Machine learning model-based design and model predictive control of a bioreactor for the improved production of mammalian cell-based bio-therapeutics

Abstract

This study is concerned with the development of reduced order machine learning (ML) and non-ML model representations of experimental and simulated bioprocesses and their implementation in model predictive control (MPC) strategies to quantify performance accuracy and computational efficiency compared with the original models. Results showed that ML models such as Long Short-Term Memory (LSTM) networks and Artificial Neural Networks (ANNs) outperformed other reduced order models such as Kriging, Multiple Linear Regression (MLR) and Random Forest (RF) in terms of performance metrics such as R 2 and RMSE for both experimental and simulated data. Experimental data were obtained from a fed-batch and perfusion-based bioprocess and an LSTM model was developed and implemented in an MPC open-loop optimal control strategy to determine optimal input trajectories to maximize key performance metrics such as product titer. For the 2 by 3 ODE simulation, results showed that an autoregressive ANN was the most accurate in terms of replicating the plant model dynamics under MPC conditions followed by the LSTM and transfer function (TF) representations, with the feedforward ANN not being able to fully capture the salient dynamics. For the 4 by 5 ODE simulation, the TF representation outperformed the feedforward ANN model which in turn was more accurate than the LSTM model. In terms of computational time, the plant model simulation time for an MPC solution is intractable for larger input-output s…