Deep neural network-based hybrid modeling and experimental validation for an industry-scale fermentation process: Identification of time-varying dependencies among parameters

Kinetic modeling of fermentation processes is difficult due to the use of micro-organisms that follow complex reaction mechanisms. Kinetic models are usually not perfect owing to incomplete knowledge of the system. Recently, there is a lot of interest towards data-driven modeling as the amount of data collected, stored, and utilized is growing tremendously due to the advent of super-computing power and data storage devices. Additionally, data-driven models are simple and easy to build but their utility is restricted by the amount and quality of data required. Therefore, hybrid modeling is an attractive alternative to purely data-based modeling, wherein it combines a kinetic model with a data-based model resulting in improved accuracy and robustness. In this work, a hybrid model is developed for an industry-scale fermentation process (> 100,000 gallons) using a three-step process. The accuracy of the kinetic model is first improved utilizing process knowledge obtained from the literature. Sensitivity analyses are then utilized to identify sensitive parameters in the kinetic model that have considerable influence on its prediction capability. Finally, a deep neural network (DNN)-based hybrid model is developed by integrating the kinetic model with a DNN trained with time-series process data to predict sensitive and uncertain model parameters. The hybrid model is shown to be more accurate and robust than the kinetic model, providing a novel capability to capture unknown time-varying dependencies among parameters.