Resolving complex physical phenomena with high fidelity and low computational cost is crucial for addressing key challenges in modern engineering. Hypersonic flows exemplify this, as accurate prediction of flowfield topology, particularly regarding shock wave location and intensity, is critical. However, traditional reduced-order models and neural emulators struggle with capturing steep gradients in flow states, especially in industrial applications. To address this, we introduce a fully GPU-based workflow that integrates accelerated data generation with the training of neural emulators augmented by uncertainty quantification and physics-aware refinement.
Our workflow utilizes a differentiable high-fidelity solver (JAX-Fluids) for rapid dataset creation and residual-based improvement of the neural emulator to enhance physical consistency. Building on this framework, we first present a suite of model architectures and analyze their scaling behavior to expose their strengths and weaknesses. We then demonstrate that residual-based refinement enables training on cases where only mesh and input parameters are available, significantly reducing residuals and improving physical consistency. Together, differentiable simulation and residual-based refinement yield physics emulators that remain reliable beyond their training distribution, a key requirement for deploying surrogates in real-world engineering design loops.
Blogger's Review: The fully GPU-based workflow proposed in this paper offers an innovative approach to physical emulation of hypersonic flows, overcoming the limitations of traditional models. By incorporating uncertainty quantification and physics-aware refinement, it significantly enhances the reliability and efficacy of the models, promising broad engineering applications.