What happened?
In November 2023, Google DeepMind published GraphCast in the journal Science, a machine learning model that forecasts global weather up to 10 days ahead DeepMind blog ↗. Instead of solving physics equations, it learned patterns from about four decades of past weather data DeepMind blog ↗. In tests it beat a leading conventional forecast on more than 90% of 1,380 checks Research paper ↗.
Normal weather forecasts come from numerical weather prediction: supercomputers solve physics equations for the atmosphere, which can take hours on hundreds of machines DeepMind blog ↗. GraphCast takes the weather now and six hours ago and predicts six hours ahead, then repeats that step to reach 10 days DeepMind blog ↗. It covers the globe in squares about 28 km wide, more than a million points, and a full 10-day forecast takes under a minute on one Google TPU v4 machine DeepMind blog ↗.
The comparison was with HRES, the main high resolution forecast from the European Centre for Medium-Range Weather Forecasts (ECMWF) DeepMind blog ↗. GraphCast was more accurate on more than 90% of 1,380 combinations of weather variables and forecast times, and on 99.7% of the targets in the troposphere, the lowest layer of the atmosphere DeepMind blog ↗. DeepMind also reports that a live version predicted Hurricane Lee's landfall in Nova Scotia about nine days ahead, versus about six days for traditional forecasts DeepMind blog ↗.
What are the three pieces?
The training data
ERA5, ECMWF's reanalysis: a reconstructed record of past weather that blends observations with physics-based models to fill gaps DeepMind blog ↗.
The model
A graph neural network, a type of AI that passes information between connected points, here points spread across the globe DeepMind blog ↗.
The starting point
Each forecast still needs the current state of the atmosphere, which at ECMWF comes from its conventional analysis ECMWF charts ↗.
A model trained on past weather matched or beat a leading physics-based forecast on most tests while using a tiny fraction of the computing time.
Leapscope interpretation of the reported result.How did AI help?
The AI replaced the most expensive step, running physics simulations forward in time, with patterns it learned from decades of data DeepMind blog ↗. Humans built the model, chose the training data and designed the tests against ECMWF's forecast Research paper ↗. ECMWF then ran GraphCast as a live experiment, starting it from its own analysis of current weather ECMWF charts ↗. Peter Dueben of ECMWF, who was not involved, said the models are "so good that we cannot avoid them anymore" MIT Technology Review ↗.
Figures from DeepMind's announcement DeepMind blog ↗ and the research paper Research paper ↗.
GraphCast has real weaknesses. Dueben noted it still trails conventional models in some areas, such as rainfall MIT Technology Review ↗. It gives one forecast without a probability, while forecasters usually run many slightly different forecasts to measure uncertainty Scientific American ↗. Its 28 km squares are too coarse for local downpours and gusts, and because it learns from the past, it may struggle with rare, record-breaking events Scientific American ↗. It also depends on reanalysis data produced by traditional physics models DeepMind blog ↗.
Which fields could this affect?
GraphCast is already part of forecasting research, with wider public benefits possible later; these connections are our assessment.
Weather forecasting research
Weather agencies and researchers can run and compare it, and ECMWF publishes experimental GraphCast forecasts. It pushed the field to take learned models seriously.
Explore scienceOpen science software
The code is public, so others can test and build on it. Google now lists GraphCast alongside newer weather models in the same repository.
Explore softwareEnergy, transport & planning
Faster forecasts could help plan power supply, shipping and flights. That depends on agencies adopting such models in daily services.
Disaster warnings
Earlier storm and heat signals could give people more time to prepare. Single-forecast models like this one do not yet replace official warnings.
What has been checked?
The evidence is a peer reviewed paper in Science, open source code, live experimental forecasts at ECMWF and news coverage with outside meteorologists. Leapscope reviewed these sources; we did not repeat the experiments.
Shown so far
- Better scores than ECMWF's HRES forecast on more than 90% of 1,380 test targets, in a peer reviewed study Research paper ↗.
- ECMWF runs GraphCast as an experimental forecast, starting from its own analysis ECMWF charts ↗.
- The model code is openly available, with trained weights downloadable GitHub repository ↗.
Still unknown
- How well it handles rare extremes such as record rainfall or sudden storm strengthening Scientific American ↗.
- How to give reliable uncertainty estimates from a single deterministic forecast Scientific American ↗.
- How it performs on local details below its 28 km grid, such as city-scale downpours Scientific American ↗.
Evidence status: Published research. Stage: Usable. The model code and trained weights were released publicly.
From research model to daily forecasts
This is our suggested way to follow the story, not a promised timetable.
Can I use it today?
Can I use it today? Researchers can download the code and trained weights from DeepMind's repository GitHub repository ↗, and anyone can view ECMWF's experimental GraphCast charts ECMWF charts ↗. Your phone's weather app is not guaranteed to use it, and official warnings still come from national weather services.
A few things you might be wondering
Does GraphCast replace meteorologists?
No. Experts quoted in coverage see it as one more tool alongside physics models and human forecasters MIT Technology Review ↗ Scientific American ↗.
Is it always more accurate than normal forecasts?
No. It beat HRES on most test targets Research paper ↗, but it trails conventional models in some areas such as rainfall MIT Technology Review ↗.
Why is it so much faster?
It does not solve physics equations at forecast time. It applies patterns learned from past data, so a 10-day forecast takes under a minute on one machine DeepMind blog ↗.
Go straight to the sources
Checked Oct 8, 2026. The first source is the original announcement or research. Later sources add independent context; background pages do not validate the result on their own.
01The developer's announcement with the main results, speed figures and Hurricane Lee example.
The research paper by Lam and colleagues describing the model and its test against HRES.
News coverage with comments from ECMWF, UCLA and MeteoSwiss researchers.
ECMWF's page for its live experimental GraphCast forecasts, initialised from ECMWF analysis.
Explains the limits of AI weather models, including extremes, uncertainty and resolution.
Open code for GraphCast and later Google weather models, with links to trained weights.