Surrogation of High-Fidelity Simulation Software

The proposed innovation is a mesh-based graph neural network (GNN) framework for training high-fidelity surrogate models. The models are highly accurate and increase computational efficiency approximately 1-2 orders of magnitude faster than the software which generated the data. This would allow use of previously generated simulations to alleviate High-Performance Computing (HPC) resources, minimize the time-to-solution for engineering and science workflows, and increase the collaboration of various initiatives by providing a leaner alternative to simulations.The mesh-based GNN, or simply MeshGraphNetwork (MGN), is based on a pioneering publication from Googlersquo;s DeepMind project shown to accurately predict the dynamics of a wide range of physical systems, including those found within computational fluid dynamics (CFD). MGNs are a novel class of GNNs that directly operate on irregular meshes with arbitrary connectivity which have difficulty scaling on hardware accelerators. Due to the neural network basis of the MGNs components, it is suitable for acceleration on hardware commonly present in HPCs.