level: research
hypersonic flows are hard to model because they contain steep gradients like shock waves. traditional reduced-order models and neural emulators often fail to capture these features accurately in industrial settings. this work presents a fully gpu-based workflow that combines fast data generation with neural emulator training. the approach uses a differentiable high-fidelity solver called jax-fluids to create datasets quickly and improve predictions through residual-based refinement.
the workflow includes uncertainty quantification and physics-aware refinement to make the emulators more reliable. by running everything on gpus, the process avoids bottlenecks from data transfer between cpu and gpu. the differentiable solver allows gradients to flow through the simulation, which helps in training and correcting the neural network. this setup aims to produce emulators that can predict full flowfield topology, including shock wave location and intensity, with high fidelity and low computational cost.
the method targets a key challenge in modern engineering: resolving complex physical phenomena efficiently. hypersonic flow prediction is critical for applications like spacecraft reentry and high-speed flight. the gpu-based workflow could make it practical to use neural emulators where traditional methods are too slow or inaccurate. by integrating data generation and training in one accelerated pipeline, the approach reduces the time and resources needed to build useful models.
why it matters: faster and more accurate flow emulators can cut design time and cost for high-speed vehicles and spacecraft.