Machine Learning Prediction of Ionic Liquid Conductivity for Sustainable Screening

Ionic liquids (ILs) have attracted increasing interest as sustainable solvents and electrolytes because of their tunable structures and physicochemical properties. However, the accurate prediction of their conductivity remains challenging. In this study, a dataset containing 537 ILs and 6800 samples was constructed based on the NIST database. Two molecular representations, Self-Referencing Embedded Strings (SELFIES) and Simplified Molecular Input Line Entry System (SMILES), were employed, and three machine learning models were developed. All models showed good predictive performance, among which Extreme Gradient Boosting (XGBoost) achieved the highest accuracy. Under the SELFIES representation, the coefficient of determination values for the training, validation, and test sets reached 0.955, 0.948, and 0.925, respectively. SHapley Additive exPlanations (SHAP) interpretation indicated that temperature was the dominant factor affecting conductivity. Furthermore, toxicity evaluation based on predicted Log(EC50) values showed that most ILs exhibit low toxicity, with the majority classified as relatively harmless or practically harmless. These results demonstrate a feasible strategy for the combined prediction of performance and environmental impact, thereby guiding the screening of sustainable ILs.