GraphCast and Pangu-Weather beat ECMWF on the 2021 winter storm

In three case studies, GraphCast and Pangu-Weather forecast the February 2021 North American winter storm better than ECMWF's HRES. They did less well on two heatwaves.

  • Not rated an organization

    The authors compare forecasts of past events, and no forecaster or utility is reported to have used them.

Depth runs from 1, a convenience or time saved, to 5, a life or a livelihood preserved.

Status
Reported and intended benefits.
Evidence

foundstudiedreportedclaimed

An independent peer-reviewed study supports the benefit.

Compared with

ECMWF HRES forecasts for the same events and lead times. Measured.

The system’s part
The authors ran the AI models themselves and compared their forecasts with HRES for the same events, places and lead times.
Limits
Not yet checked.
Costs and harms
Not yet checked.
Later
Not yet checked.

Download data Cite Suggest a correction

Benefit type
Improved Task Accuracy
System
GraphCast, Pangu-Weather and FourCastNet
Made by
Google DeepMind (GraphCast), Huawei (Pangu-Weather) and NVIDIA (FourCastNet)
Where
Texas and the southern United States (2021 winter storm), the Pacific Northwest (2021 heatwave) and South Asia (2023 humid heatwave)
Stage
Tested in a laboratory
Date
, when the source was published
Kind
Aggregate result
Broad benefit
Estimating the Quantifiable Characteristics of Products, Events, or Information
Last revised
31 October 2026

Sources

  1. Peer-reviewed study, by authors with no tie to the system.

    Pasche OC, Wider J, Zhang Z, Zscheischler J, Engelke S. Validating deep learning weather forecast models on recent high-impact extreme events. Artificial Intelligence for the Earth Systems 4(1) (2025).

    We find that ML weather prediction models locally achieve similar accuracy to HRES on the record-shattering Pacific Northwest heatwave but underperform when aggregated over space and time. However, they forecast the compound winter storm substantially better.

Details

AI system

Workflow

The authors ran the pretrained models released by their developers. Forecasts from all initialization times were compared at grid cells near affected cities.

Follow-up

Conditions for replication

Health indices need variables such as surface humidity, which the AI models did not predict. Single case studies should not be used alone to judge forecasting systems.

Entry details

Sector

Weather forecasting / extreme events.

Purpose

Google DeepMind built GraphCast, Huawei built Pangu-Weather and NVIDIA built FourCastNet to forecast global weather with deep learning models trained on reanalysis data.

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 beat ECMWF on the 2021 winter storm (AIBD-0764). Artificial Intelligence Benefits Database, edition of 31 October 2026. https://petervartanian.xyz/aibd/records/AIBD-0764.html (archived at https://doi.org/10.7910/DVN/AADSII, version 4.0)

BibTeX Data Field definitions