A hybrid XGB-ANN model integrated with Lion Optimization Algorithm for improved prediction accuracy and lactic acid yield optimization

Abstract

Machine learning (ML) has been widely applied in industrial fields in recent years. However, ML faces challenges in fermentation production due to complex features and limited data. To overcome the limitations of single models and improve the prediction accuracy and stability of lactic acid fermentation, we combined six classical ML models with Artificial Neural Networks (ANN) separately in a serial manner to construct the Classical-ANN (C-ANN) hybrid models. Meanwhile, implemented the Lion Optimization Algorithm (LOA) to systematically optimize critical fermentation parameters and conducted experimental validation. While single models showed moderate performance (Extreme Gradient Boosting (XGB) R²= 0.9197; ANN R²= 0.9295), C-ANN hybrids demonstrated superior accuracy across all metrics. The XGB-ANN model achieved exceptional performance (R²=0.9850), benefiting from XGB's robust feature selection and ANN's high-dimensional processing capabilities. SHAP analysis revealed fermentation time as the most critical yield determinant, followed by total volume and pH. Lion Optimization Algorithm parameters enabled a predicted yield of 81.21 %, with experimental validation showing only 4.72 % deviation from predictions. This study not only provides valuable assistance for optimization of fermentation production conditions in the short term but also offers an innovative and effective modeling approach for future research in this field.