Stacking ensemble learning and SHAP analysis for predicting CO2 solubility and guiding ionic liquid design

Abstract Excessive CO2 emissions constitute a primary driver of global climate change, creating an urgent need for efficient CO2 capture technologies. Ionic liquids (ILs) are considered highly promising absorbents. However, their vast chemical space poses significant challenges for experimental screening. In this study, Morgan fingerprint descriptors were combined with a stacking ensemble learning algorithm, and SHapley Additive exPlanations (SHAP) analysis was employed to develop an accurate and interpretable framework for predicting CO2 solubility. The stacking ensemble model employs the light gradient boosting machine and the backpropagation neural network as base learners. The dung beetle optimizer and the grey wolf optimizer were used to determine optimal hyperparameter combinations for these learners, while linear regression was adopted as the meta-learner to produce the final CO2 solubility prediction. The results demonstrate that the stacking ensemble algorithm achieves the highest predictive accuracy, attaining a test-set coefficient of determination (R2) of 0.9888, a mean absolute error of 0.0149 and a mean squared error of 0.0006. SHAP analysis further reveals that the [Tf2N] (bis(trifluoromethylsulfonyl)amide) anion exerts the most significant influence, followed by the alkyl chain length of imidazolium cations. Accordingly, combining the [Tf2N] anion with imidazolium cations bearing longer alkyl chains represents a promising strategy for synthesizing ILs with high CO2 absorption efficiency.