Diffusion model sharpens AI typhoon forecasts of wind and rain at low cost
FuXi-TC uses a diffusion model trained on WRF simulations to sharpen FuXi forecasts of tropical cyclone wind and rain, nearly matching ECMWF HRES on intensity and beating it on heavy rain, in about 2 minutes per forecast.
Why it matters
AI weather models forecast cyclone tracks well but underestimate wind strength. Physics-based downscaling fixes this but is slow. This approach gets much of the benefit at a small fraction of the cost, which could allow fast updates and large ensembles.
What they did
The authors ran WRF at 0.1° resolution, driven by FuXi forecasts, to build high-resolution training targets from the 2023 typhoon season. They trained a diffusion model to turn coarse FuXi forecasts into detailed fields of wind, pressure and rain. They tested it on Western North Pacific storms in 2024 and 2025 against best-track data and rain gauges. They also applied it to FuXi ensemble members and to North Atlantic hurricanes without retraining.
Key findings
- FuXi-TC had lower maximum wind errors than the base FuXi model and ERA5, and was comparable to ECMWF HRES on the 2024 subset.
- HRES gave slightly smaller intensity errors than FuXi-TC on the full set of 43 storms.
- For heavy rain (above 20 mm and 40 mm per 6 h), FuXi-TC scored higher than FuXi, HRES and WRF across nearly all lead times.
- Applied to FuXi ensemble members, it reduced under-dispersion and improved intensity probability scores, with CRPS comparable to ECMWF-ENS.
- A 120-h forecast took 2 minutes on one A100 GPU, versus about 11.5 hours on 32 CPUs for WRF at 0.1°.
- Without retraining, it beat FuXi and ERA5 on North Atlantic wind errors across 15 hurricanes, but HRES was better in the mid-range of the forecast.
Limitations
- Performance may be capped by systematic biases in the WRF physics used to make the training targets.
- The ensemble is still under-dispersed compared with ECMWF-ENS, and wind speed remains biased low.
- Rain verification covers only mainland China, using gauge data with uneven availability that makes scores oscillate.
Glossary
- Diffusion model: A generative model that learns to turn random noise into realistic data step by step.
- WRF: A physics-based regional weather model used here to produce high-resolution training targets.
- Spread-skill ratio: Ensemble spread divided by the error of the ensemble mean; a value of one means well-calibrated.
- Critical success index: A score for yes/no rain forecasts that counts hits against misses and false alarms.
Original paper
FuXi-TC: a generative framework integrating deep learning and physics-based models for improved tropical cyclone forecasts
npj Climate and Atmospheric Science · 8 September 2026
AI-generated summary of the original article; changes were made. Check the original before relying on it.
