Flow-informed clustering of bioreactor volumes to build CFD-based compartment models

• Novel clustering method accounting for hydrodynamics. • Clustering algorithm recognizes the reactor mixing patterns. • Generic method to build accurate compartment models from CFD and clustering data. • 3D approach can be used for most liquid phase reactor geometries. The design and operation of large-scale bioreactors are critical for industrial biotechnological productions, yet modeling these systems presents significant challenges due to the complex hydrodynamics and the occurrence of spatial gradients. Computational fluid dynamics (CFD)-based compartment models resolve spatial heterogeneity, by providing a simplified representation of the reactor volume to account for mixing dynamics and heterogeneous conditions. This study aims at further reducing the complexity of the compartment models by clustering the reactor volume using k-means and hierarchical clustering algorithms. By incorporating the mixing profile of a tracer into the loss functions, the method allows for a significant reduction of the number of compartments while maintaining high mixing accuracy. This work focuses on the methodology for extracting reduced-order, CFD-based compartment models from any high-resolution flow data set. A simplified CFD case is used solely as a demonstration platform for the clustering workflow, which is agnostic to the data source and can equally be applied to experimentally measured or validated CFD flow fields. The proposed method allows for the generation of accurate yet computationally light mixing models which facilitate the integration of detailed kinetic models, yielding more precise models for control and optimization of the biotechnological process. Moreover, the clustering method can be used as a basis to design scale-down experiments by generating compartment models with only a few clusters to replicate the behavior of the large-scale system, providing key parameters such as volumes and exchange flow rates.