Deep learning for urban wind prediction: an MLP-Mixer approach with 3D encoding

Abstract

Pedestrian-level wind environments are strongly influenced by urban morphology, with large and tall buildings playing a significant role. City authorities increasingly mandate assessments of pedestrian wind conditions before approving new construction. Computational fluid dynamics (CFD) models can provide a detailed understanding of the aerodynamic environment; however, in early design stages, urban morphologies are subject to change, requiring multiple simulations, adding substantial financial and time burdens to projects. To address this challenge, we develop a deep learning approach for the rapid inference of pedestrian-level wind conditions using a multi-layer perceptron (MLP)-mixer architecture. By embedding 3D structural details into the training data, our model can infer wind conditions around complex structures such as lift-up designs and skyways while maintaining inference times on the order of fractions of a second. This extends the capabilities of deep learning models that typically reduce the problem to a 2D image-to-image translation task, omitting crucial structural details. We conduct an extensive evaluation of our model and compare its performance to the widely adopted UNet architecture, demonstrating that the MLP-mixer outperforms UNet across all evaluation metrics. Notably, the MLP-Mixer achieves a mean squared error approximately 2.6 times lower, a peak signal-to-noise ratio 3.7 dB higher and the highest recorded structural similarity index of 0.991. These results indicate improved agreement with the reference CFD data. We anticipate that the MLP-mixer model will serve as a valuable tool in early-stage urban design workflows, enabling faster and more efficient wind assessments.