Machine learning-based prediction of microbial growth and acidification in yogurt fermentation at industrial temperatures

Accurate prediction of microbial growth and acidification is essential for optimizing yogurt fermentation processes. This study presents a machine learning-based framework that integrates Long Short-Term Memory (LSTM) neural networks with classical primary growth models for microbial prediction, and Support Vector Regression (SVR) for modeling pH acidification dynamics. The framework was applied to Streptococcus thermophilus and Lactobacillus delbrueckii subsp. bulgaricus fermentations in reconstituted skim milk at 37 °C, 41 °C, and 45 °C. The hybrid LSTM model showed superior accuracy at 37 °C and 41 °C in capturing strain-specific, temperature-dependent growth patterns, while the Gompertz model remained most effective at 45 °C. For acidification prediction, SVR outperformed Gompertz, Weibull, and ensemble models under all conditions. Validation using ten biological replicates revealed RMSE reductions of up to 86 % compared to classical models. These results highlight the potential of hybrid machine learning frameworks to improve real-time fermentation monitoring, reduce batch variability, and support intelligent control of industrial yogurt production. Future work will explore integration with sensor data and expansion to mixed cultures and functional dairy formulations. • Hybrid LSTM–SVR modeling accurately predicts microbial growth and acidification in yogurt fermentation. • Prediction errors reduced by up to 86 % versus traditional kinetic models across industrial temperatures. • LSTM captures bacterial growth dynamics; SVR effectively models pH changes. • Validation with ten biological replicates confirms model reliability. • Method can be coupled with real-time sensors to optimize dairy fermentation control.