Abstract
Designing re-entry vehicles requires accurate predictions of hypersonic flowaround their geometry. Rapid prediction of such flows can revolutionize vehicledesign, particularly for morphing geometries. We evaluate advanced neuraloperator models such as Deep Operator Networks (DeepONet),parameter-conditioned U-Net, Fourier Neural Operator (FNO), and MeshGraphNet,with the objective of addressing the challenge of learning geometry-dependenthypersonic flow fields with limited data. Specifically, we compare theperformance of these models for two grid types: uniform Cartesian and irregulargrids. To train these models, we use 36 unique elliptic geometries forgenerating high-fidelity simulations with a high-order entropy-stable DGSEMsolver, emphasizing the challenge of working with a scarce dataset. We evaluateand compare the four operator-based models for their efficacy in predictinghypersonic flow field around the elliptic body. Moreover, we develop a novelframework, called Fusion DeepONet, which leverages neural field concepts andgeneralizes effectively across varying geometries. Despite the scarcity oftraining data, Fusion DeepONet achieves performance comparable toparameter-conditioned U-Net on uniform grids while it outperforms MeshGraphNetand vanilla DeepONet on irregular, arbitrary grids. Fusion DeepONet requiressignificantly fewer trainable parameters as compared to U-Net, MeshGraphNet,and FNO, making it computationally efficient. We also analyze the basisfunctions of the Fusion DeepONet model using Singular Value Decomposition. Thisanalysis reveals that Fusion DeepONet generalizes effectively to unseensolutions and adapts to varying geometries and grid points, demonstrating itsrobustness in scenarios with limited training data.