Physics-Informed Neural Surrogates for Mesh-Invariant Modeling of High-Speed Flows

High-speed flight is a critical area for national defense and space exploration, yet accurately predicting vehicle behavior under such extreme conditions remains challenging. While high-fidelity simulations can capture the complicated physical characteristics, they are often too computationally demanding for early design stages or real-time decision-making. To address this, we present a data-driven surrogate modeling framework that enables rapid and accurate predictions of aerodynamic behavior. Our method uses point-wise multilayer perceptrons (MLPs) that are conditioned on flight parameters, making the model independent of any specific mesh. The training data, generated with the SPARC solver, span a range of Mach numbers, altitudes, and vehicle geometries -- including parametrically varied cones and a lifting-body waverider. By embedding the continuity equation through a vector potential formulation, our model enforces mass conservation directly. Integration with the Reactant compiler and MLIR-based differentiation eliminates redundant computation in nested automatic differentiation, yielding $1.4 - 1.5\times$ geometric mean end-to-end speedups during both training and inference. The results show accurate reconstruction of lift, drag, and flow fields across test cases, demonstrating that compiler-optimized, physics-informed surrogates can deliver rapid and reliable aerodynamic predictions suitable for design exploration and onboard applications.