AI/ML and Graph-assisted Robust Shock Detection and Processing in Numerical Simulations including Hypersonic Flows

In this talk, we present a set of procedures that provide robust shock detection and processing for accurate high-speed flow simulations involving complex geometry. Our approach is capable of starting with extremely coarse unstructured meshes and adaptively constructing highly accurate and optimized meshes. We employ a sequence of three strategies. The first strategy uses a gated recurrent neural network or transformer architecture. Such networks are trained to predict the presence and structure of shocks, and to fill in missing or fragmented detection results from traditional physics-based shock detectors/sensors, thereby improving the robustness of shock identification. This is followed by the second strategy that uses graph-based procedures to augment or enhance the shock processing due to the first strategy. Graph-based procedures provide an efficient way to detect, process and label mesh elements containing different shocks. The resulting information is used in the third strategy that employs shock alignment and fitting to construct a mesh size field (based on mesh metric tensors) and applies anisotropic meshing. The entire process results in highly accurate and optimized meshes, e.g., those with highly anisotropic elements properly aligned around shocks. We demonstrate our approach on multiple high-speed flow cases including hypersonic flows (e.g., CEV case), and showcase its robustness and utility.