Hybrid thermodynamics‐informed data‐driven models for predicting fluorinated gas solubility in deep eutectic solvents

Abstract To facilitate the efficient screening of deep eutectic solvents (DESs) for capturing fluorinated gases (F‐gases) as extremely severe greenhouse gases, this study developed thermodynamics‐informed machine learning (ML) models, based on physical descriptors, that is, distribution of surface shielding charge density ( σ‐profiles ), the molecular surface area and molecular volume, derived from the thermodynamic model COSMO‐RS to ultimately predict the solubility of F‐gases in DESs. More than 1000 experimental solubility data (6F‐gases and 42 DESs) were collected. The support vector regression (SVR) was ultimately selected as the optimal predictive ML model under the comprehensive assessments, and was successfully applied to predict the solubility of R‐32 in three DESs (not included in the original dataset), demonstrating a highly predictive accuracy (average absolute relative deviation (AARD) of 17.79%) comparable to physical model PC‐SAFT (AARD of 11.75%). This demonstrates that thermodynamics‐informed ML model has the strong generalization capability to screen DESs for F‐gases capture.