Deep Learning for Fast Flow Field Predictions in Urban Canopies

Accurate and fast prediction of urban wind flows is critical for ensuring pedestrian comfort, safety, and sustainable urban design. High wind speeds in urban canopies can cause discomfort, restrict mobility, and pose safety risks, while complex urban morphologies create interference effects that significantly alter local wind fields. Traditionally, wind tunnels and Computational Fluid Dynamics (CFD), including Large-Eddy Simulation (LES) and Reynolds-Averaged Navier–Stokes (RANS), have been used to predict these effects. While effective, these methods become prohibitively expensive and time-consuming when evaluating multiple cities at multiple wind directions, limiting their practicality for routine use.

This study presents a deep learning approach for fast and accurate prediction of urban wind fields, reducing computation time from O(10) hours to O(1) seconds for one evaluation. We employ a U-Net architecture trained on LES data from 252 synthetically generated urban configurations subject to varying wind directions. The U-Net requires 2D building footprint at a given height, augmented with a signed distance field and its gradients, and a Spatial Attention Module that enhances feature transfer through skip connections. The custom loss function combines the prediction and gradient errors with L2 regularization to improve generalizability. The model predicts the mean velocity magnitude and the turbulence intensity at multiple heights within the urban canopy.

The U-net achieves a relative error below 10% in unseen synthetic configurations, while offering significantly faster predictions compared to traditional CFD frameworks. Ongoing work extends this framework to real-world urban environments, including cities such as Chicago and Denver, by training on RANS simulation data. These results demonstrate the potential of deep learning for fast, scalable, and reliable wind assessments across diverse urban morphologies, providing urban planners and engineers with an effective tool for designing safer and more sustainable cities.