This conference paper was submitted for presentation at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

The Proper Orthogonal Decomposition (POD) has been used for several years in the post-processing of highly-resolved Computational Fluid Dynamics (CFD) simulations. While the POD can provide valuable insights into the spatial-temporal behavior of single transient flows, it can be challenging to evaluate and compare results when applied to multiple simulations. However, the analysis of bundles of simulations is a commonly needed task in the engineering design process, for example, when investigating the influence of geometrical changes or boundary conditions on the flow structure. A manual comparison of many simulations can be very time-consuming. On the other hand, studying only scalar and integral quantities of interest will not be sufficient to understand all relevant aspects of the flow. Therefore, we propose an automated workflow based on data-driven techniques, namely dimensionality reduction and clustering. The aim is to extract knowledge from simulation bundles arising from large-scale transient CFD simulations. We apply this workflow to investigate the flow around two cylinders as a practical example. In close proximity to the cylinders, complex shear layer interactions take place that lead to varying modal structures in the wake region. A parameter study is designed by changing the relative position of the two cylinders to each other. As a result, multiple clusters in the parameter space can be identified that each show a similar characteristic flow behavior. A particular emphasis lies in the introduction of in-situ algorithms to compute suitable data-driven representations efficiently and concurrently to the run of a simulation on the compute cluster. The in-situ data analysis approach reduces the amount of data in- and output, but also enables a simulation monitoring to reduce computational efforts, e.g., a data-driven early stopping or outlier analysis when running several simulations in engineering design studies. Finally, a classifier is trained to predict characteristic physical behavior in the flow only based on the input parameters, i.e., the relative positions of the two cylinders. This allows us to predict the principal flow dynamics for unseen parameter combinations without running additional, expensive CFD simulations.