Description
- Abstract:
- Accurate prediction of hypersonic aerothermodynamic flow fields is critical for atmospheric entry vehicle design, yet high-fidelity computational fluid dynamics (CFD) simulations remain prohibitively expensive due to the need of fine grids required to resolve strong shock waves, thin boundary layers, and stiff thermochemical reaction networks. This thesis investigates the use of Deep Operator Networks (DeepONets) as surrogate models for steady-state laminar hypersonic flow over a generic sphere–cone entry configuration with an 18-species thermochemical nonequilibrium reacting flow model and carbon surface ablation. Both single-fidelity (SF) and multi-fidelity (MF) neural operator architectures are developed and evaluated. High-fidelity training data are generated using the NASA's FUN3D solver, which solves the compressible Navier–Stokes equations with two-temperature non-equilibrium chemistry, while lower-cost aerodynamic data are obtained using the engineering-level CBAero code. The multi-fidelity framework combines these heterogeneous data sources through a shared trunk network encoding a common output function space representation, with separate branch networks capturing fidelity-specific corrections. The MF-DeepONet achieves the largest improvement for pressure, reducing the relative $L^2$ error from 9.77\% to 2.04\%, with consistent improvement also observed for density. Temperature fields show more modest gains due to weaker cross-fidelity correlation arising from nonequilibrium thermochemical processes not captured by the low-fidelity model. Sensitivity studies confirm that incorporating even modest amounts of low-fidelity data substantially stabilizes training when high-fidelity supervision is sparse. The influence of multi-fidelity loss weighting strategies on convergence and accuracy is also investigated. Results indicate that pressure and density benefit from stronger high-fidelity supervision, while temperature fields respond better to curriculum-style schedules. Once trained, both surrogate models provide inference times orders of magnitude faster than CFD, supporting their use in parametric studies and design-space exploration for atmospheric entry applications.
- Notes:
- Thesis (Sc. M.)--Brown University, 2026
Citation
Ali, Saleem Adam,
"Learning Hypersonic Flow Fields From Sparse Data Using a Multi-Fidelity Neural Operator"
(2026).
Engineering Theses and Dissertations.
Brown Digital Repository. Brown University Library.
https://repository.library.brown.edu/studio/item/bdr:3nggavku/
