A GAN-Based Surrogate Model for Instantaneous Urban Wind Flow Prediction

Abstract

Urban form impacts the airflow patterns in cities and the resulting urban microclimate. This has significant implications for ventilation, overheating, wind chill, and safety concerns such as down drafts from skyscrapers. While Computational Fluid Dynamics (CFD) simulations are the best practice for analyzing urban airflow patterns in design, they are computationally expensive and require a high level of expertise, making them underutilized in the early design process. This paper presents a surrogate model for CFD using a Generative Adversarial Network (GAN) that can process arbitrary building geometries. The model is trained using an automated end-to-end pipeline based on Eddy3D and implemented within the Rhino and Grasshopper environment as an Open Neural Network Exchange (ONNX)-based CFD-GAN predictor. This workflow provides instantaneous simulation feedback within the design software, reduces the risk of user error, and allows for appropriate spatial resolution in early design. The CFD-GAN demonstrates promising accuracy, with a Structural Similarity Index Measure (SSIM) ranging from 75%–97% on a limited training dataset of 564 unique urban geometries. Although the model currently has limitations regarding accuracy in complex urban wake regions, we show that these are likely not of concern for outdoor thermal comfort analyses. While it cannot replace CFD in later design stages, the CFD-GAN facilitates the incorporation of urban airflow analysis in early design with minimal effort and instantaneous performance feedback.