What happened
Two researchers at Uppsala University compared Pangu-Weather and GraphCast with ECMWF's high-resolution deterministic model, IFS HRES. All three started from the same IFS HRES analysis, in a semi-operational setting. Forecasts began every 12 hours from 1 January to 16 December 2020 and were scored against ERA5. The authors report that the AI models mostly outperformed IFS HRES on global error 1 to 10 days ahead. They could also compete on temperature and wind extremes in most regions. GraphCast beat IFS HRES in the tropics and the North Pacific for all variables at all lead times. IFS HRES did better on cold extremes in Australia, New Zealand and Antarctica, and mostly on hot extremes in Europe, North America and East Asia. For extremes, the AI models lost ground at longer lead times. Geoscientific Model Development published the study on 7 November 2024. The evaluation covers one year and uses forecasts from the public WeatherBench 2 dataset.
Sources
-
Peer-reviewed study, by authors with no tie to the system.
Olivetti L, Messori G. Do data-driven models beat numerical models in forecasting weather extremes? A comparison of IFS HRES, Pangu-Weather, and GraphCast. Geoscientific Model Development 17, 7915-7962 (2024).
We find that data-driven models mostly outperform ECMWF’s physics-based deterministic model in terms of global RMSE for forecasts made 1–10 d ahead and that they can also compete in terms of extreme weather predictions in most regions.
Details
AI system
- Workflow
-
The operational versions of both AI models took the IFS HRES analysis as input. Forecasts every 12 hours through 2020 were scored against ERA5 at 1.5 degrees.
Follow-up
- Conditions for replication
-
The AI model must start from the same operational analysis as the physics-based model. Performance varies by region, type of extreme and lead time.
Entry details
- Sector
-
Weather forecasting / temperature and wind extremes.
- Purpose
-
Huawei built Pangu-Weather and Google DeepMind built GraphCast to forecast global weather with deep learning models trained on ERA5 reanalysis.
Not yet checked: model, release, provider, role, configuration, inputs and outputs, autonomy, human oversight, beneficiaries, reach, distribution and access, duration, every measurement question, every money question, ownership and control, dependencies, cross-border connections, geopolitical implications, replication attempts, availability, institutional decision, grounds, action and result, success criterion, every evidence question, observed or intended, alternative assessments, related entries, other identifiers, category mappings.
Cite this record
Peter H. Vartanian. (2026). Pangu-Weather and GraphCast rival ECMWF on weather extremes (AIBD-0763). Artificial Intelligence Benefits Database, edition of 31 October 2026. https://petervartanian.xyz/aibd/records/AIBD-0763.html (archived at https://doi.org/10.7910/DVN/AADSII, version 4.0)