Machine learning is now a working tool for one of the hardest practical forecasting problems, global weather. ECMWF, the European medium-range forecasting center, made its AI system AIFS fully operational on February 25, 2025, and the U.S. National Oceanic and Atmospheric Administration (NOAA) followed with three AI-driven global models on December 17, 2025. Their results come with conditions that show where AI helps on a complex problem and where it still depends on older methods, starting with GraphCast, a 2023 research model that beat the leading physics-based deterministic system on 90 percent of 1,380 verification targets.
Key takeaways
- Small errors in the starting state grow, so physics-based forecasts need large supercomputers. GraphCast learns from historical reanalysis data and produces a 10-day global forecast at 0.25 degrees in under a minute.
- NOAA ran its own AI global models operationally from December 17, 2025, and ECMWF moved AIFS to version 2 on May 12, 2026.
- ECMWF reports AIFS gains of up to 20 percent on some measures and about 1,000 times less energy per forecast.
- AI weather models still start from physics-based estimates of the current state from observations, and their grids are coarser than the best physics model.
- Most headline accuracy figures come from the developers' own evaluations, so independent checking over time matters.
Why weather counts as a complex problem
The atmosphere is a fluid governed by known physical laws, but those laws cannot be solved exactly for the whole planet. Traditional numerical weather prediction (NWP) divides the globe into a grid, estimates the present state from observations, and steps the equations forward in time. Accuracy improves mostly by adding computing power, which the GraphCast authors point out in their paper. The method also cannot learn directly from decades of past weather records.
The difficulty is not only size. Forecasts are sensitive to the starting conditions, so a forecast is useful only for a limited number of days, and forecasters need a sense of uncertainty as well as a single best guess. Any AI approach has to deal with both: be fast enough to be useful and honest about what it does not know.
What GraphCast did
GraphCast, described in a paper first posted on arXiv on December 24, 2022 and published in Science in 2023, is a machine learning model trained directly on reanalysis data, a historical reconstruction of the atmosphere. According to the paper's abstract, it predicts hundreds of weather variables up to 10 days ahead at 0.25 degree resolution worldwide in under one minute. The authors report that it outperformed the most accurate operational deterministic system on 90 percent of 1,380 verification targets and that it better predicted severe events, including tropical cyclones, atmospheric rivers, and extreme temperatures.
For Canadian readers, one reported example is Hurricane Lee in September 2023. The developers state that, in a live experiment run by ECMWF, GraphCast predicted the storm's landfall in Nova Scotia about nine days ahead, while traditional forecasts settled on Nova Scotia about six days out. That figure is the developers' own account of a single storm, so it illustrates the approach and does not measure average performance.
GraphCast was trained on ERA5 reanalysis, which itself is produced with traditional NWP to fill gaps in observations. The learning-based model therefore inherits information from the physics-based system it competes with.
From research model to operations
On February 25, 2025, ECMWF made AIFS operational, describing it as the first fully operational open weather prediction model using machine learning with the widest range of parameters. It runs next to the physics-based Integrated Forecasting System (IFS), and the two are presented as complementary. ECMWF reports gains of up to 20 percent over state-of-the-art physics-based models on many measures, including tropical cyclone tracks, and a reduction of about 1,000 times in the energy needed for a forecast.
A follow-up came on July 1, 2025 with AIFS ENS, an ensemble of 51 forecasts at 31 kilometers resolution, intended to express uncertainty. ECMWF says it uses physics-based data assimilation for its initial conditions and is lower in resolution than the physics-based ensemble. A September 2025 paper from ECMWF describes an update to the single-forecast model that adds physical consistency constraints through bounding layers, which it says substantially improve precipitation forecasts.
What changed in 2025 and 2026
Two later steps matter for anyone reading the 2025 numbers. On December 17, 2025, NOAA made three global models operational: AIGFS, AIGEFS, and HGEFS. NOAA says it started from Google DeepMind's GraphCast and fine-tuned it on its own Global Data Assimilation System data. By NOAA's account, AIGFS uses up to 99.7 percent less computing than the traditional GFS and cut tropical cyclone track errors at longer lead times, though version 1.0 is worse on cyclone intensity. The AI ensemble AIGEFS has 31 members and skill comparable to the physics-based GEFS at 9 percent of the computing, and the hybrid HGEFS pairs 31 physics-based members with 31 AI members (62 in total) and, per NOAA, beat both parent ensembles on most major metrics.
