Predictive Design of Deep Eutectic Solvents for Lignin Solubility: A Machine Learning-Calibrated COSMO-RS Framework

Deep eutectic solvents (DES) are promising green media for lignin dissolution due to their tunable hydrogen-bond networks. However, the vast compositional space of DES makes the rational design of efficient lignin solvents challenging. To address this, this study developed a hybrid COSMO-RS and machine learning (ML) framework for accurate lignin solubility prediction and rapid solvent screening. By shifting the focus from direct numerical fitting to residual learning, prediction accuracy can be substantially improved, which effectively provides a high-order correction for many–body interaction energy contributions within the thermodynamic cycles. Among three prediction targets tested, the deviation-based model achieved the highest accuracy. The multilayer perceptron (MLP) yielded the best performance, increasing test R2 from 0.9182 to 0.9941 and reducing the average absolute relative deviation (AARD %) to below 8.94%. Interpretability results revealed that hydrogen bond, the solubility calculated by COSMO-RS, and temperature are the dominant factors governing solubility. This work presents a hybrid modeling framework that integrates physical insights with data-driven calibration, enabling precise and interpretable predictions for complex lignin-solvent systems.