Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 66 of the 66 references it cites were found in a registry or an index; 12 of the 38 sources its claims cite were checked for support.

Learning skillful medium-range global weather forecasting

Open access

Abstract

Global medium-range weather forecasting is critical to decision-making across many social and economic domains. Traditional numerical weather prediction uses increased compute resources to improve forecast accuracy but does not directly use historical weather data to improve the underlying model. Here, we introduce GraphCast, a machine learning–based method trained directly from reanalysis data. It predicts hundreds of weather variables for the next 10 days at 0.25° resolution globally in under 1 minute. 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 cyclone tracking, atmospheric rivers, and extreme temperatures. GraphCast is a key advance in accurate and efficient weather forecasting and helps realize the promise of machine learning for modeling complex dynamical systems.

Overview

CoChat Summary

TL;DR

GraphCast is a graph neural network trained directly on ERA5 reanalysis that produces 10-day global weather forecasts at 0.25° resolution in under one minute. It outperforms the operational HRES deterministic system on most evaluated targets and supports better prediction of tropical cyclones, atmospheric rivers, and extreme temperatures.

Key findings

  • GraphCast significantly outperformed HRES on 90.3% of 1380 verification targets, and significantly on 89.9% of targets; excluding 50 hPa it significantly outperformed HRES on 96.9% of 1280 targets, and excluding 50 and 100 hPa on 99.7% of 1180 targets.
  • On z500, GraphCast had better skill across all lead times, with skill score improvement around 7%–14%; it also outperformed Pangu-Weather on 99.2% of the 252 targets Pangu-Weather presented.
  • For severe events, GraphCast had lower median cyclone track error and was significantly better than HRES for lead times 18 hours to 4.75 days; it improved atmospheric river ivt prediction from 25% at short lead time to 10% at longer horizons; extreme heat precision-recall curves were above HRES at 5- and 10-day lead times, while HRES was better at 12 hours.
  • Training on more recent data improved GraphCast’s skill scores, with variants trained up to before 2021 generally outperforming earlier-trained variants on 2021 test data.

Methods

  • GraphCast is a 36.7-million-parameter graph neural network in an encode-process-decode configuration, using a multi-mesh icosahedral representation with 40,962 nodes and 327,660 mesh edges; it takes two most recent weather states and predicts the next 6-hour state autoregressively.
  • It was trained on 39 years (1979–2017) of ERA5 reanalysis, using a weighted MSE objective over 12 autoregressive steps, with training taking roughly four weeks on 32 Cloud TPU v4 devices; evaluation used held-out 2018 data, with some analyses on 2018–2021 and 2021.
  • Forecast skill was evaluated against HRES using RMSE and ACC, with HRES-fc0 as HRES ground truth and ERA5 as GraphCast ground truth; comparisons used 06z/18z initializations and 12-hour target intervals to match data assimilation lookahead.

Limitations

  • The paper states that GraphCast is deterministic and does not explicitly model uncertainty like ECMWF’s ensemble system ENS; its MSE training objective encourages spatial blurring, which may limit value for some applications.
  • The paper states GraphCast operates at 0.25° resolution, 37 vertical levels, and 6-hour time steps, versus HRES at 0.1°, 137 levels, and up to 1-hour time steps, and that it depends critically on large quantities of high-quality data assimilated via NWP, so it should not be regarded as a replacement for traditional weather forecasting.
  • The paper reports reduced skill in the stratosphere, especially at 50 hPa, and at high surface elevations; it also notes possible ERA5 versus HRES-fc0 domain differences and spurious spectral artifacts around 100 km wavelengths.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1126/science.adi2336

