Machine learning assisted Structure-based models for predicting electrical conductivity of ionic liquids

Abstract

Ionic liquids (ILs) have attracted a great deal of attention as vital compounds that are widely used in various industries, such as the chemical and petroleum industries. Therefore, determining the properties of ionic liquids, such as electrical conductivity (EC), is an important issue. In this study, four different machine learning approaches including K-nearest neighbours (KNN), extreme gradient boosting (XGB), adaptive boosting-decision tree (Ada-DT), and adaptive boosting-support vector regression (Ada-SVR) were established to predict EC of ILs based on two categories namely: (I) smart models based on ILs chemical structure and temperature, and (II) smart models based on ILs thermodynamic properties and temperature. For this purpose, 2625 laboratory data points corresponding to 225 ILs have been used. The results showed that based on Model (I), KNN model with root mean square error (RMSE) and coefficient of determination (R2) values of 0.1976 and 0.9935, respectively, is the best model in predicting EC. Also, based on Model (II), the XGB model with RMSE and R2 values of 0.1447 and 0.9965, respectively, is the most powerful and accurate model for estimating electrical conductivity. Sensitivity analysis revealed that temperature and molecular weight have a direct and inverse effect on increasing the electrical conductivity of ILs, respectively. The investigation of the effects of different chemical bonds (in the chemical structure of ionic liquids) on ILs EC showed that –CH2- and > N- substructures have the most negative and positive effect on EC, respectively. Also, the proposed smart paradigms were capable of accurately predicting the effect cation alkyl chain length on electrical conductivity. Based on the leverage technique assessment more than 95% of total data points are in a valid region which indicates that intelligent models are reliable. Consequently, the introduction of intelligent models based on the chemical structure of ILs, the much broader data bank, and the higher accuracy of the proposed models are among the innovations and advantages of this study compared to the previous researches. Findings of this study highlights that chemical structure-intelligence based approaches provide robust, reliable and accurate way to predict EC of ILs.