Weather agencies in Europe and North America have started running artificial intelligence models side by side with decades-old physics-based forecasting systems, and in some cases are blending the two. The shift is driven by AI's ability to turn out a global 10-day forecast in under a minute on a single processor, compared with hours of supercomputer time for traditional numerical weather prediction, according to Google DeepMind and the European Centre for Medium-Range Weather Forecasts (ECMWF).
What GraphCast and AIFS actually do
GraphCast, built by Google DeepMind, is a machine-learning model that uses graph neural networks to process weather data across more than a million grid points covering the Earth's surface at a resolution of 28 by 28 kilometres. It takes the atmospheric state from six hours earlier and the current state, then predicts the next six hours; running that step repeatedly produces a 10-day forecast. The model and its results were published in the journal Science on November 14, 2023, with its code released as open source, according to Google DeepMind's own announcement.
ECMWF built its own system, the Artificial Intelligence Forecasting System (AIFS), which became operational on February 25, 2025, running alongside the centre's traditional physics-based Integrated Forecasting System (IFS). AIFS uses the same starting data as IFS — about 60 million quality-controlled observations from satellites, aircraft, buoys and ground stations collected every six hours — but applies a machine-learning model trained on historical weather patterns instead of solving physics equations directly, according to ECMWF.
How accurate are the AI models, according to the organisations that built them
Google DeepMind reported that GraphCast produced more accurate predictions than ECMWF's High Resolution Forecast (HRES) — described by the company as the industry benchmark — on more than 90% of 1,380 test variables and lead times, and on 99.7% of variables specific to the troposphere, the layer of atmosphere where most weather occurs. DeepMind also said GraphCast predicted the Nova Scotia landfall of Hurricane Lee nine days in advance.
ECMWF reported that AIFS "outperforms state-of-the-art physics-based models for many measures, including tropical cyclone tracks, with gains of up to 20%," while using roughly 1,000 times less energy per forecast than conventional methods.
Three AI weather systems compared, by source| System | Built by | Status | Reported performance edge |
|---|
| GraphCast | Google DeepMind | Published in Science, Nov. 2023; open-sourced | More accurate than ECMWF's HRES on over 90% of 1,380 test variables |
|---|
| AIFS | ECMWF | Operational since Feb. 25, 2025 | Up to 20% better on tropical cyclone tracks; ~1,000x less energy per forecast |
|---|
| GEML (hybrid) | Environment and Climate Change Canada | Operational since May 27, 2026 | Built on GraphCast; combines AI with physics via "spectral nudging" |
|---|
Canada's hybrid approach: AI guiding a physics model
Environment and Climate Change Canada took a different route, building a Canadian AI model called the Global Environmental eMuLator (GEML) on top of DeepMind's open-source GraphCast code and training it on Canadian historical weather data. Rather than replacing the physics-based model, the agency combines the two through a technique it calls "spectral nudging": the traditional model is gently guided toward the atmospheric state generated by GEML, while finer-scale details relevant to severe weather still evolve using established physical equations, according to the agency's official announcement. The hybrid system became operational on May 27, 2026, after more than a year of testing.
Environment and Climate Change Canada said AI models can outperform traditional computer models for large-scale atmospheric patterns, which matters most at longer lead times, but did not publish a specific improvement figure in the announcement.
Where AI still falls short
Both ECMWF and Environment and Climate Change Canada flagged limits. ECMWF's AIFS currently runs only as a deterministic single forecast; an ensemble version offering 50 forecast variations, which is standard for assessing forecast uncertainty, is still being developed, according to ECMWF. Environment and Climate Change Canada said that, on their own, current AI models forecast only a limited number of variables, which is not enough for operational use, and that they smooth out finer atmospheric details in ways that can lower accuracy for certain types of extreme weather — one reason the agency chose a hybrid design instead of an AI-only system.
Questions
- What is GraphCast?
- GraphCast is a machine-learning weather model built by Google DeepMind that uses graph neural networks to generate 10-day global forecasts in under a minute, published in the journal Science in November 2023.
- Is AI more accurate than traditional weather forecasting?
- According to the organisations that built them, AI models like GraphCast and ECMWF's AIFS beat physics-based benchmarks on most test variables and lead times, including tropical cyclone tracks, but agencies say physics-based models still handle some extreme-weather details more reliably.
- Is ECMWF's AI forecasting system available to the public?
- Yes. ECMWF runs AIFS alongside its physics-based IFS model, and the centre says AIFS forecast charts are openly available on its website.
- What is Canada's GEML model?
- GEML, the Global Environmental eMuLator, is Environment and Climate Change Canada's AI weather model built on Google DeepMind's open-source GraphCast and trained on Canadian data; it became operational on May 27, 2026, as part of a hybrid system combined with the agency's traditional physics-based model.
- Why don't agencies replace physics-based models entirely with AI?
- Environment and Climate Change Canada and ECMWF both say current AI models smooth out finer atmospheric details important for certain extreme-weather situations, so they combine AI with physics-based modelling rather than using AI alone.
This piece was prepared with artificial intelligence and passed an automated source check before publication. If you see an error, ask for a correction. Request a correction or reply · How we make this
