Robust and adaptive mechanistic modelling in bioprocessing

Abstract

Mechanistic models play an essential role in the development of bioprocesses. Despite their rigorous structure, they describe the processes with interpretable model parameters and provide a mathematical representation of the underlying dynamics. That’s why they are employed extensively in process experimental design, monitoring and control.However, many obstacles still hinder their effective utilization; models are formalized according to their goals. In an industrial context, models are built with simplifications, some of which lead to discontinuous models. This issue alongside the insufficiency and peculiarities of the analytics such as the off-gas and dissolved oxygen signals, and the differences in processes conditions such as the working volume, can lead to non-adaptive (inflexible structures) non-robust (unreliable output) models. Further, similarly to the processes they describe, models must be adapted along the development life cycle. This thesis hypothesizes that proper sound scientific methods can address these challenges.The thesis aims to achieve adaptive robust models by proposing methods to: overcome the reduced predictive capabilities of the discontinuous models, extract (latent) unexploited information from already-existing analytics, properly account for analytics peculiarities, and facilitate the transferability between different scales. The thesis impact can be measured on two levels. On a scientific level, it highlights the importance of proper handling of discontinuities in bioprocessing. It provides methods to analyze and properly integrate the off-gas and dissolved oxygen signals in yeast and E. coli cultivation models. For E. coli cultivations, it reveals a new dissolved oxygen signal segment (characteristic), not reported before, and likely to be linked to metabolic adaptation behavior.On an industrial level, this work provides tools that help to reduce the number of required runs to calibrate yeast and E. coli models, and to help address milliliter scale reactor issues of oxygen supply and intermittent feeding planning, leading to a reduction in process development cost and time.The results of the proposed methods encourage further investigations on different organisms and platforms, to evaluate their generic applicability on a wider set of variable conditions.