Effective process control is a basic requirement for biopharmaceutical manufacturing to achieve high efficiency and enhanced quality control. High non-linearity and uncertainties associated with bioprocesses challenge the ability of traditional controllers to deliver satisfactory performance, thereby creating a need for advanced model based control strategies for efficient bioprocess control. Unlike the chemical or traditional pharmaceutical sector, bioprocess control is a complex process that involves controlling the behaviour of billions of cells in a bioreactor that evolves with non-linear dynamics over time. Variability in cell culture behaviour arising from heterogeneous culture conditions, product types, and cell types make developing a generalised control action challenging. Historic bioprocess data can provide valuable insights into the underlying dynamics of cell culture. Process Analytical Technology (PAT) initiative has highlighted the importance of identifying critical process parameters (CPP) of a bioprocess that influence critical quality attributes (CQA), to achieve seamless integration of analytical data with real-time monitoring and control for enhanced process understanding and to overcome manufacturing challenges. This thesis explores the different stages of development of a novel data-driven dynamic control scheme for bioprocesses in the context of antibody producing CHO cell cultures in fed-batch bioreactors.

Multivariate monitoring techniques in biopharmaceuticals has resulted in the generation of large amounts of data comprising real-time measurements of critical quality and performance attributes, and if exploited efficiently can provide opportunity for developing superior control action. This work employs a novel methodology for condensing online data into an offline data matrix to achieve a comprehensive view of the critical process parameter landscape. The methodology was found superior to traditionally employed averaging and helped increase the number of variables available for representing the design space of the process. The augmented data set was then used to extract novel parameter relationships by applying concepts from network theory. The ability of these newly identified parameter relationship in describing the process efficiently was tested by constructing symbolic regression models. The performance of newly identified variables in predicting process behaviour was found superior to variable relationships found in literature. Following this, new models were generated using support vector machines (SVM) to predict the performance of these cultures at a future time point. These models helped decide the expected trajectory of a culture, based on process knowledge derived from historical bioprocess data. Customised control strategies were developed, to acknowledge the process dynamics of different days of the culture, by employing an optimisation algorithm along with the different models developed throughout the course of this study. The control scheme dynamically recalculates the expected trajectory and proposes customised reactive control action when encountered with a deviation from the expected trajectory. The proposed control scheme was able to recommend sensible control actions during the different test cases designed for control. This closed-loop model-based multi-attribute control scheme, that employs concepts from data science, network theory and control theory, ensure that the cultures remain on a pre-defined well established trajectory thereby minimising variability.