Data-driven modeling of ionic liquids extractive desulfurization integrating machine learning and quantum chemistry

Conventional hydrodesulfurization is ineffective against refractory thiophenic compounds, making ionic liquids (ILs) attractive for extractive desulfurization (EDS) despite their vast chemical space. Here, an end-to-end design framework was established using a database of 244 EDS records. Among five machine learning models, the multilayer perceptron showed superior external generalization for structurally distinct ILs. SHAP analysis and quantum chemical calculations revealed that cationic features and π-stacking interactions dominate desulfurization performance. Guided by these insights, IL structures were directionally optimized and experimentally validated. The two newly designed ILs showed strong agreement with model predictions, demonstrating the reliability of the proposed data-driven design strategy. This work provides an extensible framework for accelerating the discovery of efficient solvents.