Artificial Intelligence-Guided Optimization of Food Fermentation: Fermentation Kinetics, Microbial Dynamics and Process Control

Abstract

Food fermentation is a nonlinear biological process in which temperature, pH, inoculum size, substrate availability, oxygen transfer and microbial physiology interact over time. Conventional optimization commonly relies on one-factor-at-a-time experimentation or deterministic kinetic models, which can obscure parameter uncertainty and batch-to-batch variability. The present study proposes and demonstrates an artificial intelligence (AI)-guided stochastic framework that couples a mechanistic fermentation model with Monte Carlo simulation, machine-learning surrogate modeling and risk-aware optimization. A synthetic fermentation space covering temperature, initial pH, inoculum density, sugar concentration and agitation was generated. A mechanistic kinetic model represented biomass growth, substrate consumption and acidification, while 5,000 simulated fermentation trajectories were used to train random-forest surrogate models for a composite fermentation-performance score and time to target acidification. The optimized operating region was approximately 37.36 °C, pH 5.27, inoculum 7.45 log10 units, sugar 99.38 g/L and agitation 136 rpm. The surrogate model achieved R² = 0.965 for the composite performance score and R² = 0.972 for fermentation time on a held-out test set. Relative to a conventional baseline (35 °C, pH 5.5, inoculum 6.0 log10 units, 60 g/L sugar and 100 rpm), the optimized point increased the predicted composite score from 81.58 to 91.94 while reducing predicted time from 20.75 to 7.89 h. A 10,000-run Monte Carlo analysis around the optimized region produced a mean composite score of 90.94 (SD 5.39) and mean fermentation time of 8.75 h (SD 1.24). The estimated probability of simultaneously achieving a score ≥85 and fermentation time ≤12 h was 84.2%. Temperature and initial pH were the dominant drivers of model response. These findings demonstrate the potential of combining mechanistic knowledge, AI surrogates and stochastic simulation for fermentation design; however, because the dataset is simulated, the numerical results should be interpreted as a methodological proof-of-concept requiring experimental validation before process deployment.