From global AI predictions to city-scale extreme winds

Abstract Rapid advances in global AI weather models offer new opportunities for efficient tropical cyclone prediction, yet their coarse resolution limits direct prediction of urban extreme winds. This study presents an AI-physics hybrid framework that couples five AI weather prediction models with Weather Research and Forecasting (WRF)-Urban Canopy Model (UCM) for high-resolution city-scale extreme wind prediction. The framework provides skillful track and intensity predictions. Further, city-scale wind predictions are evaluated against simultaneous observations from weather stations across Hong Kong during Typhoons Ragasa (2025) and Yagi (2024). The results show that the framework generally captures the urban wind variability, while the prediction errors vary with AI driver models and local surface conditions. Sensitivity experiments indicate that enhanced land-use representation based on Local Climate Zone (LCZ) improves urban wind predictions. This study demonstrates the potential of hybrid AI-physics frameworks for reliable wind hazard warning and resilience planning in densely populated coastal cities.