---
title: GraphCast (Google DeepMind)
summary: A machine-learned global model published with code and weights; trained on ECMWF reanalysis.
mount: <div class="card"><p class="kicker">Models</p><div class="wx-scroll" tabindex="0"><svg class="mg" viewBox="0 0 994 56" width="994" height="56" role="img" aria-label="GraphCast (Google DeepMind): what it is connected to"><style>.mg text{font:12px var(--sans);fill:var(--ink)}.mg .mg-sub,.mg .mg-verb{font-size:10px;fill:var(--ink-2)}.mg-node rect{fill:var(--surface);stroke:var(--line-2)}.mg-node.me rect{stroke:var(--focus);stroke-width:2}.mg-node.unstated rect{stroke-dasharray:2 3;fill:var(--paper)}a.mg-node:hover rect,a.mg-node:focus rect{stroke:var(--focus);stroke-width:2}.mg-edge{fill:none;stroke:var(--ink-2);stroke-opacity:.55;stroke-width:1}.mg-edge.unv{stroke-dasharray:5 4}</style><path class="mg-edge" d="M232 23C306 23 306 28 380 28"><title>stated on the page: raw.githubusercontent.com</title></path><text class="mg-verb" x="240" y="19" text-anchor="start">published</text><a class="mg-node " href="body-google-deepmind.html"><rect x="2" y="8" width="230" height="30" rx="4"/><text x="9" y="21">Google DeepMind</text><text class="mg-sub" x="9" y="33">Bodies</text></a><path class="mg-edge" d="M610 28C685 28 685 23 760 23"><title>stated on the page: raw.githubusercontent.com</title></path><text class="mg-verb" x="752" y="19" text-anchor="end">is trained on</text><a class="mg-node " href="data-era5.html"><rect x="760" y="8" width="230" height="30" rx="4"/><text x="767" y="21">ERA5 (Copernicus Climate Change Ser…</text><text class="mg-sub" x="767" y="33">Open data and APIs</text></a><g class="mg-node me"><rect x="380" y="13" width="230" height="30" rx="4"/><text x="387" y="26">GraphCast (Google DeepMind)</text><text class="mg-sub" x="387" y="38">Models</text></g></svg></div><p class="wx-legend meta">Left: what points at it. Right: what it points at. Each line carries its verb; a dashed line was not confirmed on the page cited, or the site does not say; the table below gives the page and the time read for each. <span class="prov">From <a href="model-graphcast.json">model-graphcast.json</a></span></p></div>
---
# GraphCast (Google DeepMind)

> Kept by incident.meteorology@web-summit.dinis, 11 November 2026, 11:36 UTC

[Where a forecast comes from](../index.md) · Models · id `model-graphcast` · [as data](model-graphcast.json)

A machine-learned global model published with code and weights; trained on ECMWF reanalysis.

**Address:** [https://github.com/google-deepmind/graphcast](https://github.com/google-deepmind/graphcast) (leaves this site)

## What it is connected to, and who says so

| This | Verb | That | The page that says so | Read | Check |
|---|---|---|---|---|---|
| GraphCast (Google DeepMind) | **is trained on** | [ERA5 (Copernicus Climate Change Service)](data-era5.md) | [raw.githubusercontent.com](https://raw.githubusercontent.com/google-deepmind/graphcast/main/README.md): "the per-model READMEs for model-specific code. ### Training Data Full model training requires downloading the [ERA5] (link) dataset from [ECMWF] (link) best accessed as Zarr via [WeatherBench2] (link)" | 11:28 UTC, 11 Nov | stated on the page |
| GraphCast (Google DeepMind) | **is what Google DeepMind published** | [Google DeepMind](body-google-deepmind.md) | [raw.githubusercontent.com](https://raw.githubusercontent.com/google-deepmind/graphcast/main/README.md): "forecasting model developed by Google DeepMind and Google Research. It also contains code for prior generation models [GraphCast] (link) and [GenCast] (link)" | 11:28 UTC, 11 Nov | stated on the page |

## Licence: what we may do with it

**GraphCast (Google DeepMind)**: Code: Apache 2.0. All other materials, the trained weights included: CC BY 4.0 ([raw.githubusercontent.com](https://raw.githubusercontent.com/google-deepmind/graphcast/main/README.md), read 11:28 UTC, 11 Nov, stated on the page).

| Read | Store | Show | Transform | Attribution |
|---|---|---|---|---|
| yes | yes | yes, with attribution | yes, with attribution | Google DeepMind, with the paper cited. |

The README says the training data (ERA5, HRES) "may be governed by separate terms". Read the weights' licence line before any use: we have not decided "show" or "transform" from it.


---
Not a warning service: this page explains where data comes from. For warnings follow [IPMA](https://www.ipma.pt/pt/otempo/prev-sam/) and ANEPC.