Physics‐informed neural networks guided modelling and multiobjective optimization of a mAb production process

In this paper, we aim to correlate various process and product quality attributes of a mammalian cell culture process with process parameters. To achieve this, we employed physics‐informed neural networks that solve the governing ordinary differential equations comprising independent variables (inputs‐ time, flow rates, and volume) and dependent variables (outputs‐ viable cell density, dead cell density, glucose concentration, lactate concentration, and monoclonal antibody concentration). The proposed model surpasses the prediction and accuracy capabilities of other commonly used modelling approaches, such as the multilayer perceptron model. It has higher R‐squared (R2), lower root mean square error, and lower mean absolute error than the multilayer perceptron model for all output variables (viable cell density, viability, glucose concentration, lactate concentration, and monoclonal antibody concentration). Furthermore, we incorporate a Bayesian optimization study to maximize viable cell density and monoclonal antibody concentration. Single objective optimization and weighted sum multiobjective optimization were carried out for viable cell density and monoclonal antibody concentration in separate (single objective optimization) and combined (multiobjective optimization) forms. An increment of 13.01% and 18.57% for viable cell density and monoclonal antibody concentration, respectively, were projected under single objective optimization, and 46.32% and 67.86%, respectively, for multiobjective optimization as compared to the base case. This study highlights the potential of the physics‐informed neural networks‐based modelling and optimization of upstream processing of mammalian cell‐based monoclonal antibodies in biopharmaceutical operations.