RP: AeroGNN: A Multi-Fidelity Physics-Informed Graph Neural Network for Aerodynamic Design Optimization of Tall Buildings
Proceedings of the AAAI Symposium Series · 06.11.2026 · 10 (1) · 612-618
DOI: 10.1609/aaaiss.v10i1.43463 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Symposium Series, online erschienen am 06.11.2026. Verfasst von Dhanya Ganesh und 1 weiteren Autor:innen.
Abstract (Originalsprache)
Tall buildings are subject to unsteady wind-induced lateral forces that cause vortex-induced vibrations (VIVs), leading to structural fatigue and occupant discomfort. Nonetheless, exploring the large design spaces needed to identify optimal geometries with high-resolution computational fluid dynamics (CFD) or wind-tunnel testing is resource-intensive. This study presents AeroGNN, a physics-informed multi-fidelity graph neural network for rapid aerodynamic optimization of tall building designs. To enable wide design exploration, geometries are generated across a nine-dimensional parameter space defined by the Gielis superformula and three-dimensional modifications. These geometries are evaluated with OpenFOAM Reynolds-averaged Navier–Stokes simulations at coarse and fine fidelity levels. The model exploits the resulting inexpensive coarse-resolution data to learn the underlying flow physics, while fewer high-resolution simulations are used to refine the predictions. Aerodynamic force coefficients are predicted alongside the pressure and velocity fields, and these coefficients are combined into a scalar aerodynamic performance score. Relative to fine-resolution CFD, the framework accelerates design evaluation by a factor of 364. The final model achieved a cross-validated R-squared of 0.849 for the aerodynamic score. Coupled with differential evolution, the model identified a geometry that reduced the aerodynamic loading by 75.9% relative to a sharp square baseline.
Zitieren
APA
Ganesh, D., & Wall, D. (2026). RP: AeroGNN: A Multi-Fidelity Physics-Informed Graph Neural Network for Aerodynamic Design Optimization of Tall Buildings. Proceedings of the AAAI Symposium Series, 10(1), 612-618. https://doi.org/10.1609/aaaiss.v10i1.43463
Vancouver
Ganesh D, Wall D. RP: AeroGNN: A Multi-Fidelity Physics-Informed Graph Neural Network for Aerodynamic Design Optimization of Tall Buildings. Proceedings of the AAAI Symposium Series. 2026;10(1):612-618. doi:10.1609/aaaiss.v10i1.43463
BibTeX
@article{Ganesh2026,
author = {Ganesh, Dhanya and Wall, Danny},
title = {RP: AeroGNN: A Multi-Fidelity Physics-Informed Graph Neural Network for Aerodynamic Design Optimization of Tall Buildings},
journal = {Proceedings of the AAAI Symposium Series},
year = {2026},
volume = {10},
number = {1},
pages = {612--618},
doi = {10.1609/aaaiss.v10i1.43463}
}
EUVIMED