Machine learning prediction and SHAP interpretation of N ₂ O solubility in ionic liquids

Abstract

The integration of machine learning (ML) into chemical research provides an efficient strategy to mitigate the high costs and extended timelines of conventional experiments. This study develops a quantitative structure–property relationship (QSPR) model using ML to predict N ₂ O solubility in ILs. Three algorithms – Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN) and Support Vector Regression (SVR) – were employed. Beyond temperature and pressure, molecular descriptors generated from SMILES structures via RDKit were innovatively incorporated as inputs, representing their first systematic application for this purpose. This approach significantly enhances the model’s capability to characterize the structural features of ionic liquids. Hyperparameters were optimized using the Optuna Bayesian framework, which was employed for tuning in this study. Compared to the grid search and random search methods used in prior work, this approach demonstrates clear advantages in both efficiency and accuracy, thereby enhancing overall model performance. All models performed effectively, with LSTM achieving superior test-set results ( R ²  = 0.9935, MSE = 0.0002). SHAP analysis of the optimal model identified pressure, the descriptor Cation_PEOE_VSA13 and temperature as the most influential features. Furthermore, two cation structures, [(ETO) ₂ IM] ⁺ and [(OH) ₂ IM] ⁺ , were identified as promising for N ₂ O absorption based on descriptor analysis.