A fully GPU-based workflow for building physics emulators of hypersonic flows

The article presents a fully GPU-based workflow for developing physics emulators for hypersonic flows, utilizing a differentiable high-fidelity solver called JAX-Fluids for efficient data generation and training. It introduces various model architectures and demonstrates that residual-based refinement significantly enhances physical consistency and reliability of the emulators, even outside their training distribution. This approach addresses the limitations of traditional reduced-order models and neural emulators, making it a valuable tool for engineers tackling complex flow phenomena in industrial applications.