Physics-Aware Generative Modeling of Hypersonic Aerodynamics
Please login to view abstract download link
The accurate prediction of shock wave location and intensity is paramount in hypersonic vehicle design. Traditional surrogate models often fail to capture the steep gradients associated with shocks. In this study, we propose a physics-aware generative approach utilizing diffusion models fine-tuned via variants of Direct Preference Optimization (DPO). We leverage the high-performance JAX-Fluids solver to synthesize a comprehensive dataset of high-fidelity flow fields for model training. While diffusion models have recently gained significant attention for forecasting physical systems, they in general focus on all spatial features equally, which is not the best approach for sharp discontinuities. To enhance their capabilities, we introduce a preference-based fine-tuning stage. Unlike traditional Reinforcement Learning approaches that require an explicit reward model, DPO or its variants optimize the diffusion policy solely using a dataset of preferred and non-preferred flow reconstructions. We generate this dataset automatically, identifying preferred samples based on their gradient sharpness and adherence to the numerically discretized governing PDE of the system. We validate our approach against high-fidelity CFD data, demonstrating that DPO-tuned diffusion models successfully recover shocks and maintain conservation properties better than baseline generative methods.
