Predicting the Properties of IL-Solvent Electrolytes Using a Hybrid Support Vector Classification and Gradient Boosting Regression (SVC-GBR) Framework

Harnessing the vast design space of ionic liquid (IL) systems requires efficient, high-throughput computer-aided molecular design (CAMD) strategies. We present a hybrid classification-regression machine learning framework for predicting viscosity and conductivity of IL-solvent systems over wide temperature and composition ranges. Gradient boosting regression (GBR) models were trained using COSMO-RS sigma profiles and group contribution (GC) features, with GC features showing superior predictive performance. Additionally, SHapley Additive exPlanations analysis was implemented to provide insight into feature importance. Preclassifying property data sets into narrower ranges (three for viscosity, two for conductivity) enhanced accuracy, motivating the development of a hybrid support vector classification SVC-GBR framework. This framework achieved AARD (%) values of 19.2% for viscosity and 25.4% for conductivity on IL-solvent systems previously unseen by the model, representing improvements of 34.0% and 21.4% over baseline GBR models.