On May 12, 2026, ECMWF put AIFS Single v2 and AIFS ENS v2 into service together with a new physics model cycle, IFS Cycle 50r1. Both AI versions add a wave component with 11 variables, ECMWF's first operational data-driven wave forecasts. AIFS Single v2 also adds snow cover forecasts. An ECMWF newsletter article reports that significant wave height errors in the medium range are about 10 percent lower than in IFS Cycle 50r1 and that skill in the lower stratosphere gained about three days over AIFS Single v1. The ensemble now uses a new multi-scale training loss. These are ECMWF's own verification results, and the physics upgrade shipped on the same day, so credit for any single gain is easy to misplace.
Where AI helps and where it leans on older methods
The comparison below summarizes the documented trade-offs. It is a simplified reading of the sources.
| Aspect | Physics-based NWP | AI forecast model |
|---|---|---|
| Source of skill | Equations of atmospheric physics | Patterns learned from past weather |
| Speed per forecast | Needs a large supercomputer | GraphCast: under a minute for a 10-day global run |
| Energy per forecast | Baseline | ECMWF reports about 1,000 times less for AIFS |
| Starting state | Built by assimilating observations | AIFS uses the same initial conditions as the IFS |
| Grid spacing at ECMWF | 9 kilometers (IFS) | 28 kilometers (AIFS Single), 31 kilometers (AIFS ENS), version 1 figures; v2 grids not stated in the sources read |
| Explanation of results | Physical processes can be traced | Learned, harder to inspect |
Reading AI accuracy claims
- Check what the model is compared with and on which variables; "better on most targets" does not mean better on every one.
- Look for the verification period and whether the test data were kept out of training.
- Separate a research result from an operational service that runs every day.
- Treat single-event success stories as illustrations.
- Watch for updates: the AIFS single-forecast model changed in 2025 to improve precipitation and moved to version 2 in May 2026, and NOAA flags weaker cyclone intensity in AIGFS version 1.0. These systems are revised as weaknesses appear.
The questions above are general, drawn from the cases in this article, and not an official standard. A conclusion that AI replaces physics-based forecasting is not supported by the sources: ECMWF runs both and states they complement each other.
FAQ
Is AI weather forecasting already in use?
Yes. ECMWF's AIFS has been fully operational since February 25, 2025, alongside its physics-based model, and NOAA has run AI global models operationally since December 17, 2025.
Does an AI model understand atmospheric physics?
It learns statistical patterns from historical data and does not solve physical equations directly. Newer ECMWF work adds physical consistency constraints to its outputs, which shows researchers are working to build physical knowledge into such models.
What data did GraphCast learn from?
Reanalysis data, a reconstruction of past atmospheric conditions. The authors describe training directly on it, and the developers note that ERA5 reanalysis itself relies on traditional forecasting methods to fill gaps.
Why does ECMWF run an ensemble?
A single forecast hides uncertainty. An ensemble of slightly varied forecasts shows the range of plausible outcomes, and AIFS ENS provides 51 of them.
Related reading
- How AI approaches problem solving
- Challenges in AI research
- What makes AI problems hard
- AI and efficiency gains
More in AI Basics.
Last verified: October 2026
Sources
- Lam R et al., GraphCast: Learning skillful medium-range global weather forecasting, arXiv:2212.12794 and Science, 2023
- ECMWF's AI forecasts become operational, ECMWF, February 25, 2025
- ECMWF press release on the AIFS ensemble model, July 1, 2025
- An update to ECMWF's machine-learned weather forecast model AIFS, arXiv:2509.18994, September 2025
- NOAA deploys new generation of AI-driven global weather models, National Oceanic and Atmospheric Administration, December 17, 2025
- Significant update to ECMWF's key forecasting systems IFS and AIFS goes live, ECMWF, May 12, 2026
- Implementation of AIFS v2, ECMWF Newsletter 187, 2026
- GraphCast: AI model for faster and more accurate global weather forecasting, Google DeepMind, November 14, 2023