Toward Sustainable Biorefining: Interpretable Machine Learning Unravels Performance Degradation Mechanisms during Deep Eutectic Solvent Recycling in Lignocellulosic Biomass Pretreatment

Abstract The long-term recyclability of deep eutectic solvent (DES) is central to its sustainability in lignocellulosic biomass pretreatment, yet quantitative rules governing recycling-induced performance loss remain limited. Here, we constructed a literature-derived, cycle-resolved dataset for DES-based biomass pretreatment and introduced a delignification retention metric to normalize solvent performance over repeated use. Three machine-learning models, including XGBoost, random forest, and multilayer perceptron, were evaluated using molecular descriptors, biomass composition, DES formulation variables, and operating parameters. XGBoost delivered the best predictive performance. SHAP analysis reveals that solvent recycling impairs solvent performance to varying degrees, manifesting as reduced delignification efficiency and elevated solid recovery. Further, SHAP analysis identified solvent recycling times, hydrogen bond donor (HBD) logP, DES molar ratio, temperature, and liquid-to-solid ratio as key variables governing pretreatment performance and retention. The model suggests that hydrophilic HBD, moderate temperatures, and sufficiently high liquid-to-solid ratios favor delignification retention, which corresponds to more stable pretreatment performance during cyclic utilization, whereas high molar ratio DES may provide strong initial delignification but poorer long-cycle stability in some cases. These results provide an interpretable data-driven framework for balancing initial pretreatment efficiency, solvent recyclability, and process sustainability in closed-loop DES biorefining.