Continuous reactor metamodel: Optimization of biodiesel reaction systems using machine learning techniques

This work introduces a machine learning-based systematic methodology for constructing and optimizing metamodels of the biodiesel transesterification process, leveraging an augmented reaction network to enhance the predictive understanding of both primary and intermediate reaction pathways. Linear regression-based surrogate models were adopted due to their simplicity and exceptional predictive performance (Q2>0.99), making them ideal for feature selection and capturing complex process behavior. The metamodels effectively represented nonlinear effects and interactions, enabling accurate prediction of the key performance indicators such as conversion, selectivity, and yield. Optimization studies were carried out across different oil feedstocks with widely varied feed compositions, revealing that soybean oil, olive oil, and beef tallow were most effective in enhancing performance metrics. In contrast, sunflower oil consistently contributed negatively, likely due to its less favorable transesterification behavior. The manipulated variables NaOH and methanol flowrates as well as reactor temperature converged to their upper operational limits, emphasizing their critical influence in boosting reaction rates and product yields. The optimized solutions were consistently found near or within the boundaries of the experimental training data, affirming the robustness and validity of the modeling approach. Overall, the proposed framework offers a reliable tool for guiding experimental efforts, optimizing biodiesel production conditions, and advancing process understanding through interpretable and data-efficient metamodels.