Prediction of SO2 solubility in ionic liquids based on machine learning and analysis of SHAP

Abstract

Sulfur dioxide (SO₂) is a major atmospheric pollutant, and ionic liquids (ILs) have shown great potential for SO₂ capture due to their unique physicochemical properties. However, the complex and designable structures of ILs make it challenging to establish accurate structure–solubility relationships. In this study, a prediction model based on molecular descriptors and a Long Short-Term Memory (LSTM) neural network was developed to predict SO₂ solubility in ILs. A dataset of 381 experimental solubility data points covering 48 ILs was collected and augmented to 1905 points using a physics-rule-based method. Molecular descriptors for anions and cations were computed from SMILES strings using RDKit, and an intelligent feature selection method reduced the feature dimensionality from 89 to 50. The proposed Z-Score-LSTM model achieved excellent predictive performance, with a coefficient of determination (R²) of 0.9500, a root mean square error (RMSE) of 0.0521, and a mean absolute error (MAE) of 0.0252 on the test set. SHAP analysis revealed that temperature and cation partial charge distribution (Cation_PEOE_VSA13) are the most influential features, with anion polar surface area inhibiting solubility and cation shape indices promoting it. This work provides a high-precision, interpretable tool for the molecular design of novel SO₂-absorbing ionic liquids.