3D High Resolution Urban Microclimate Reconstruction and Sensor Placement Optimization Based on Diffusion-based Deep Learning

3D high resolution urban microclimate is essential for optimizing sustainable city design and operation, pedestrian thermal and wind comfort, and district-scale energy resilience. However, real-world microclimate sensing is limited by sparse measurements, and high-fidelity computational fluid dynamics (CFD) simulations are computationally expensive. To overcome these challenges, this study proposes a hierarchical generative-refinement (HGR) framework for urban microclimate modeling with limited computing resources. In Step 1, an initial diffusion-based generative model is trained to predict low-resolution flow and temperature fields, which is enhanced through maximum a posterior gradient ascent (MAP-GA) to incorporate sparse measurements as conditional constraints. To enable a more efficient and explainable selection of locations for measurements (virtual measurements) in MAP-GA, a spatiotemporal variability analysis combined with SHAP-based sensitivity evaluation is conducted. In Step 2, a CNN-based model refines the low-resolution predictions in Step 1 into high-resolution microclimate fields by integrating urban morphology information, allowing the model to reconstruct fine-scale circulation patterns around complex urban geometries. On a 4 m × 4 m resolution dataset, the proposed framework achieves reconstruction errors on unseen test dataset (RMSEs) of 0.32 m/s, 0.28 m/s, 0.34 m/s, and 0.41°C for U, V, W, and T, respectively, while requiring approximately 90-120 seconds of computation. The results demonstrate that, for the tested Montréal heat-wave case, the HGR framework can efficiently reproduce high-resolution microclimate fields from limited measurements, supporting its potential for fast urban microclimate reconstruction.