Deep Generative Modeling for Predicting Turbulence Structure in Urban Flows

This study introduces a deep generative modeling approach for predicting turbulence structures in urban street-canyon flows, grounded in extensive experimental data. Utilizing high-fidelity particle image velocimetry measurements from wind tunnel experiments, we develop a novel deep learning framework that combines convolutional encoder-decoder architectures with transformer models. Our approach is tailored to capture the complex spatio-temporal dynamics of urban turbulence across various canyon geometries and upstream roughness conditions. The model is trained on detailed flow measurements at the roof level of street canyons, encompassing different width-to-height ratios and flow regimes. By integrating autoregressive training strategies and exploring diffusion model techniques, we enhance the model's ability to generate realistic flow field snapshots and predict key turbulent statistics, two-point correlations, and dominant flow structures. This research demonstrates the potential of deep generative modeling in bridging experimental fluid dynamics with advanced predictive capabilities, offering new insights into urban flow phenomena and turbulence prediction.