Ultrasonic-Assisted Biodiesel Production from Waste Animal fat: Process Optimization and Interpretable Machine Learning

Abstract

Waste animal fat is an attractive low-cost feedstock for sustainable biodiesel production. However, its efficient conversion requires appropriate process optimization and reliable prediction of operating conditions. This study presents an integrated framework combining ultrasonic-assisted transesterification, response surface methodology (RSM), and interpretable machine learning (ML) to optimize biodiesel production and identify the relative importance of process variables. A five-factor, three-level Box-Behnken design (BBD) containing 46 experimental runs was employed to investigate the effects of ultrasonic power, frequency, methanol-to-oil molar ratio, reaction temperature, and reaction time. Three tree-based ML models, including Random Forest (RF), Gradient Boosting (GB), and Extreme Gradient Boosting (XGBoost), were developed and evaluated using the experimental dataset. Among the evaluated models, XGBoost attained the best predictive performance for the test dataset, with an R² of 0.8097 and an RMSE of 0.5467. SHAP analysis, permutation importance, and partial dependence plots consistently identified reaction time and molar ratio as the most significant parameters governing biodiesel yield. RSM optimization predicted optimum operating conditions of 134.74 W ultrasonic power, 26.17 kHz frequency, 6.23:1 molar ratio, 35.98 °C reaction temperature, and 8.17 min reaction time, resulting in a predicted biodiesel yield of 95.53%. Experimental validation showed less than 2% deviation from the predicted value. Physicochemical characterization, FTIR, and GC-MS analyses confirmed successful biodiesel production with fuel properties meeting ASTM D6751 requirements.