Transformer–Diffusion Modeling of Unsteady Flow over an Iced Airfoil
Ice accretion can induce large-scale flow separation and severely degrade aerodynamic performance, while high-fidelity simulations of the unsteady separated flows remain computationally expensive. This study investigates the capability of generative learning for rapid prediction of such complex flow over a realistic iced airfoil. The generative learning framework consists of a training-free encoder compressing flowfields into latent representations, a transformer propagating the temporal evolution in latent space, and a diffusion-based decoder reconstructing full-resolution flowfields. The framework is trained on unsteady simulation data of an iced airfoil and evaluated at unseen angles of attack and Reynolds numbers. Results demonstrate accurate prediction in three key aspects. First, the model captures the flow separation structures across different regimes. Second, it predicts integral aerodynamic loads, including lift, drag, and moment coefficients, with high accuracy across all tested conditions. Third, it reliably reproduces the unsteady temporal evolution, including the vortex-shedding period and pointwise flow history. The trained generative learning model achieves reasonable speedup compared with conventional unsteady Reynolds-averaged Navier–Stokes simulations. These results highlight the potential of generative learning as an efficient surrogate for rapid prediction of unsteady flows under icing conditions.
Authors
- Hui Hu (ORCID: https://orcid.org/0000-0001-7928-4138)
- Ping He (ORCID: https://orcid.org/0000-0001-5074-7215)
- Jiawei Chen
Institutions
- Iowa State University (US)
Publication Details
- Journal
- AIAA Journal
- Published
- 2026-11-16
- DOI
- https://doi.org/10.2514/1.j067206
- Primary Topic
- Model Reduction and Neural Networks
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Directorate for Engineering