AI Weather Models Show Competitive Skill at Extremes, New Study Finds
A new study posted on arXiv (2608.09972) challenges the idea that first-generation AI weather models struggle during extreme weather. Over ten months, researchers evaluated eleven AI and physical forecasting systems using European synoptic, solar, and rain-gauge data. They found that AI models don’t consistently underperform in severe conditions. The study analyzed metrics like 10 m wind and hourly precipitation, measuring mean absolute error (MAE) against ECMWF IFS within ERA5 1991-2020. Notably, Jua EPT-2.1 Europa showed an 8.4% improvement in wind forecasts, while Jua EPT-2 HRRR had a 12.1% gain in temperature predictions. The results suggest that AI models are competitive with traditional ones, even during extreme events, and the researchers' names remain undisclosed.
Key facts
- Study verifies 11 physical and AI forecast systems against European stations over 10 months.
- Metrics: 10 m wind, 2 m temperature, hourly shortwave, hourly precipitation; scored MAE vs ECMWF IFS.
- AI models do not show uniform relative-skill deficit in extremes.
- Jua EPT-2.1 Europa leads all-conditions wind (+8.4%).
- Jua EPT-2 HRRR leads temperature overall (+12.1%) and heat regime (+19.6 ± 2.2%).
- EPT-2.1 Europa and DWD ICON Global lead at gale-force wind.
- Jua EPT-2.1 Helios leads solar overall (+10.2 ± 1.7%), overcast (+16.4 ± 3.4%), clear-sky tail (+24.8 ± 5.4%).
- Three Jua models gain for precipitation (details incomplete).
Entities
Institutions
- arXiv
- ECMWF
- DWD
- Jua
Locations
- Europe