Advanced machine learning approach using Kolmogorov-Arnold Networks (KANs) and comparison with RSM for optimization of biodiesel yield from palm based waste cooking oil

The increasing environmental concerns associated with fossil fuels have intensified the demand for sustainable and renewable energy sources such as biodiesel. In this study, biodiesel production from palm-based waste cooking oil was optimized using Response Surface Methodology (RSM) and a Kolmogorov-Arnold Networks (KANs) model. A Central Composite Design (CCD) was employed to evaluate the effects of reaction temperature (55-70 °C), reaction time (60-120 min), agitation speed (500-700 rpm), catalyst concentration (1-5 wt%), and methanol-to-oil ratio (6:1-12:1) on biodiesel yield. The maximum experimental biodiesel yield of 99.7% was obtained under optimized reaction conditions, demonstrating the effectiveness of the transesterification process. The RSM model exhibited coefficient of determination (R 2 = 0.9597) and identified the optimal conditions as 55 °C reaction temperature, 120 min reaction time, 500 rpm agitation speed, 1 wt% catalyst loading, and 12:1 methanol-to-oil ratio. To further enhance prediction accuracy, a KAN model was developed and evaluated. The KAN model achieved an overall coefficient of determination of R 2 = 0.9576, with superior training performance (R 2 = 0.9893). The model exhibited low prediction error (RMSEP = 5.6664), with most relative errors within ±10% indicating strong capability in capturing nonlinear relationships. Comparative analysis revealed that while RSM provides reliable and interpretable optimization, the KAN model offers enhanced flexibility and improved prediction of complex nonlinear interactions. Therefore, KAN presents a promising approach for advanced optimization of biodiesel production processes. This study demonstrates the potential of KAN as a next-generation modeling tool for complex biofuel production systems.