A geometry-conditioned transformer solver for predicting three-dimensional wind distribution in urban microclimate design

Sustainable urban microclimate design requires fast and accurate assessments of urban wind environments. While computational fluid dynamics (CFD) simulations have proven effective for detailed assessments, they remain computationally intensive for iterative design processes. Recently, deep learning-based methods have emerged as promising alternatives for solving parametric partial differential equations (PDEs). However, their effectiveness remains largely unexplored in real-world urban scenarios characterized by complex building geometries. In this study, we proposed the Urban Geometry-conditioned Wind Transformer (UrbanGWT), a novel deep learning model for efficient prediction of three-dimensional urban wind environment. We evaluated UrbanGWT on a dataset of real-world building layouts in Hong Kong, which is a typical high-density metropolis with complex building forms and arrangements. To demonstrate the effectiveness of the proposed model, we benchmarked the model against several existing grid-based and geometric deep learning (GDL)-based baselines. The results showed that UrbanGWT outperformed all baselines in terms of overall accuracy, reducing the mean absolute error (MAE) and root mean square error (RMSE) by 6.91 % and 6.98 %, respectively, compared to the state-of-the-art deep learning model, while achieving a speedup of up to 85,125× over the traditional CFD simulations. This research highlights the potential of conducting efficient and accurate wind assessments for urban microclimate design.