Neural ordinary differential equation-based model predictive controller for regulating glucose concentration in a fed-batch CHO cell bioreactor

Abstract

Chinese hamster ovary (CHO) cells are widely used in the biopharmaceutical industry to produce recombinant proteins. Effective process control is crucial for managing biomanufactured product production in response to increasing market demand. Model predictive control (MPC) is an advanced controller compared to the traditional proportional integral derivative (PID) controller for handling complex nonlinear systems. However, existing MPC controllers fail to address challenges related to control accuracy, model plant mismatch (MPM), and computational load simultaneously. Neural ordinary differential equation (ODE), capable of effectively modelling dynamics within complex systems at high computational efficiency, has the potential to tackle these limitations. This study developed a neural ODE model-based MPC to dynamically maintain glucose concentration in a fed-batch CHO cell bioreactor simulation system. Additionally, benchmark studies were conducted to compare the control performance of neural ODE-MPC with neural network (NN)-based MPC and long short-term memory (LSTM)-based MPC. The results demonstrate that neural ODE-MPC can provide reliable control performance in managing glucose concentration with lower control errors, small MPM, and higher computational efficiency compared to the benchmark systems. In conclusion, neural ODE-MPC has the potential to address MPC challenges and enhance production efficiency in future industrial applications.