Machine learning-based predictive modeling and optimization: Artificial neural network-genetic algorithm vs. response surface methodology for black soldier fly (Hermetia illucens) farm waste fermentation

Recognizing the complexity of non-linear and interdependent biological processes, this study compared the predictive performance of artificial neural network (ANN) models with response surface methodology regression based (RB) models. The research focused on the biological transformation of black soldier fly (Hermetia illucens) farm waste into chitin, facilitated by Lactobacillus paracasei. Key parameters of time (1–7 days), temperature (30–40 °C), substrate concentration (7.5–20 wt%), and inoculum concentration (5–15 v/v%), were evaluated for their impact on demineralization and deproteinization subprocesses and subsequently optimized. It was determined that the ANN models outperformed RB models, with R² values of 0.950 and 0.959 for DP% and DD%, compared to 0.677 and 0.720 for RB models. While both models, optimized using a multi-objective genetic algorithm (MOGA) and a desirability function respectively, produced comparable optimal results, differences emerged in process variable analysis. Main effects plots (RB) and one way partial dependence plots (ANN) revealed conflicting parameter influences, highlighting the limitations of regression models in complex systems. This study highlights the superiority of ANN-MOGA in addressing biological complexity and recommends its use especially if RB models show suboptimal predictive capabilities.