Overview
A workflow for constructing physics emulators of hypersonic flows has been developed, leveraging a fully GPU-based architecture. This approach aims to address challenges in resolving complex physical phenomena with fidelity and computational efficiency, particularly in contexts like hypersonic flows where accurate prediction of flowfield topology, including shock wave location and intensity, is critical.
The workflow integrates accelerated data generation with the training of neural emulators. It incorporates uncertainty quantification and a physics-aware refinement process. This integration is supported by a differentiable high-fidelity solver, JAX-Fluids, which facilitates rapid dataset creation and a residual-based improvement mechanism for the neural emulator, intending to enhance physical consistency.
Research Context
Traditional reduced-order models and neural emulators have historically encountered difficulties in accurately capturing steep gradients within flow states, particularly when applied to supersonic and hypersonic flows in industrial settings. These limitations pose a barrier to achieving physical consistency in such applications. The ability to resolve complex physical phenomena while maintaining low computational costs is a central challenge in modern engineering.
Approach
The developed workflow is entirely GPU-based. It utilizes JAX-Fluids, a differentiable high-fidelity solver, for two primary functions: enabling rapid dataset creation and facilitating residual-based improvement of the neural emulator. This residual-based refinement is designed to enhance the physical consistency of the emulator.
The research first involved presenting a suite of model architectures. Subsequently, an analysis of their scaling behavior was conducted to identify their strengths and limitations. Following this, the efficacy of residual-based refinement was investigated in scenarios where only mesh data and input parameters were available. The objective was to determine if this refinement could reduce residuals and improve physical consistency under these conditions.
Findings
- The fully GPU-based workflow enabled accelerated data generation and training of neural emulators augmented by uncertainty quantification and physics-aware refinement.
- JAX-Fluids, used as a differentiable high-fidelity solver, supported rapid dataset creation and residual-based improvement of the neural emulator.
- Residual-based refinement enabled training in cases where only mesh and input parameters were known.
- This residual-based refinement led to a substantial reduction in residuals.
- The refinement process also improved physical consistency.
- The combination of differentiable simulation and residual-based refinement resulted in physics emulators that maintained reliability even when applied beyond their initial training distribution.
Why This Matters
The development of physics emulators that maintain reliability beyond their training distribution is identified as a key requirement for deploying surrogate models within real-world engineering design loops. The methodology presented, which allows for substantial reduction in residuals and improved physical consistency, is relevant for industrial applications requiring precise prediction of flowfield topology, especially regarding shock wave location and intensity, in hypersonic flow regimes.