A critical methodological revisit on group-contribution based property prediction of ionic liquids with machine learning

Abstract

As group-contribution (GC) concept is convenient and pragmatic for predictive property modeling, exploring the limits of GC-based models by pairing different machine learning (ML) algorithms in a critical modeling procedure is of great interest. In this work, based on four temperature-dependent and two temperature-independent ionic liquid (IL) properties (viscosity, surface tension, refractive index, thermal conductivity, IPC-81 cytotoxicity, and thermal decomposition temperature), GC modeling is critically performed and analyzed by including six ML algorithms (multivariate linear regression, ridge regression, artificial neural network, support vector regression, random forest, and extreme gradient boosting). Following rigorous cross-validation, the significant effect of the point-based and IL-based dataset split schemes on modeling the four temperature-dependent properties is clearly demonstrated. The optimal algorithm for modeling each of the six IL properties is determined, wherein the effect of random dataset splits is also evaluated. Finally, the effect of molecular occurrences of group in GC modeling is proposed and exemplified.