Fast prediction of wind fields around an isolated building with FiLM U-Net
Abstract
CFD can describe the wind-speed distribution around an isolated building at different heights in considerable detail. However, a complete CFD calculation usually requires geometric modeling, mesh generation, boundary-condition setting, iterative solution, and post-processing, which makes it difficult to compare many building-mass alternatives during conceptual design. To address this problem, this study uses horizontal wind-field images exported from Fluent to construct trainable samples from a building footprint mask, 10-channel spatial condition maps, and a 22-dimensional scalar condition vector. A FiLM-conditioned U-Net is then developed for rapid prediction of horizontal wind-field slices around isolated buildings. The spatial condition maps describe the relationships between pixels and the building boundary, wake direction, and height level, while the scalar condition vector provides global geometric constraints including W, L, H, z, and z/H. On the complete building-mass test set, the model achieved MAE = 0.0223, Global R2 = 0.9415, SSIM = 0.9447, and IoU4 = 0.9637. The model preserves the wake region, corner-flow region, and transition structure near the roof with good stability, and can therefore serve as a rapid auxiliary tool for early-stage assessment of the building wind environment.