Response surface methodology–machine learning optimization of Ziziphus mauritiana biodiesel over UiO‐66/SO ₃

Abstract

Biodiesel derived from non‐edible feedstocks via heterogeneous catalysis offers a sustainable pathway to displace petroleum diesel in compression ignition engines. This study reports, for the first time, the application of a sulfonated zirconium metal–organic framework (UiO‐66/SO ₃ H) catalyst to transesterify Ziziphus mauritiana seed oil, optimized through an integrated Response Surface Methodology (RSM)–machine learning (ML)–SHAP (SHapley Additive exPlanations) framework and validated through multi‐blend CI engine testing. A Box–Behnken Design (BBD) (30 runs, 4 factors) identified optimal conditions (10.5:1 methanol:oil ratio, 3 wt% catalyst, 60°C, 120 min) yielding 96.2% biodiesel ( R ²  = 0.9866, in‐sample MAPE = 0.41%). Among six machine learning algorithms cross‐validated via Leave‐One‐Out Cross‐Validation, XGBoost performed best ( R ²  = 0.9344, MAPE = 2.21%), while SHAP analysis confirmed that the machine‐learning‐ and ANOVA‐derived factor importance rankings converge exactly (methanol ratio > catalyst loading > temperature > time). Engine evaluation of B10, B20, and B30 blends in a Kirloskar TV1 single‐cylinder CI engine (5.2 kW, 1500 rpm) showed carbon monoxide (CO), hydrocarbon (HC), and smoke reductions of up to 25.81%, 16.67%, and 16.41%, respectively, with a 9.32% (oxides of nitrogen) NO _x penalty for B30 at full load. Combined Kline–McClintock and Moffat uncertainty analysis yielded ±3.24%. These findings establish UiO‐66/SO ₃ H as an effective, reusable catalyst for a previously unexplored non‐edible feedstock and demonstrate a reproducible RSM–ML–SHAP validation framework applicable to future biodiesel optimization studies.