What happened
Researchers at the University of Reading compared forecasts of Storm Ciarán, which caused sixteen deaths in Northern Europe in November 2023. They ran FourCastNet, Pangu-Weather, GraphCast and FourCastNet-v2 through ECMWF's ai-models toolbox. All four started from the same operational ECMWF analysis as the IFS high-resolution forecast. The storm was an out-of-sample test, since the models' training data end before 2023. The authors report that the four models captured the cloud head, the warm sector and the jet of the warm conveyor belt. More than 48 hours ahead, their forecasts of the storm's rapid deepening and track matched an ensemble of physics-based models. All four underestimated the peak winds, at 25-26 m/s against 30 m/s in ERA5 and 34 m/s in the IFS analysis. None captured the sharp bent-back warm front. The journal npj Climate and Atmospheric Science published the study on 22 April 2024. The authors caution that one case cannot rank the models.
Sources
-
Peer-reviewed study, by authors with no tie to the system.
Charlton-Perez AJ, Dacre HF, Driscoll S, et al. Do AI models produce better weather forecasts than physics-based models? A quantitative evaluation case study of Storm Ciarán. npj Climate and Atmospheric Science 7, 93 (2024).
The four machine learning models considered (FourCastNet, Pangu-Weather, GraphCast and FourCastNet-v2) produce forecasts that accurately capture the synoptic-scale structure of the cyclone including the position of the cloud head, shape of the warm sector and location of the warm conveyor belt jet
Details
AI system
- Workflow
-
The models ran through ECMWF's ai-models toolbox from the operational ECMWF analysis. Forecasts were compared at lead times when forecasters were issuing warnings.
Follow-up
- Conditions for replication
-
The storm must develop like storms in the training record, with similar drivers such as the upper-level jet. Warnings of damaging winds need physics-based guidance, since all four models underestimated the peak winds.
Entry details
- Sector
-
Weather forecasting / windstorms.
- Purpose
-
NVIDIA built FourCastNet and FourCastNet-v2, Huawei built Pangu-Weather and Google DeepMind built GraphCast to forecast global weather with deep learning 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 three other AI models capture Storm Ciarán's structure (AIBD-0761). Artificial Intelligence Benefits Database, edition of 31 October 2026. https://petervartanian.xyz/aibd/records/AIBD-0761.html (archived at https://doi.org/10.7910/DVN/AADSII, version 4.0)