AI-uWRF: A Physics-Informed Spatiotemporal Diffusion Transformer for High-Resolution Urban Weather Emulation

The urbanized Weather Research and Forecasting (uWRF) model is widely used for high-resolution urban climate modeling, yet its excessive computational cost restricts real-time forecasting and long-term climate assessment. To overcome this bottleneck, we propose AI-uWRF, a novel physics-informed generative emulator designed to bypass the computational demands of dynamical downscaling. The core architecture is a Hybrid Spatiotemporal Conditional Diffusion Model that integrates a Spatial Transformer within a U-Net backbone. A key innovation is the dual-stream condition encoder, which effectively fuses static urban surface heterogeneity (e.g., land use, topography) with dynamic large-scale atmospheric forcing. Unlike purely data-driven approaches, AI-uWRF incorporates physical constraints, including hydrostatic balance and continuity equations, into the training process to ensure thermodynamically consistent outputs. Validated against high-resolution (333 m) uWRF simulations in Wuhan, China, our emulator accelerates the generation of key meteorological fields (e.g., 2m temperature, 10m wind, surface pressure) by three orders of magnitude. The results demonstrate that AI-uWRF captures complex urban land-atmosphere interactions with high fidelity, offering a transformative tool for time-sensitive applications such as building energy optimization and probabilistic heatwave risk management.