 binary-liquid mixtures (relevant to 156 chemical components) at different temperatures and various concentrations. Secondly, to select the optimum and the most effective descriptors, a repeated search technique based on genetic algorithm (GA) and partial least squares (PLS) is applied. Thirdly, contrasting with the ST of the binary mixtures, the best construction of multi-layer perceptron neural network (MLPNN) is employed for examination of the nonlinear behavior of the selected descriptors other than temperature and mole fraction of the light components (empirical descriptors). Comparing the results to the experimental data confirms the appropriate and well-meaning accuracy of the ultimate model (4.89% mean relative error). This would be of significant importance as the proposed model, which directly uses the molecular structures without requiring any experiment, has a proper precision while the accuracy of empirical or thermodynamic-based models is limited and depends on specific model factors and/or other parameters obtained from experimental data.