Metabolic-Informed Neural Network for Multistage Soft-Sensing Modeling of Microbial Fermentation Processes

To effectively monitor and control the fermentation process, an accurate real-time measurement of important variables is necessary. However, given the complexity of microbial fermentation and the significant differences in process characteristics at different stages, a single data-driven model for soft sensors cannot fully capture microbial growth patterns. Therefore, we propose a multistage microbial fermentation soft-sensing modeling method that integrates microscopic metabolic information. Initially, this paper adds consideration to the transition phase between the various stages of the fermentation process. Fuzzy C-mean clustering is used to partition the microbial growth stages. Next, unlike conventional physics-informed neural networks that only utilize macroscopic physical information, we constrain the training of the neural network by using dynamic metabolic flux analysis to obtain the rate of change of intracellular and extracellular metabolite concentrations and then incorporate it into the network to model each growth phase. Finally, dynamic time warping is applied to identify the current stage of the variable data during online measurements. The multistage model is then used for real-time measurement, and a case study of penicillin production validates the effectiveness of the proposed method.