# GraphCast: Google DeepMind's AI weather model beat the ECMWF gold standard on 90% of 1,380 targets, and made a 10-day forecast in under a minute

> In a paper published in Science in November 2023, Google DeepMind's GraphCast machine-learning model outperformed ECMWF's operational High Resolution Forecast on 90% of 1,380 verification targets while producing a 10-day global forecast in under one minute, a result ECMWF's own head of Earth-system modelling publicly acknowledged.

- Verification status: verified
- Case type: deployment
- Provider: GraphCast (Google DeepMind) (https://deepmind.google/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/)
- Client: Google DeepMind (research result, benchmarked against ECMWF's operational HRES), tech (named)
- Sector: tech / UK / ops
- Canonical URL: https://theinternetninja.com/stories/google-deepmind-graphcast-outperforms-ecmwf-hres-90pct-1380-targets-2023/
- Source: The Internet Ninja (theinternetninja.com), independent verified-proof platform

## Outcomes

| Metric | Before | After |
| --- | --- | --- |
| Share of 1,380 verification targets where GraphCast beat ECMWF HRES | ECMWF HRES (operational deterministic gold standard) | GraphCast more accurate on 90% of 1,380 targets |
| Time to produce a 10-day global forecast | hours on a supercomputer (traditional numerical weather prediction) | under one minute on a single Google TPU v4 |

## Verification method

Retrieved live this session (2026-09-01) from the peer-reviewed Science paper (DOI 10.1126/science.adi2336, confirmed verbatim via its identical arXiv manuscript 2212.12794), cross-checked against two independent outlets (MIT Technology Review and The Register) and Google DeepMind's own blog. Every figure quoted verbatim and archived to the Wayback Machine, with local copies in sources/. The two critical figures are stated by an independent peer-reviewed publication and reproduced by independent press, not solely by the developer.

## FAQ

**How much more accurate was GraphCast than ECMWF's HRES model?**

In the Science paper, GraphCast significantly outperformed the most accurate operational deterministic system, ECMWF's HRES, on 90% of 1,380 verification targets. Independent reporting by MIT Technology Review and The Register reproduced the same figure.

**How fast is GraphCast compared with traditional weather prediction?**

GraphCast predicts hundreds of weather variables over 10 days at 0.25-degree global resolution in under one minute on a single Google TPU v4 machine, against the hours of supercomputer time that conventional numerical weather prediction needs.

## Full case file

**Verification status: checker-verified against the public record, awaiting the owner's final
validation.** This is an independent, peer-reviewed research finding, checked against primary and
independent sources; the green badge follows the owner's sign-off and is not claimed here.

## The problem
Traditional numerical weather prediction improves accuracy by adding compute, but "cannot
directly use historical weather data to improve the underlying model"
([source](https://arxiv.org/abs/2212.12794)). The open question was whether a machine-learning
model trained on decades of reanalysis data could match, or beat, the operational deterministic
gold standard, ECMWF's High Resolution Forecast (HRES), which The Register describes as the
benchmark GraphCast was measured against
([source](https://www.theregister.com/2023/11/15/google_deepmind_graphcast/)).

## What was built
GraphCast is "a machine learning-based method" that "can be trained directly from reanalysis
data" and "predicts hundreds of weather variables, over 10 days at 0.25 degree resolution
globally, in under one minute" ([source](https://arxiv.org/abs/2212.12794)). Google DeepMind
introduced it "in a paper published in Science" as "a state-of-the-art AI model able to make
medium-range weather forecasts with unprecedented accuracy"
([source](https://deepmind.google/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/)).
The Register reported that the research "was published in the journal Science on Tuesday,
November 15, 2023" ([source](https://www.theregister.com/2023/11/15/google_deepmind_graphcast/)).

## The outcome
The headline result is stated in the peer-reviewed abstract: "GraphCast significantly
outperforms the most accurate operational deterministic systems on 90% of 1380 verification
targets, and its forecasts support better severe event prediction, including tropical cyclones,
atmospheric rivers, and extreme temperatures" ([source](https://arxiv.org/abs/2212.12794)).
Two independent outlets reproduced the same figure: MIT Technology Review reported that
"GraphCast outperformed the model from the European Centre for Medium-Range Weather Forecasts
(ECMWF) in more than 90% of over 1,300 test areas"
([source](https://www.technologyreview.com/2023/11/14/1083366/google-deepminds-weather-ai-can-forecast-extreme-weather-quicker-and-more-accurately/)),
and The Register reported that "GraphCast provided more accurate predictions on more than 90
percent of 1,380 test variables and forecast lead times"
([source](https://www.theregister.com/2023/11/15/google_deepmind_graphcast/)).

