High-Resolution Urban Microclimate Reconstruction with Generative Diffusion-Based AI Incorporating Sparse Measurements

Monitoring urban microclimate is essential for sustainable urban design, pedestrian wind comfort assessment, and district-scale energy resilience. However, real-world microclimate sensing is constrained by sparse measurements, while high-fidelity computational fluid dynamics (CFD) simulations remain prohibitively expensive. To address these challenges, this study introduces a fast and accurate two-step urban microclimate reconstruction framework driven by sparse sensor observations. In the first step, a diffusion-based generative model predicts low-resolution wind and temperature fields while assimilating sparse measurements as conditional inputs. In the second step, a CNN-based super-resolution model enhances these coarse predictions into high-resolution microclimate fields by leveraging detailed building-distribution information, enabling the recovery of fine-scale flow structures around complex urban geometries. The framework was trained and evaluated using a realistic urban CFD dataset. On a dataset with 4 m × 4 m spatial resolution originally generated by CityFFD simulations with 54 million grids, the proposed method achieves reconstruction RMSEs of 0.34 m/s, 0.29 m/s, 0.35 m/s, and 0.41 °C for U, V, W, and T, respectively, while requiring 90-120 seconds of computation for one prediction. These results demonstrate that the framework can efficiently reproduce highresolution urban microclimate fields from sparse measurements, enabling fast reconstruction for urban planning and resilience applications.