Tri-objective co-optimization of waste-cooking-oil biodiesel using machine learning, NSGA-II, and life-cycle assessment

Turning waste cooking oil (WCO) into biodiesel reduces waste and fossil fuel dependence, yet performance and environmental impacts are typically assessed in isolation, obscuring critical trade-offs. We identify operating conditions that simultaneously maximize biodiesel yield, increase higher heating value (HHV), and minimize process energy per kilogram of fuel. Using 500 designed experiments varying seven factors ( methanol-to-oil ratio, catalyst loading, temperature, reaction time, mixing speed, free fatty acid content, and water content). We train multiple machine learning models and apply multi-objective optimization. XGBoost delivers the best predictive accuracy across all targets; SVR performs worst for HHV. A gate-to-gate life cycle assessment (per 1 kg biodiesel) quantifies environmental impacts. The optimizer yields 58 Pareto-optimal solutions; a balanced point achieves 85.9 wt% yield, 38.2 MJ kg −1 HHV, and 5.1 MJ kg −1 process energy. Compared to a conventional baseline, this solution reduces global warming potential, acidification, and eutrophication by 11.3% on average. Unlike sequential approaches that optimize performance first and assess sustainability afterward, our integrated ML–optimization–LCA framework unifies prediction, search, and environmental evaluation in a single workflow, explicitly revealing trade-offs among yield, fuel quality, and energy intensity. This enables producers to select operating points aligned with priorities – higher output, better fuel, or lower energy – while quantifying environmental co-benefits. Policymakers can leverage these insights to align biodiesel production with carbon reduction and efficiency targets, supporting cleaner, scalable deployment at plant scale. • ML–NSGA-II–LCA workflow co-optimizes yield, HHV, and process energy. • Balanced point: 85.9% yield, 38.2 MJ/kg HHV, 5.1 MJ/kg energy. • Average 11.30% impact reduction vs. baseline (GWP, AP, EP). • 58 Pareto-optimal settings quantify actionable trade-offs. • Transparent, reproducible operators and instrument specs provided.