GraphCast
Last reviewed
Sources
9 citations
Review status
Source-backed
Revision
v2 · 1,440 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
9 citations
Review status
Source-backed
Revision
v2 · 1,440 words
Add missing citations, update stale details, or suggest a clearer explanation.
GraphCast is a machine-learning weather-forecasting model from Google DeepMind that produces a 10-day global forecast at 0.25-degree resolution in under one minute on a single Google Cloud TPU v4 device, and was more accurate than the European Centre for Medium-Range Weather Forecasts (ECMWF) High-Resolution Forecast (HRES), the leading operational deterministic physics model, on 90.3% of 1,380 verification targets.[1][3] Rather than solving the equations of atmospheric physics, it uses a graph neural network trained on roughly 39 years of historical reanalysis data. The work was described in "Learning skillful medium-range global weather forecasting," by Remi Lam and colleagues, published in the journal Science on 14 November 2023.[1][2] DeepMind summarised the system as "a weather forecasting system based on machine learning and Graph Neural Networks (GNNs), which are a particularly useful architecture for processing spatially structured data."[2] GraphCast was a landmark result in AI for science, and was named a runner-up for Science's 2023 Breakthrough of the Year.[1]
Operational weather forecasting has long relied on numerical weather prediction (NWP). These systems discretise the atmosphere into a three-dimensional grid and integrate the governing equations of fluid dynamics and thermodynamics forward in time, starting from an analysis of the current state of the atmosphere assembled from satellite, radiosonde, aircraft, and surface observations.[3] The gold-standard deterministic product in this category is the ECMWF HRES, run on the centre's supercomputers in Reading, England.
NWP is accurate but expensive. A 10-day global forecast at high resolution can require hours of computation across a supercomputer with hundreds or thousands of processor cores.[2][3] The approach also depends on hand-built representations of physical processes that are too small or too complex to resolve directly, such as cloud microphysics and turbulence. GraphCast takes a different route: rather than solving physical equations, it learns the mapping from one atmospheric state to the next directly from decades of historical reanalysis data.[1]
GraphCast is a graph neural network organised in an "encode-process-decode" configuration, with roughly 36.7 million learned parameters.[3][4] The model takes the two most recent states of the atmosphere, each sampled on a 0.25-degree latitude-longitude grid of 721 by 1,440 points (about 1.04 million points), and predicts the state six hours later. Longer forecasts are produced autoregressively, feeding each prediction back as input, so a 10-day forecast is generated as a sequence of forty 6-hour steps.[3][4]
A central design choice is the internal "multi-mesh" on which the network operates. The mesh is built by refining a regular icosahedron six times; the finest mesh contains 40,962 nodes, and the multi-mesh superimposes the edges from every refinement level.[3][4] This structure gives the graph both short-range edges, for local detail, and long-range edges, which let information propagate efficiently across the globe in relatively few message-passing steps. The three components work as follows:
| Component | Role |
|---|---|
| Encoder | Maps input variables from the latitude-longitude grid onto the internal multi-mesh using a single GNN layer |
| Processor | Performs learned message passing on the multi-mesh with 16 unshared GNN layers |
| Decoder | Maps the processed features from the mesh back to the latitude-longitude grid |
Each grid point carries 227 variables: five surface variables and six atmospheric variables, the latter resolved across 37 vertical pressure levels.[2][3][4] GraphCast was trained on 39 years of the ECMWF ERA5 reanalysis, covering 1979 to 2017, and evaluated on held-out data from 2018 onward.[1][3] The published high-resolution model is accompanied by a 1-degree variant and an operational 0.25-degree variant that uses 13 pressure levels.[5]
The paper's principal claim is that GraphCast beats ECMWF HRES on 90.3% of 1,380 verification targets, where a target is a specific combination of variable, vertical level, and forecast lead time.[1][3] In DeepMind's words, "GraphCast provided more accurate predictions on more than 90% of 1380 test variables and forecast lead times."[2] When the comparison is restricted to upper-air (tropospheric) variables most relevant to general forecasting skill, the advantage is larger still, with GraphCast outperforming HRES on the large majority of those targets.[2]
