Model Predictive Controller Design For Bioprocesses Based On Machine Learning Algorithms
Mohammad Rashedi, Hamid Khodabandehlou, Matthew Demers, Tony Wang, Christopher Garvin
Abstract
The optimization of critical quality attributes in biopharmaceutical processes demands the development of a scalable and optimal control scheme to meet the process constraints and objectives. In this paper, we designed a model predictive controller (MPC) to find the optimal feeding strategy to maximize cell growth and metabolite production in fed-batch bioprocesses. Due to high complexity of bioprocesses and lack of high-fidelity first principle models, we evaluated the use of machine learning algorithms in the forecast model to aid in our development. By taking advantage of the bioprocess model, this controller aims to maximize the protein production daily for each batch. The control scheme of the bioprocess is defined as an optimization problem to be solved while all metabolites and cell culture process variables are maintained within the specification. To evaluate the performance of the controller, we designed and implemented MPC with the best model to a bioreactor in a real experiment. The experimental validation confirms more than 2% improvement in final protein production compared to average historical experiments.