Unsupervised learning bioreactor regimes

Abstract

Efficient operation of bioreactors is crucial for the success of biomanufacturing processes. Traditional Computational Fluid Dynamics (CFD) simulations provide detailed insights but often involve lengthy computation times and complexity, hindering their practicality for real-time applications. This study introduces a novel multivariate unsupervised learning algorithm that clusters bioreactors into physically meaningful regions based on CFD-generated and real-world data. These clusters not only facilitate the determination of internal reactor regimes but also serve as a foundational step for developing compartment models. Our approach utilizes a custom k-means clustering algorithm, which ensures spatial continuity of clusters by incorporating geometric data, and optimizes the number of compartments to maximize physical significance and data retention. This optimization is guided by a Pareto front analysis, balancing the need for clear compartment definition with the preservation of maximum information from the dataset. The effectiveness and versatility of this methodology were verified through case studies involving a 202 m³ Rushton impeller bioreactor (steady state simulation) and an 840 m³ airlift reactor (dynamic simulation). In the airlift reactor, the clustering algorithm accounted for dynamic fluctuations by averaging the simulation results, providing a robust method for incorporating temporal variations into the compartment analysis. The findings highlight the advantages of 3-D compartmentalization in capturing the intricate dynamics of fluid motion and cellular activities, thereby advancing the design of bioreactors and scaling down experiments for more efficient industrial applications.