DeepMind also reported that GraphCast supports better prediction of severe events than its training objective alone would suggest, even though it was not specifically trained for them. The authors highlighted improved tracking of tropical cyclone paths, the ability to characterise atmospheric rivers that carry large amounts of water vapour, and earlier signals of extreme heat.[1][2] A widely cited example came during testing: "a live version of our publicly available GraphCast model, deployed on the ECMWF website, accurately predicted about nine days in advance that Hurricane Lee would make landfall in Nova Scotia," several days before traditional forecasts converged on that outcome.[2]
These results carry caveats. GraphCast is trained against and verified largely on ERA5 reanalysis, and tends to produce somewhat blurrier fields than physics-based models, which can understate the intensity of the most extreme, rarest events. Subsequent independent studies have found that physics-based systems can still outperform data-driven models on certain record-breaking extremes.[6]
The most dramatic difference from conventional NWP is speed. Once trained, GraphCast generates a 10-day forecast in under one minute on a single Google TPU v4 machine, compared with the hours of supercomputer time a comparable HRES run demands.[2][3] Training the model was itself substantial, taking about four weeks on 32 TPU v4 devices, but that is a one-time cost; inference afterwards is cheap and fast.[3]
That efficiency makes it practical to run large ensembles, re-forecast quickly as new observations arrive, or deploy forecasting on modest hardware. It does not remove the dependence on NWP infrastructure entirely, because GraphCast still needs an initial analysis of the current atmospheric state, which is itself produced by an observation-driven data-assimilation system.[3]
DeepMind released the GraphCast model code on GitHub, along with pretrained weights for the high-resolution, operational, and small variants.[5] The code is published under the Apache License 2.0, while the model weights are released under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 license.[5] This openness let outside groups reproduce, fine-tune, and benchmark the model.
ECMWF began running experimental, data-driven forecasts including GraphCast and displayed them alongside its own products for comparison and evaluation. The centre treated these external models as a complement to, rather than a replacement for, its physics-based operations, and they helped inspire ECMWF's own machine-learning system.[2][3] In February 2025 ECMWF took its Artificial Intelligence Forecasting System (AIFS), itself built on graph neural networks and transformers, into operational service alongside its physics-based Integrated Forecasting System (IFS), and an ensemble version, AIFS ENS, followed on 1 July 2025.[8] In 2026, as part of the AIFS v2 upgrade that went live on 12 May 2026, ECMWF stopped producing graphical products from the external experimental models GraphCast, Pangu-Weather, Aurora, and FourCastNet, noting that GraphCast had been "superseded by probabilistic models developed by their respective research groups," such as GenCast.[9]
GraphCast was an influential demonstration that a learned model could match or beat the best operational physics-based forecasts across a broad range of variables and lead times, while running orders of magnitude faster. It arrived alongside related efforts such as Huawei's Pangu-Weather and NVIDIA's FourCastNet, and it helped move machine-learning weather prediction from a research curiosity toward an operational tool.
Within DeepMind, GraphCast laid the groundwork for GenCast, a diffusion-based ensemble model published in Nature on 4 December 2024 that produces probabilistic 15-day forecasts. A GenCast forecast comprises an ensemble of 50 or more predictions, takes about 8 minutes to generate on a single Google Cloud TPU v5, and reported beating the ECMWF ENS ensemble on 97.2% of evaluated targets, rising to 99.8% at lead times greater than 36 hours.[7] These models in turn feed into Google's WeatherNext family of forecasting products, which package the research for wider use. Together they mark a shift in numerical meteorology, in which data-driven methods increasingly sit beside, and in some cases ahead of, traditional simulation in the forecasting pipeline.