P34 - Interpretable Compression of Fluid Flows Using Graph Neural Networks
DescriptionNeural network (NN) based reduced-order models (ROMs) via autoencoding have been shown to drastically accelerate traditional computational fluid dynamics (CFD) simulations for rapid design optimization and prediction of fluid flows. However, many real-world applications (e.g. hypersonic propulsion, pollutant dispersion in cities, wildfire spread) rely on complex geometry treatment and unstructured mesh representations in the simulation workflow — in this setting, conventional NN-based modeling approaches break down. Instead, it is necessary to use frameworks that (a) easily interface with unstructured grid data, and (b) are not restricted to single geometric configurations after training. The goal here is to address this through the development of an interpretable autoencoding strategy based on the graph neural network (GNN) paradigm. More specifically, a novel graph autoencoder architecture is developed for ROM-amenable autoencoding. An adaptive graph pooling strategy, combined with multiscale message passing operations, is shown to produce interpretable latent spaces through the identification of coherent structures. With this notion of interpretability established, analysis is then conducted on effects of compression factors, physical significance of identified coherent structures, and impact of multi-scale message passing on reconstruction errors.
TimeTuesday, June 2719:30 - 21:30 CEST