Anchored-Branched Steady-state WInd Flow Transformer (AB-SWIFT): a metamodel for 3D atmospheric flow in urban environments

In order to estimate pollutant plume dispersion at a local scale in accidental release scenarios, it is necessary to estimate the air flow behavior around the affected site. This flow can be computed using Computational Fluid Dynamics (CFD), but such an approach can be computationally intensive. As a promising alternative, deep learning surrogates learned on CFD-generated data usually require cheaper resources at inference time. However, local air flows depend strongly on urban geometry, which is challenging to take into account in deep learning surrogate models. Additionally, deep learning approaches tend to struggle to scale to large meshes required by real-case scenarios. Finally, flow behavior in the Atmospheric Boundary Layer is influenced by atmospheric stratification stability, which modifies the turbulence level in the flow and must be taken into account [2].

To tackle these challenges, we propose an Anchored Branched Steady-state WInd Flow Transformer (AB-SWIFT), a transformer-based model with an internal branched structure uniquely designed for atmospheric flow modeling. AB-SWIFT relies on the anchor attention mechanism [1], allowing scalability to hundreds of millions of mesh points. To the best of the authors’ knowledge, AB-SWIFT is among the first works to apply transformer-type neural networks to atmospheric modeling. It also explicitly accounts for variable atmospheric stratification stability, which is typically neglected in existing models.

We challenge our model on a specially designed database of atmospheric simulations around randomised urban geometries and with a mixture of unstable, neutral, and stable atmospheric stratification. Urban geometries are determined by randomly sampling buildings and positioning them in space. Additionally, for each simulation, the atmospheric stratification stability is varied by sampling values of the Monin-Obukhov length and of the ground roughness.  Our model reaches the best accuracy on all predicted fields compared to the state-of-the-art transformers and graph-based models.

Figure 1: Horizontal slice at h = 2m above ground of an AB-SWIFT prediction on an unseen geometry under stable atmospheric conditions.

[1] B. Alkin, M. Bleeker, R. Kurle, T. Kronlachner, R. Sonnleitner, M. Dorfer, and J. Brandstetter. Ab-upt: Scaling neural cfd surrogates for high-fidelity automotive aerodynamics simulations via anchored-branched universal physics transformers. arXiv preprint arXiv:2502.09692, 2025.
[2] S. R. Hanna, G. A. Briggs, and R. P. Hosker Jr. Handbook on atmospheric diffusion. Technical report, National Oceanic and Atmospheric Administration, Oak Ridge, TN (USA . . . ,) 1982