TC-Diffusion: A diffusion-based probabilistic tropical cyclone model with application to typhoon wind hazard assessment

This paper introduces TC-Diffusion, a novel probabilistic tropical cyclone (TC) full track simulation method empowered by the diffusion generative model. Leveraging diffusion’s exceptional capability to process high-dimensional space, our model naturally considers the spatial heterogeneity of TC dynamics as conditional information, avoiding the need of segmentation or clustering. Additionally, it can incorporate various sources of prior information to guide the TC simulation. For validation, 100 realizations of 74-year synthetic TCs are drawn from the proposed diffusion model and compared against historical TCs in the China Meteorological Administration (CMA) dataset. The synthetic TCs have excellent agreement with historical ones in terms of their key statistics, spatial distributions, and intensity development. Simulated TCs are ultimately combined with a parametric wind field to estimate wind hazards along the China southeastern coastline. Our estimated design wind speed generally agrees with alternative full-track methods, verifying the effectiveness of the proposed method for typhoon wind hazard assessment.