Machine‐learning‐assisted systematic screening of non‐ionic deep eutectic solvents for CO 2 capture

Abstract

Machine learning (ML) offers a promising approach for predicting CO 2 solubility, yet its application to the systematic screening of non‐ionic deep eutectic solvents (DESs) remains limited. Herein, a ML‐assisted framework integrating COSMO‐RS‐derived molecular descriptors, theoretical screening, and experimental validation is developed for CO 2 capture. Random forest and support vector regression (SVR) models are systematically compared, with the optimized SVR model exhibiting superior predictive performance. Strict leave‐one‐DES‐out validation further demonstrates its generalization capability. Virtual screening of 1051 DES candidates identifies three imidazole/ethylene glycol DESs with favorable absorption performance, which are experimentally validated with Henry's law constants of 8.53–10.02 MPa at 303.15 K. A residual‐correction transfer‐learning strategy is further employed to evaluate the transferability of the established model to SO 2 solubility prediction. Molecular‐level analysis reveals that electrostatic complementarity and vdW interactions govern CO 2 absorption. This work provides an efficient strategy for systematic non‐ionic DES screening and CO 2 capture.