Cites

66 of 66 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 | Model schematic. (a) The input weather state(s) are defined on a 0.25° latitude-longitude grid comprising a total of 721 × 1440 = 1, 038, 240 points. Yellow layers in the closeup pop-out window represent the 5 surface variables, and blue layers represent the 6 atmospheric variables that are repeated at 37 pressure levels (5 + 6 × 37 = 227 variables per point in total), resulting in a state representation of 235, 680, 480 values. (b) GraphCast predicts the next state of the weather on the grid. (c) A forecast is made by iteratively applying GraphCast to each previous predicted state, to produce a sequence of states which represent the weather at successive lead times. (d) The Encoder component of the GraphCast architecture maps local regions of the input (green boxes) into nodes of the multi-mesh graph representation (green, upward arrows which terminate in the green-blue node). (e) The Processor component updates each multi-mesh node using learned message-passing (heavy blue arrows that terminate at a node). (f) The Decoder component maps the processed multi-mesh features (purple nodes) back onto the grid representation (red, downward arrows which terminate at a red box). (g) The multi-mesh is derived from icosahedral meshes of increasing resolution, from the base mesh (𝑀 0 , 12 nodes) to the finest resolution (𝑀 6 , 40, 962 nodes), which has uniform resolution across the globe. It contains the set of nodes from 𝑀 6 , and all the edges from 𝑀 0 to 𝑀 6 . The learned message-passing over the different meshes' edges happens simultaneously, so that each node is updated by all of its incoming edges.
Figure 1 | Model schematic. (a) The input weather state(s) are defined on a 0.25° latitude-longitude grid comprising a total of 721 × 1440 = 1, 038, 240 points. Yellow layers in the closeup pop-out window represent the 5 surface variables, and blue layers represent the 6 atmospheric variables that are repeated at 37 pressure levels (5 + 6 × 37 = 227 variables per point in total), resulting in a state representation of 235, 680, 480 values. (b) GraphCast predicts the next state of the weather on the grid. (c) A forecast is made by iteratively applying GraphCast to each previous predicted state, to produce a sequence of states which represent the weather at successive lead times. (d) The Encoder component of the GraphCast architecture maps local regions of the input (green boxes) into nodes of the multi-mesh graph representation (green, upward arrows which terminate in the green-blue node). (e) The Processor component updates each multi-mesh node using learned message-passing (heavy blue arrows that terminate at a node). (f) The Decoder component maps the processed multi-mesh features (purple nodes) back onto the grid representation (red, downward arrows which terminate at a red box). (g) The multi-mesh is derived from icosahedral meshes of increasing resolution, from the base mesh (𝑀 0 , 12 nodes) to the finest resolution (𝑀 6 , 40, 962 nodes), which has uniform resolution across the globe. It contains the set of nodes from 𝑀 6 , and all the edges from 𝑀 0 to 𝑀 6 . The learned message-passing over the different meshes' edges happens simultaneously, so that each node is updated by all of its incoming edges.CC-BY · From the PDF
Figure 16 | Figure16| Region specification for the regional analysis. We use the same regions and naming convention as in the ECMWF scorecards (https://sites.ecmwf.int/ifs/scorecards/scorecards-47r3HRES. html), and add some additional regions for better coverage of the entire planet. Per-region evaluation is provided in Figure17and Figure18.
Figure 16 | Figure16| Region specification for the regional analysis. We use the same regions and naming convention as in the ECMWF scorecards (https://sites.ecmwf.int/ifs/scorecards/scorecards-47r3HRES. html), and add some additional regions for better coverage of the entire planet. Per-region evaluation is provided in Figure17and Figure18.CC-BY · From the PDF
Figure 20 | Mean bias error for GraphCast at 12 hour lead time, over the 2018 test set.
Figure 20 | Mean bias error for GraphCast at 12 hour lead time, over the 2018 test set.CC-BY · From the PDF
Figure 21 | Mean bias error for GraphCast at 2 day lead time, over the 2018 test set.
Figure 21 | Mean bias error for GraphCast at 2 day lead time, over the 2018 test set.CC-BY · From the PDF
Figure 22 | Mean bias error for GraphCast at 10 day lead time, over the 2018 test set.
Figure 22 | Mean bias error for GraphCast at 10 day lead time, over the 2018 test set.CC-BY · From the PDF
Figure 25 | Normalized RMSE difference of GraphCast relative to HRES, by location, at 12 hours. Blue indicates that GraphCast has greater skill than HRES, Red that HRES has greater skill.
Figure 25 | Normalized RMSE difference of GraphCast relative to HRES, by location, at 12 hours. Blue indicates that GraphCast has greater skill than HRES, Red that HRES has greater skill.CC-BY · From the PDF
Figure 26 | Normalized RMSE difference of GraphCast relative to HRES, by location, at 2 days. Blue indicates that GraphCast has greater skill than HRES, Red that HRES has greater skill.
Figure 26 | Normalized RMSE difference of GraphCast relative to HRES, by location, at 2 days. Blue indicates that GraphCast has greater skill than HRES, Red that HRES has greater skill.CC-BY · From the PDF
Figure 27 | Normalized RMSE difference of GraphCast relative to HRES, by location, at 10 days. Blue indicates that GraphCast has greater skill than HRES, Red that HRES has greater skill. In these 10 day plots we must compare GraphCast at 06z/18z with HRES at 00z/12z (see Section 5). This difference in time-of-day may confound some comparisons, e.g. of 2 t.
Figure 27 | Normalized RMSE difference of GraphCast relative to HRES, by location, at 10 days. Blue indicates that GraphCast has greater skill than HRES, Red that HRES has greater skill. In these 10 day plots we must compare GraphCast at 06z/18z with HRES at 00z/12z (see Section 5). This difference in time-of-day may confound some comparisons, e.g. of 2 t.CC-BY · From the PDF
Figure 48 |Figure 49 | 2 /s 2 z500: Figure 48 | Forecast visualization: 1 0 u. Forecast initialized at 2018-12-22 00:00 UTC, with plots corresponding to 2, 6, and 10 day lead times.
Figure 48 |Figure 49 | 2 /s 2 z500: Figure 48 | Forecast visualization: 1 0 u. Forecast initialized at 2018-12-22 00:00 UTC, with plots corresponding to 2, 6, and 10 day lead times.CC-BY · From the PDF
Image 10
CC-BY · From the PDF
Image 11
CC-BY · From the PDF
Image 12
CC-BY · From the PDF
Image 13
CC-BY · From the PDF
Image 14
CC-BY · From the PDF
Image 15
CC-BY · From the PDF
Image 16
CC-BY · From the PDF
Image 17
CC-BY · From the PDF
Image 18
CC-BY · From the PDF
Image 19
CC-BY · From the PDF
Image 20
CC-BY · From the PDF
Image 21
CC-BY · From the PDF
Image 22
CC-BY · From the PDF
Image 23
CC-BY · From the PDF
Image 24
CC-BY · From the PDF
Image 25
CC-BY · From the PDF
Image 26
CC-BY · From the PDF
Image 27
CC-BY · From the PDF