The speed gain is the second measured result: GraphCast makes those "10-day forecasts" in a
run that "takes less than a minute on a single Google TPU v4 machine"
([source](https://deepmind.google/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/)),
which MIT Technology Review put plainly as GraphCast doing "these calculations in under a
minute" ([source](https://www.technologyreview.com/2023/11/14/1083366/google-deepminds-weather-ai-can-forecast-extreme-weather-quicker-and-more-accurately/)).
Google DeepMind also reports a narrower, stronger sub-result: in the troposphere, "our model
outperformed HRES on 99.7% of the test variables for future weather"
([source](https://deepmind.google/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/)).

## What the finding does and does not show
The 99.7% troposphere figure is the weakest load-bearing claim on this page: it comes only from
Google DeepMind's own blog, a Tier 3 first-party source, not from the independently verifiable
abstract or the two independent outlets, so it is reported here as the developer's own figure and
is not treated as a verified headline number
([source](https://deepmind.google/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/)).
The two figures that are independently corroborated, the 90%-of-1,380 accuracy result and the
under-one-minute speed, carry unusual weight because the institution whose model was beaten
acknowledged the result: MIT Technology Review quotes ECMWF's head of Earth-system modelling,
Peter Dueben, saying "it showed that these models are so good that we cannot avoid them anymore"
([source](https://www.technologyreview.com/2023/11/14/1083366/google-deepminds-weather-ai-can-forecast-extreme-weather-quicker-and-more-accurately/)).

> ## How this was verified
> Method: every figure was retrieved live on 2026-09-01 from the peer-reviewed Science article
> (DOI 10.1126/science.adi2336, published 2023-11-14), confirmed verbatim against its identical
> arXiv manuscript (2212.12794), and cross-checked against MIT Technology Review's and The
> Register's independent 2023 reports plus Google DeepMind's own blog. Each quotation is
> verbatim; all four sources were captured on the Wayback Machine and saved locally in
> `sources/`. The two critical figures are stated by an independent peer-reviewed publication
> and reproduced by two independent outlets, the strongest available form of independence for a
> benchmark result.

## Related case files
- [AlphaFold, the same lab's protein-structure model that topped an independent CASP benchmark, another DeepMind result graded against a neutral yardstick](/stories/deepmind-alphafold-casp14-92-4-gdt-independent-top-rank-2024-nobel-chemistry/)
- [DeepMind's machine-learning system that raised the value of Google's wind energy about 20%, an applied-science outcome from the same team](/stories/google-deepmind-machine-learning-wind-energy-value-up-20pct-700mw-2019/)
- [AlphaProof and AlphaGeometry 2 reaching an IMO silver-medal score, another DeepMind result measured against an external competition standard](/stories/deepmind-alphaproof-alphageometry2-imo-2024-silver-medal-28-of-42/)

## Sources
1. Lam, Sanchez-Gonzalez, Willson, Wirnsberger, Fortunato, Alet, Ravuri, Ewalds, Eaton-Rosen, Hu, Merose, Hoyer, Holland, Vinyals, Stott, Pritzel, Mohamed, Battaglia · "Learning skillful medium-range global weather forecasting" · Science, Vol. 382, Issue 6677, pp. 1416-1421 · 2023-11-14 · DOI 10.1126/science.adi2336; verbatim manuscript at arXiv:2212.12794 · **Tier 1 (primary, peer-reviewed publication)** · https://arxiv.org/abs/2212.12794
2. MIT Technology Review (June Kim) · "Google DeepMind's weather AI can forecast extreme weather faster and more accurately" · 2023-11-14 · **Tier 2 (independent press reproducing the 90%/1,380 and under-a-minute figures)** · https://www.technologyreview.com/2023/11/14/1083366/google-deepminds-weather-ai-can-forecast-extreme-weather-quicker-and-more-accurately/
3. The Register (Katyanna Quach) · "Google DeepMind's GraphCast AI predicts the weather faster and more accurately than gold-standard system" · 2023-11-15 · **Tier 2 (second independent press outlet for the 90%/1,380 figure and the Science publication)** · https://www.theregister.com/2023/11/15/google_deepmind_graphcast/
4. Google DeepMind · "GraphCast: AI model for faster and more accurate global weather forecasting" · 2023-11-14 · **Tier 3 (developer's own blog; sole source for the 99.7% troposphere sub-claim, used only qualitatively)** · https://deepmind.google/blog/graphcast-ai-model-for-faster-and-more-accurate-global-weather-forecasting/