Graph Neural Networks for Interpretable Data-Based Modeling of Fluid Flows
DescriptionReduced-order modeling strategies based on neural networks can accelerate traditional computational fluid dynamics simulations for rapid design optimization and prediction of a wide range of fluid flows. To realize this vision of improved modeling, key limitations -- namely, incompatibility with unstructured data representations and latent space interpretability -- prohibiting their extension into practical flow configurations must be tackled. This work addresses these limitations with a novel graph neural network (GNN) architecture. In the context of fluid flow compression, it is shown how the method produces a latent graph that (a) can be visualized in physical space directly, (b) identifies coherent structures in the domain, and (c) is described by an adjacency matrix that adapts in time with the evolution of the flow. Model outputs are assessed on an unsteady and unstructured turbulent fluid flow dataset for both autoencoding and forecasting applications, and additional emphasis is placed on demonstrating the scalability of underlying graph operations on modern GPU-based compute nodes. Through the ability to unify autoencoder-based reduction and physical interpretability into a single framework, this work presents a pathway for improved data-driven modeling in complex geometry configurations.
TimeMonday, June 2615:00 - 15:30 CEST
Computer Science, Machine Learning, and Applied Mathematics