Verification

Claims the paper makes

9 Valid · 0 Not valid · 18 Not checked · 150 Own results and interpretations

65 more claims

Create a free account to see all 177 claims and to verify the 13 references still open.

See all claims — free account
  • Probably not a substantial claim. 54
  • It may state two separate claims as one. 70
  • The quote only partly states this claim. 2

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

10 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 48 references

Research works printed but not linked: 3

  1. D. H. Levinson, H. J. Diamond, K. R. Knapp, M. C. Kruk, E. J. Gibney, Toward a homogenous global tropical cyclone best-track dataset. Bull. Am. Meteorol. Soc. 91, 377–380 (2010).
  2. ECMWF IFS Documentation CY45R1 – Part II: Data assimilation (European Centre for Medium-Range Weather Forecasts 2018); .10.21957/a3ri44ig4 · doi:10.21957/a3ri44ig4
  3. T. Haiden M. Janousek F. Vitart L. Ferranti F. Prates “Evaluation of ECMWF forecasts including the 2019 upgrade ” ECMWF Technical Memorandum No. 853 (European Centre for Medium-Range Weather Forecasts 2019); .10.21957/mlvapkke · DOI

Cited by

See all 2,423

Metadata

Authors
Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Alet, F., Ravuri, S., Ewalds, T., Eaton-Rosen, Z., Hu, W., Merose, A., Hoyer, S., Holland, G., Vinyals, O., Stott, J., Pritzel, A., Mohamed, S., & Battaglia, P.
Affiliations
  • Google DeepMind, London, UK.
  • Google DeepMind (United Kingdom) (GB)
  • Google (United States) (US)
Corresponding author
  • Willson, Matthew
  • Wirnsberger, Peter
  • Fortunato, Meire
  • Alet, Ferran
  • Ravuri, Suman
  • Mohamed, Shakir
  • Battaglia, Peter
Venue
Science
Published
2023-12-22
Volume
382
Issue
6677
Pages
1416-1421
ISSN
  • 0036-8075 (ISSN-L)
  • 1095-9203
OpenAlex
W4388654737
Cited by
1,466
References
48
Citations by year
2026: 599 · 2025: 522 · 2024: 319 · 2023: 25
FWCI
161.54
Type
journal article
Language
en
Open access
Licence
cc-by
Locations
  • Science (publishedVersion, cc-by)
  • PubMed (publishedVersion)
Topics
  • Physical Sciences › Earth and Planetary Sciences › Atmospheric Science › Meteorological Phenomena and Simulations
  • Physical Sciences › Environmental Science › Environmental Engineering › Hydrological Forecasting Using AI
  • Social Sciences › Transportation › Traffic Prediction and Management Techniques
Keywords
medium-range weather forecasting, GraphCast, reanalysis data, machine learning weather prediction, deterministic forecasting, tropical cyclone tracking, atmospheric rivers, severe weather forecasting

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-11-08, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-08 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Learning skillful medium-range global weather forecasting. https://app.cochat.ai/papers/10.1126/science.adi2336
The paper
Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Alet, F., Ravuri, S., Ewalds, T., Eaton-Rosen, Z., Hu, W., Merose, A., Hoyer, S., Holland, G., Vinyals, O., Stott, J., Pritzel, A., Mohamed, S., & Battaglia, P. (2023). Learning skillful medium-range global weather forecasting. Science. https://doi.org/10.1126/science.adi2336

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.