Hybrid Modeling of Fed-Batch Cell Culture Using Physics-Informed Neural Network

Abstract

Fed-batch cell culture is widely used in the biopharmaceutical industry to manufacture life-saving therapeutics. Biopharmaceutical manufacturing processes can be variable due to the inherent volatility and complexity of biological components and cells. Representative and robust mathematical models of cell culture are desired for the real-time monitoring and advanced process control of cell culture processes to ensure product quality and improve productivity. In this study, a novel hybrid model for a large-scale pilot Chinese hamster ovary (CHO) fed-batch cell culture was developed using physics-informed neural networks (PINNs), where first-principle biological knowledge garnered in small-scale studies is integrated with deep learning modeling of the large-scale bioreactor process data. The proposed model demonstrates both the robustness of first-principle models and the descriptive power of Neural Networks. The proposed model was tested using real-world process data and compared against a range of benchmark models, including pure data-driven models, pure mechanistic models, and other hybrid models. The model was also evaluated under various instrumentation scenarios with differing data availability. It was shown that the PINN-based model can generate accurate predictions in both scenarios and that daily calibration using measurements of cell culture analyzers can further improve the model predictions.