Inpainting U-Net for seamless pedestrian-level wind prediction across urban morphologies

Jingzi Huang, Claire E. Heaney, Tao Li, Xinzhe Li, Graham O. Hughes, Maarten van Reeuwijk

2026

Abstract

Pedestrian-level wind prediction is essential for urban wind-comfort assessment, but LES remains computationally expensive, preventing its use for rapid evaluation. This study develops a two-stage U-Net framework for efficient prediction of time-averaged wind speed over urban morphologies. The model is trained and evaluated using the UrbanTALES dataset, which contains city configurations under different wind directions. In the first stage, a baseline U-Net model (M1) predicts the wind field patch-by-patch from building height and fetch information. This patch-wise formulation allows the framework to be applied to arbitrary-size urban domains, but it can introduce discontinuities at patch boundaries because each patch is inferred independently, while limited attention has been paid to assembling patch-wise predictions into seamless full-field predictions. To address this, a second U-Net model (M2) is introduced as an inpainting-based refinement model. M2 takes a larger contextual window containing the M1 prediction and morphology, and reduces discontinuities in the M1 prediction using neighbouring flow information. During full-field inference, M2 is iteratively applied until convergence. The results show that M1 captures the spatial distribution of pedestrian-level wind speed and performs well for low- and moderate-velocity regions, although high-velocity peaks are less accurate. M2 substantially reduces patch-boundary artefacts and improves the spatial coherence of the full-field prediction. Across unseen cases, the proposed framework reproduces the mean velocity and spatial variability well, while maximum velocities remain underestimated. Overall, the proposed framework provides an efficient and flexible surrogate model for high-resolution pedestrian-level wind prediction across realistic urban morphologies.