Abstract

To train a neural network as a surrogate model for simulations on an unstructured mesh, such as computational fluid dynamics (CFD) simulations, it is helpful to maintain the structure of the mesh, rather than interpolating onto a regular grid. We describe our recent work using graph neural networks and graph transformers for unstructured mesh data. Large-scale simulations on supercomputers often involve meshes that do not fit on one processor and are split with a domain decomposition. We maintain the same mesh partitioning while training the neural network, enforcing "physical consistency" via halo nodes. This work is demonstrated using data from NekRS, a spectral-element CFD solver, and we demonstrate efficient scaling of training across many GPUs.