Downscaling & super-resolution★ Highlight

Intelligent Shanghai Typhoon Model (ISTM): A Diffusion‐Based Downscaling Emulator From AIFS to Kilometer‐Scale Regional Reanalysis for Typhoon Forecasts

Zeyi Niu, Wei Huang, Sirong Huang, Bo Qin, Mengqi Yang, Haofei Sun, Zhaoyang Huo, Haixia Xiao, Kairui Feng

Journal of Geophysical Research Machine Learning and Computation · 01 April 2026 · China Meteorological Administration (CMA), Fudan University (FuXi)

Paper

Why it matters. Diffusion downscaling of AIFS to 9-km typhoon fields cuts 72-hr intensity error ~39.5%, addressing AI models' known weak-typhoon bias.

In one sentence. Diffusion residual emulator downscales AIFS to 9-km typhoon fields, cutting 72-hour typhoon intensity error by about 39.5% relative to AIFS.

Abstract

Abstract Artificial‐intelligence weather prediction models have recently surpassed numerical models in large‐scale skill, but they still systematically underestimate typhoon intensity due to their reliance on coarse‐resolution training data from ERA5. To overcome this limitation, we constructed a bespoke 9‐km high‐resolution typhoon reanalysis (HiRes) by dynamically downscaling ERA5 with the operational Shanghai Typhoon Model (SHTM). The evaluation shows that HiRes has a clear advantage over ERA5 in aspects such as surface wind and land precipitation. HiRes inherently combines a large‐scale deterministic background with small‐scale residual corrections, motivating a learning framework that mirrors this structure. We therefore introduce the Intelligent Shanghai Typhoon Model (ISTM), a regression‐residual diffusion emulator that first predicts the conditional mean fields and then stochastically reconstructs the residuals with a conditional diffusion model. Trained on HiRes, ISTM substantially improves wind and precipitation intensity and organization compared with ERA5 and deterministic U‐Net baselines. After fine‐tuning, ISTM can downscale AIFS forecast fields to produce SHTM‐consistent, physically coherent fields; while maintaining track errors comparable to AIFS, it reduces the 72‐hr mean typhoon intensity forecast error by about 39.5% relative to AIFS.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.