Quantum Chemistry-Guided Machine Learning for Accelerated Design of CO2-Solubilizing Deep Eutectic Solvents

Deep eutectic solvents (DESs) offer promise for CO2 capture due to their tunability and low cost, yet their development is constrained by limited predictive accuracy for CO2 solubility. In this work, we present the first machine learning framework that integrates quantum chemical descriptors for solubility prediction. We compiled 2287 experimental measurements from 119 DESs over wide temperature (293.15–353.15 K) and pressure (26.3–7620 kPa) ranges. Through density functional theory (DFT) calculations, 48 electronic structure parameters for hydrogen bond donors (HBDs) and acceptors (HBAs) were extracted and combined with operational variables to build predictive models. Among various algorithms, CatBoost achieved superior performance (test R2 = 0.9921; RMSE = 0.1152), reducing residual deviation by 72% compared to conventional σ-profile methods. SHAP interpretability analysis revealed that CO2 solubility is primarily governed by the electrostatic potential distribution of HBAs and the electron delocalization of HBDs. Leveraging this model, we screened 2176 potential DES combinations and identified several high-performance candidates. This quantum chemistry-informed machine learning framework offers an interpretable and efficient computational tool to guide molecular design of DESs, significantly reducing experimental efforts and establishing a new paradigm for rational solvent development in carbon capture technologies.