Three-dimensional primitive prediction with building embedded Swin Transformer for urban turbulence

Large-eddy simulation (LES) is the benchmark for resolving canonical urban canopy physics critical for pedestrian comfort, pollutant dispersion modeling, and Urban Air Mobility operations. However, generating long-horizon LES data requires prohibitively long simulation times, making it impractical for real-time or large-scale applications. To overcome these limitations, we develop 3DSwinUrbanNet, a three-dimensional (3D) Swin-Transformer-based model that directly predicts the full primitive variables (u,v,w,p) using time steps orders of magnitude larger than those in the computational fluid dynamics solver (Δt=1 min vs LES Δt=0.05 s). 3D building geometry embeddings are integrated into the encoder, backbone, and decoder via average-pooling over patches and linear projection into the latent space. Model accuracy is evaluated using time-averaged fields, wall-normal profiles, turbulence statistics, and scale-resolved diagnostics (probability density functions and energy spectra). This approach yields a 1984× speedup (2.58 s vs LES 5120 s for 20 min forecasting) without precursor overhead, enabling real-time predictions on a single Ray Tracing Texel eXtreme 4090 graphic processing unit.