Machine Learning-Augmented COSMO-SAC Model for Accurate Screening of Green Deep Eutectic Solvents in Sustainable CO ₂ Capture

The discovery of efficient and environmentally benign deep eutectic solvents (DESs) is crucial for advancing sustainable carbon capture, utilization, and storage (CCUS). However, the experimental screening of vast DES candidates is prohibitively costly and time-consuming, while existing predictive models suffer from poor accuracy or fail to extrapolate to unseen molecular structures. To bridge this gap, this study introduces a hybrid physics-machine learning framework that synergizes a first-principles-based COSMO-SAC thermodynamic model with a Δ-learning correction module for the high-accuracy prediction of CO2 solubility in DESs. By optimizing the COSMO-SAC physical parameters specifically for DES systems and employing a symbolic regression-based Δ-KAN model, exceptional predictive accuracy (R2 = 0.915, AARD = 4.3%) is achieved on a rigorously split data set containing unseen DES types─a 45% reduction in relative error against the state-of-the-art pure machine learning model. This high extrapolation capability is critical for the de novo design of green solvents. Furthermore, a web-based platform (COSAC-X) is deployed to facilitate the rational screening of novel DESs and other solvent systems (e.g., organic solvents and ionic liquids) for carbon capture and other phase equilibrium applications (e.g., extraction and crystallization). This work provides a reliable and transferable toolbox that significantly accelerates the development of energy-efficient and sustainable CCUS processes, moving from serendipitous discovery to rational design.