• • The gradient boosting regression (GBR) model achieved a test-set R² of 0.94 and the lowest RMSE among six ML models, enabling accurate prediction of levulinic acid yield from cellulose conversion.
• • SHAP and partial dependence analyses revealed that water fraction, catalyst dosage, and reaction temperature are the dominant factors controlling LA yield, with synergistic effects that must be optimized jointly.
• • Integration of GBR with particle swarm optimization identified RuCl₃ as an efficient catalyst under high-temperature and short-reaction-time conditions, offering a data-driven route to catalyst selection.
• • The data-driven framework reduces reliance on exhaustive experimental trials, accelerating the discovery of optimal reaction conditions for industrial LA production.