Abstract Summary (Max 250 words)
Predicting heterogeneities in bioreactors is crucial for scaling up fermentations, as spatial gradients can strongly influence microorganisms and fermentation performance. Although Computational Fluid Dynamics (CFD) coupled with biokinetic models can capture these effects, such simulations are computationally intensive and expensive [1-3], limiting their use for rapid assessment of dynamically operated systems such as fed-batch fermentations. Compartment models (CM) provide an attractive alternative, offering a computationally efficient yet sufficiently detailed representation of hydrodynamics and mass transport in stirred tank bioreactors [4-7]. When combined with machine-learning techniques, these models can further capture the impact of varying operating conditions without relying on CFD-derived flow fields [8]. In this presentation, we discuss our approach to integrating compartment models with machine learning to construct spatially resolved surrogate models for exploring bioreactor design and operation. Depending on the application, supervised or unsupervised methods can be used – either to approximate the flow field and decouple hydrodynamics from biokinetics [8], or to directly infer substrate gradients and metabolic regimes resulting from their interaction [9-10]. Furthermore, deep learning enables the incorporation of categorical variables, expanding the range of design and operational scenarios that can be explored. Our surrogate models achieve a runtime reduction of up to three orders of magnitude compared to a fully coupled CFD model, while maintaining resolution suitable for industrial decision-making. We also highlight key considerations for deployment, including enforcing linear constraints to ensure mass conservation and mitigating spurious predictions outside expected operating ranges.
