The deployment of mechanistic models for advanced bioprocess monitoring and control

Abstract

Producers of biopharmaceuticals and other biotechnological products are obliged to guarantee the quality of their products. To achieve high quality, critical process parameters need to be measured in real-time and maintained within predefined limits. Especially during biological catalysis, where the target product is synthesized by living organisms, the control of multiple interacting parameters is necessary. Unfortunately, not all influential process parameters of these multi-analyte solutions can be directly measured. Corrective measures, after the timeconsuming analysis of a sample, are often not effective and lead to irreversible changes. Process models, consisting of transferable and applicable knowledge, can be used to estimate non-measured states and to optimally control these interacting parameters. In this regard, mechanistic models, describing the mechanism of a system, with a distinct model structure and characteristic parameters are very promising. However, mechanistic models are still rarely applied in biotechnological production. The complexity of living organisms make the establishment of useful process models a non-trivial task. Often, the models are over-fitted, which compromises their applicability. In addition to that, the availability of software tools including modeling workflows and suitable real-time communication systems for their deployment are limited. By building upon recent developments in state estimation, model predictive control and today’s possibilities offered by the ongoing digitization, the thesis aims at offering viable modelbased solutions for both academia and industry. The focus hereby is the deployment of process models for bioprocess monitoring and control. Therefore, generic methods for model analysis, state estimation and control were developed and used for different organisms. Based on a Penicillium chrysogenum fed-batch process, it could be shown that: • Particle filters are versatile state estimation algorithms and well suited for bioprocesses. • By inclusion of time-delayed, offline data and infrared spectroscopy, difficult-to-measure process variables were made accessible in real-time. • Based on model-based state observation different control strategies could be tested and experimentally verified. The use of model predictive control process lead to a significant increase in process performance. Through lab-scale verification experiments and the successful transfer of the developed tools on Escherichia coli, the generic functionality could be proved. In order to test advanced controllers in fast-growing E. coli processes, a system with a backup controller was set up to guarantee ongoing control. The results of the thesis demonstrate that the usage of mechanistic models enables the estimation of difficult to measure process parameters and allows to keep them under control. This is a prerequisite to guarantee product quality and safety during production and helps biopharmaceutical producers to fulfill regulatory requirements and to continuously improve their production. Towards an increased utilization of mechanistic models in academia and industry, the thesis provides a basis for further developments. Increasing software functionalities and suitable real-time communication systems, driven by the current wave of digitization and Industry 4.0 will further contribute to their diffusion.