GPU Workflow Accelerates Physics Emulators for Hypersonic Flow Simulation.

Accurately resolving complex physical phenomena at low computational cost is a critical challenge in modern engineering, particularly for hypersonic flows where precise prediction of flowfield topology, including shock waves, is essential. Traditional reduced-order models and neural emulators often struggle to capture steep gradients with physical consistency in industrial applications. This research introduces a novel, fully GPU-based workflow designed to build physics emulators for hypersonic flows. The workflow integrates accelerated data generation with the training of neural emulators, enhanced by uncertainty quantification and physics-aware refinement. A key enabler is JAX-Fluids, a differentiable high-fidelity solver used for rapid dataset creation and residual-based improvement of the neural emulator, ensuring greater physical consistency. The framework allows for training even when only mesh and input parameters are available, significantly reducing residuals and improving physical consistency. By combining differentiable simulation with residual-based refinement, the resulting physics emulators maintain reliability even when applied to conditions outside their original training distribution, which is crucial for real-world engineering design loops.