Problem description

In large-scale fermentation processes, microbes experience and respond to heterogenous conditions when they move around the bioreactor, which may impact process performance.

A fermentation digital twin would be able to simulate and predict the microbial response, allowing ultimately for monitoring, control and optimization of the fermentation process. Currently, a prototype of such a digital twin is available, by combining first-principal spatially resolved hydrodynamic models and dynamic metabolic models.  

The prototype is however computationally very intensive and a real-time online application, which should also include process dynamics (such as changes in volume, aeration, and rheology, etc.), is infeasible. Recent studies show that data-driven approaches, such as Physics-Informed Neural Networks, incorporating physical constraints (Navier-Stokes equation), is able to reconstruct flow fields and can be employed even in direct flow modelling, allowing for substitution of the computationally intensive fluid dynamics calculation. However, these promising approaches have not been extended to industrially-relevant geometry (3D bioreactor with stirrers) and process conditions (e.g., gas and liquid phase) and to include microbial response.