Probabilistic Bayesian Deep Learning Approach for Online Forecasting of Fed-Batch Fermentation.

Area of Science:

• Biotechnology
• Chemical Engineering
• Data Science

Background:

• Microbial fermentation involves complex metabolic and chemical reactions.
• Mixed bacterial cultures for 2-keto-l-gulonic acid production exhibit nonlinear and time-varying dynamics.
• Accurate prediction of product formation is crucial for process optimization and control.

Purpose of the Study:

• To develop a highly accurate and robust prediction model for microbial fermentation product formation.
• To apply a probabilistic Bayesian deep learning approach to address the complexities of fermentation processes.
• To enable effective online process monitoring through reliable forecasting.

Main Methods:

• Utilized a Bayesian optimized deep neural network (BODNN) as the core prediction model.
• Optimized structural parameters of the BODNN models.
• Employed a Bayesian hybrid method for weighted combination of BODNN models, using posterior probabilities for forecasting.
• Classified training datasets based on prior prediction error evaluation.

Main Results:

• Achieved average root mean square errors of 1.51% for 4-hour ahead and 2.01% for 8-hour ahead predictions.
• Validated the model on 95 industrial fermentation batches.
• Demonstrated the model's capability to capture fermentation batch dynamics.
• Confirmed suitability for online process monitoring applications.

Conclusions:

• The proposed Bayesian deep learning approach provides accurate and robust predictions for microbial fermentation.
• The method effectively handles the nonlinear and time-varying characteristics of mixed bacterial cultures.
• This approach offers a valuable tool for real-time monitoring and optimization of industrial fermentation processes.