Prediction of CO2 Solubility in Deep Eutectic Solvents Using Machine Learning with Compact σ-Profile-Derived Descriptors

Abstract

Deep eutectic solvents (DESs) are promising for CO2 capture, yet rapid screening remains challenging because of their chemical diversity. This study develops an interpretable machine learning framework integrating experimental conditions with compact molecular descriptors derived from COSMO solvation calculations and the resulting σ-profiles. A data set of 2,398 CO2 solubility measurements across 114 unique DESs was used to develop tree-based machine learning models. XGBoost achieved the highest testing accuracy (R2 = 0.990, RMSE = 0.070, MAE = 0.043, AARD = 10.05%) using only nine input features. Benchmark comparisons showed performance comparable to or better than models using full σ-profiles or larger feature sets. SHAP analysis identified pressure as the dominant predictor and highlighted contributions from molecular geometry and hydrogen-bonding-related descriptors. The proposed framework therefore provides an accurate, efficient, and interpretable approach for screening DESs within chemically related design spaces and supports future data-driven solvent design.