Machine learning assisted QSPR model for prediction of ionic liquid’s refractive index and viscosity: The effect of representations of ionic liquid and ensemble model development

Abstract

In this study, we used four ways to represent ionic liquids (ILs), namely, molecular fingerprint (MF), molecular descriptor (MD), the addition of MF (MF + MF) and the combination of MF and MD (MF_MD), to develop quantitative structure–property relationship (QSPR) models for predicting the refractive index and viscosity of ILs. Results showed that the predictive performance of QSPR models followed the order: MD < MF + MF < MF < MF_MD, indicating combining the chemical structure information and the physicochemical properties of ILs was beneficial to enhancing the predictive performance of the QSPR model. We also investigated the effect of the data splitting way on the predictive performance of the QSPR model, and the results showed that the group-based random splitting way was more reasonable than the random splitting way. The shapely additive explanation (SHAP) method was used to interpret MF_MD-based QSPR models. Results showed that different MDs play important role in prediction of refractive index and viscosity and the effects of conditions (temperature and/or pressure) were correctly identified. The QSPR model also correctly “learned” how MF affect the viscosity but wrongly “identified” how MF affect the refractive index. Finally, we developed the ensemble models by combining these single QSPR models to develop the final more accurate QSPR model. This study demonstrated that how to represent ILs plays important role in obtaining QSPR models with high predictive performance and developing the ensemble model was the possible efficient approach to further enhance the predictive performance of the QSPR model for ILs.