Comparative machine learning strategies for improving antioxidant properties and aroma quality in fermented mung bean milkby Lactobacillus plantarum PC4

This study compares least squares support vector machine (LSSVM) and artificial neural network (ANN) models, integrated with the NSGA-II algorithm, to optimize the fermentation of mung bean milk by Lactobacillus plantarum PC4. Given its superior predictive accuracy and generalization, LSSVM was selected as the final model for multi-objective optimization and experimental validation. LSSVM consistently outperformed ANN in predictive accuracy and generalization, particularly under data-scarce conditions, yielding R² values exceeding 0.97 across all responses. Optimal fermentation conditions predicted by LSSVM (6.0 h, 37.0 °C, 1.99 % inoculum) were experimentally validated, showing minimal error (<5 %) across most parameters. GC–MS analysis confirmed that the LSSVM-NSGA-II optimized fermentation conditions effectively suppressed off-flavor aldehydes (e.g., hexanal, nonanal) while promoting the formation of favorable volatiles, including 1-hexanol, acetoin, esters, and aromatic compounds. These targeted improvements in antioxidant and aroma profiles underscore the efficacy of this data-driven approach in enhancing both the functional and sensory attributes of plant-based fermented beverages.