Data Driven Approaches to Predict and Control Chinese Hamster Ovary Cell Growth in a Bioreactor System

Chinese Hamster Ovary (CHO) cells exhibit limited vulnerability to various viruses and can be grown in suspension to efficiently produce recombinant monoclonal antibodies (mAb). As a result, this cell line has become predominant in the biopharmaceutical industry, constituting 70% of therapeutic protein production. As the need for therapeutic products increases, enhancing manufacturing efficiency and maintaining consistent process stability pose significant challenges. Modeling methods that capture the dynamics of variables and process parameters in cell culture significantly enhance manufacturing efficiency. Meanwhile, automated process controllers stabilize operations and mitigate batch-to-batch variations. Despite considerable research in these areas, there remain gaps that require further exploration to improve process modeling and control effectiveness. Therefore, this dissertation aims to address research gaps in the fields of process modeling and control for a CHO cell bioreactor system, focusing on the following research tasks: (1)There is a lack of predictive models capable of accurately capturing the nonlinear dynamics of variables and process parameters under irregular sampling times, while also providing flexibility to incorporate additional variables for scaling up processes in a CHO cell bioreactor system. Many research efforts have focused on developing mechanistic models for key variables such as viable cell concentration (Xv), glucose concentration (GLC), and lactate concentration (LAC). However, these models' frameworks do not support the integration of extra variables required for process scale-up. Alternatively, several studies have employed data-driven models for variable modeling, but these models are typically limited in handling nonlinearity, irregularly sampled time-series data, and long-term forecasting. To address these gaps, a Neural ODE model was developed using experimental data from a CHO cell bioreactor to model Xv, GLC, LAC, dissolved oxygen (DO), and pH. (2)There is a lack of a Model Predictive Controller (MPC) to efficiently regulate process variables in a complex nonlinear bioreactor system. While various types of MPC have been explored in the literature, existing MPCs primarily aim to enhance control performance and often overlook addressing issues such as model plant mismatch (MPM) and computational load simultaneously. This research aims to develop a Neural ODE-based MPC to fill these research gaps concurrently. Specifically, the proposed Neural ODE-MPC is trained using a dataset and subsequently applied to regulate glucose concentration under both high and low glucose conditions in the culture medium. The control performance, MPM, and computational load of the Neural ODE-MPC are assessed using corresponding metrics and compared against established MPC benchmarks. The proposed solutions in these research tasks can enhance the accuracy of process modeling with a Neural ODE model and offer enhanced control capability through the Neural ODE-MPC in a complex bioreactor system.