Machine learning for process monitoring and adaptive control in continuous chromatography
File(s)
Author(s)
Michalopoulou, Foteini
Type
Thesis
Abstract
Process efficiency and adaptability are becoming growing priorities in biopharmaceutical manufacturing. In recent years, the industry has moved towards intensified and continuous operations to meet rising demand for monoclonal antibodies and other therapeutic proteins. While such approaches promise higher productivity and reduced variability, they also introduce cyclic dynamics and non-linear behaviour that complicate downstream purification. Within this context, chromatographic separations remain a critical challenge. Mechanistic models offer detailed predictions for these systems, but their high computational cost can restrict their use in online applications. This highlights an opportunity for machine learning approaches that combine mechanistic fidelity with the computational efficiency required for real-time decision-making.
This thesis presents modelling and control frameworks that address these challenges. Firstly, the potential of data-driven models is examined, using neural networks to approximate full elution profiles and enable rapid evaluation. Building on this, a hybrid modelling approach is introduced, embedding mechanistic structure within a neural network to improve robustness while avoiding spatial discretisation. Both data-driven and hybrid models are then integrated into Bayesian optimisation workflows, demonstrating how surrogate-based methods can sustain reliable performance under online horizons where mechanistic models fail to converge. Finally, reinforcement learning is explored as a complementary paradigm for adaptive control, trained to regulate flowrate and switching times under variability, disturbances, and measurement noise.
To conclude, the presented findings establish a toolbox of machine learning methods that extend the scope of mechanistic models into online applications. The results point towards future integration of optimisation and adaptive control within digital biomanufacturing, supporting the transition to more efficient and resilient chromatographic processes.
This thesis presents modelling and control frameworks that address these challenges. Firstly, the potential of data-driven models is examined, using neural networks to approximate full elution profiles and enable rapid evaluation. Building on this, a hybrid modelling approach is introduced, embedding mechanistic structure within a neural network to improve robustness while avoiding spatial discretisation. Both data-driven and hybrid models are then integrated into Bayesian optimisation workflows, demonstrating how surrogate-based methods can sustain reliable performance under online horizons where mechanistic models fail to converge. Finally, reinforcement learning is explored as a complementary paradigm for adaptive control, trained to regulate flowrate and switching times under variability, disturbances, and measurement noise.
To conclude, the presented findings establish a toolbox of machine learning methods that extend the scope of mechanistic models into online applications. The results point towards future integration of optimisation and adaptive control within digital biomanufacturing, supporting the transition to more efficient and resilient chromatographic processes.
Version
Open Access
Date Issued
2025-10-18
Date Awarded
01/02/2026
License URL
Advisor
Papathanasiou, Maria
Publisher Department
Department of Chemical Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)