Prediction of Hydrogen Solubility in Ionic Liquids by Machine Learning Models Based on COSMO-RS Descriptors

Understanding hydrogen (H 2 ) solubility in ionic liquids (ILs) is crucial for promoting the performance of ILs in many important applications of green H 2 value-chains including low-energy H 2 purification, IL-based H 2 compressors, and solvents for biomass hydrogenation. However, experimental screening of desirable ILs is extremely challenging, and computational modeling together with intelligent algorithms can offer valuable support for ILs’ screening and design. Although thermodynamic equations that heavily rely on substance-specific parameters and reported intelligent algorithms achieve good fitting accuracy, they neglect the electrostatic microenvironment of the ILs themselves and fail to provide interpretable, physically meaningful descriptors for large-scale screening tasks. In this study, we employ the conductor-like screening model for real solvents to extract the cumulative surface charge density distribution (σ-profiles) of ionic fragments as key feature descriptors, establishing the relationship between the electrostatic microenvironment of ILs and H 2 dissolution behavior. Four machine learning (ML) models, including Gradient Boosting Regression (GBR), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Multilayer Perceptron (MLP), are developed to predict H 2 solubility in ILs. Among all models, the MLP model exhibits the best performance, achieving a coefficient of determination ( R 2 ) of 0.9928 and a root-mean-square error of 0.00629 on the test set. Moreover, its predictions on the external independent validation set are in good agreement with experimental values. Therefore, the MLP model is recommended for predicting H 2 solubility in ILs. All models are validated via 5-fold cross validation and Bayesian hyperparameter optimization to ensure robustness and reliability. Analysis of the input features highlights the influence of σ-profiles on H 2 dissolution behavior, conferring strong generalization capability for solubility prediction.