Accurate and Interpretable Prediction of CO2 Solubility in Physical Absorbents Based on Sequential Models

Abstract Methanol-based physical absorption processes for industrial decarbonization usually require harsh low-temperature conditions, highlighting the need for screening efficient absorbents under mild conditions. Machine learning has accelerated solvent screening; however, challenges in prediction accuracy and interpretability still remain. To this end, sequential models (RNN and GRU) were developed by treating the temperature and pressure as ordered inputs to capture the thermodynamic dependence of CO2 solubility. Compared with baseline models (multilayer perceptron (MLP) and XGBoost), the sequential models exhibit excellent performance in the high-CO2 solubility region, particularly the GRU model, which achieved a high R2 of 0.9866. To better understand the structure governing CO2 solubility, Shapley additive explanation (SHAP) analysis based on molecular descriptor (MD) and group contribution (GC) representations was conducted, which revealed negative contributions from hydroxyl groups and positive effects from carbonyl groups and ether bonds. Furthermore, we identified several promising absorbents (e.g., methyl acetate, dimethoxyethane, and acetylacetone) under relatively mild conditions, providing a reliable framework for absorbent screening.