What happened
Researchers at ETH Zurich, the University of Bern, the University of Lausanne and Columbia University studied three storms that hit the United Kingdom in February 2020. Storms Ciara, Dennis and Jorge came in close succession, with strong winds and heavy rain. The authors compared GraphCast and Pangu-Weather with the mean of ECMWF's 50-member operational ensemble, cycle 46r1. They checked all forecasts against ERA5, and the storms fell outside the AI models' training periods. The authors report that the AI models tended to outperform the ensemble mean for pressure at sea level at weekly timescales, and to a lesser extent for surface winds. GraphCast had the lowest pressure error, an average of 3.23 hPa. Some single ensemble members matched or occasionally beat the AI models. The authors add that weaker links between errors in different variables may mean that the AI models overlook physical constraints. The journal npj Natural Hazards published the study on 27 May 2026. The authors recommend AI forecasts for impact warnings only alongside a physics-based forecast.
Sources
-
Peer-reviewed study, by authors with no tie to the system.
Afargan-Gerstman H, Wu RWY, Ferrini A, Domeisen DIV. Evaluating the skill of AI and physics-based models for February 2020 North Atlantic storm hazards. npj Natural Hazards 3, 76 (2026).
Our results show that, for these case studies, AI models tend to outperform the numerical model in predicting mean sea level pressure (MSLP) on weekly timescales, and, to a lesser extent, surface winds.
Details
AI system
- Workflow
-
Forecasts of the three storms were scored against ERA5 at 1.5 degrees. The two AI models were compared with the ensemble mean and with its best member.
Follow-up
- Conditions for replication
-
The storms must fall outside the AI models' training periods. Impact warnings need a physics-based forecast alongside the AI forecast.
Entry details
- Sector
-
Weather forecasting / windstorms.
- Purpose
-
Google DeepMind built GraphCast and Huawei built Pangu-Weather 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). GraphCast and Pangu-Weather foresee Britain's February 2020 storms (AIBD-0762). Artificial Intelligence Benefits Database, edition of 31 October 2026. https://petervartanian.xyz/aibd/records/AIBD-0762.html (archived at https://doi.org/10.7910/DVN/AADSII, version 4.0)