How AI Weather Forecasting Works, and What Forecasters Still Do
Since 2025 two of the world's big forecasting centers, ECMWF in Europe and NOAA in the United States, have run AI models every day beside their physics-based ones. Here is how those models are built and tested, what they cost to run, where they fall short, and the steps still done by physics and by people.
Checked against ECMWF's AIFS v2 pages (upgrade of May 12, 2026), NOAA's December 2025 release on AIGFS, AIGEFS and HGEFS, and the GraphCast and GenCast papers
Role scorecard
Atmospheric and Space Scientists SOC 19-2021.00
- Core tasks AI does in production0 of 3
- Most mature evidenceIn production
- CostOutput free and open under CC-BY-4.0; operating costs not disclosed (ECMWF)
- BenefitOver 10 times faster, about 1,000 times less energy per forecast (ECMWF-stated)
As of · Tasks: O*NET 31.01
| Core task (O*NET) | What AI or robots do | Extent | Evidence grade | Example case | Source |
|---|---|---|---|---|---|
| Develop or use mathematical or computer models for weather forecasting. | AI models generate global forecasts every run alongside the physics-based ones; people choose, tune and verify the models | Assists | In production | ECMWF AIFS Single (operational since February 25, 2025) and AIFS ENS (July 1, 2025); NOAA AIGFS, AIGEFS and HGEFS (announced December 2025) | 2, 3, 5 |
| Interpret data, reports, maps, photographs, or charts to predict long- or short-range weather conditions, using computer models and knowledge of climate theory, physics, and mathematics. | The models supply faster guidance for forecasters and national weather services to interpret; no source says an AI model issues a forecast or warning on its own | Assists | In production | NOAA: "faster delivery of forecast products to meteorologists and the public"; ECMWF: users "decide what best suits their needs" | 3, 5 |
| Analyze climate data sets, using techniques such as geophysical fluid dynamics, data assimilation, or numerical modeling. | The operational AI models still start from physics-based data assimilation; ECMWF has introduced forecasting directly from observations with machine learning, and we found no statement that it is operational | Not yet | Prototype | ECMWF AI-DOP (direct observation prediction), research presentation, February 2026 | 3, 6 |
Extent and grade follow our evidence grades. Both operational rows are "Assists" because the centers describe the AI as guidance for forecasters, so no task is counted as done by AI alone. Grades come from the centers' own statements. We show three of the 27 core tasks O*NET lists for the occupation; the rest (balloon launches, teaching, climate research and similar) have no AI case in the sources we opened.1
Costs and benefits
| Data access | AIFS output is released under CC-BY-4.0 and ECMWF's terms; real-time open data are free (ECMWF-stated)4 |
|---|---|
| Speed and energy | AIFS ENS makes forecasts over 10 times faster than the physics-based system, with approximately 1,000 times less energy (ECMWF-stated); NOAA says one 16-day AIGFS forecast uses 0.3% of the computing resources of its operational GFS and takes about 40 minutes (NOAA-stated)35 |
| Accuracy | Gains of up to 20% on some measures versus physics-based models, with limits (see below); center-stated, not independent235 |
As of · Sources: ECMWF, NOAA (as reproduced), and the GraphCast and GenCast papers235
| Model | Operator | Type | Evidence grade | Status | Members | What is stated | Source |
|---|---|---|---|---|---|---|---|
| AIFS Single | ECMWF | AI, single forecast | In production | Operational since Feb 25, 2025; v2 since May 12, 2026 | 1 | Initial conditions from physics-based analysis; 0.25° output to 15 days, four runs a day | 2, 4 |
| AIFS ENS | ECMWF | AI, ensemble | In production | Operational since Jul 1, 2025; v2 since May 12, 2026 | 51 | 31 km (v1) against 9 km for the physics-based ensemble; ECMWF says gains of up to 20% on measures such as surface temperature | 3, 4 |
| AIGFS | NOAA | AI, single forecast | In production | Deployed December 2025 (v1.0) | 1 | One 16-day run in about 40 minutes; v1.0 degrades tropical-cyclone intensity forecasts, track errors reduced at longer lead times | 5 |
| AIGEFS | NOAA | AI, ensemble | In production | Deployed December 2025 | 31 | Skill comparable to the operational GEFS on 9% of its computing; NOAA's early-results release also says it extends skill 18 to 24 hours | 5 |
| HGEFS | NOAA | Hybrid grand ensemble | In production | Deployed December 2025 | 62 | 31 physics-based GEFS plus 31 AIGEFS members; NOAA says it beats both across most major verification metrics | 5 |
| GraphCast | Google DeepMind | AI, single forecast | Prototype | Research paper (arXiv, 2022, revised 2023); NOAA used it as the foundation for AIGFS | 1 | Paper reports better than the best operational deterministic system on 90% of 1,380 verification targets; 10-day forecast in under a minute | 7, 8 |
| GenCast | Google DeepMind | AI, ensemble | Prototype | Research paper (arXiv, 2023, revised 2024) | Ensemble | Paper reports greater skill than ECMWF's ENS on 97.4% of 1,320 targets; 15-day ensemble in 8 minutes | 9 |
Every performance claim in the last column is made by the operator or the model's authors and measured by them on their own tests, so the figures do not compare across rows: the centers and the papers use different verification targets, baselines and lead times. "In production" means the center runs the model operationally; "Prototype" is a published research model, although NOAA says it fine-tuned GraphCast to build AIGFS.578 GenCast's ensemble size is not stated in the abstract we opened.9 Resolution and member counts are for the versions named; AIFS was upgraded to v2 in May 2026 and v2's grid spacing is not given in the pages we opened.4
How it works
The traditional chain, and where the AI fits
Traditional forecasting runs observations through data assimilation into a physics-based model and then post-processing and dissemination.6 ECMWF describes its starting conditions as a previous short-term forecast combined with around 60 million quality-controlled observations from satellites, aircraft, ships, buoys and ground stations, which feed its AI model every six hours.2 The AI model is trained on how the weather has evolved in the past and works out how today's starting conditions will influence the next days; ECMWF's physics-based system integrates the laws of physics in its code.2 The AIFS paper describes a model built from a graph neural network encoder and decoder with a sliding-window transformer processor, trained on ECMWF's ERA5 reanalysis and its operational analyses.7
Ensembles: many forecasts instead of one
A single forecast hides uncertainty, so centers run ensembles. ECMWF's AIFS ENS is 51 forecasts with slight variations, and NOAA's AIGEFS is 31; NOAA's HGEFS merges the 31 AI members with the 31 members of its physics-based GEFS into a 62-member grand ensemble, which NOAA says it believes is the first such hybrid at an operational center.35
What it costs to run
This is where the centers report the clearest gains: ECMWF says AIFS ENS makes forecasts over 10 times faster than its physics-based system with approximately 1,000 times less energy.3 NOAA says one 16-day AIGFS forecast uses 0.3% of the computing resources of its operational GFS and finishes in about 40 minutes, and its AI ensemble needs 9% of the resources of the GEFS.5 Training the models is a separate cost that neither center's release quantifies.
How accurate it is, and where it isn't
ECMWF says AIFS Single outperforms state-of-the-art physics-based models on many measures, including tropical cyclone tracks, with gains of up to 20%, and that AIFS ENS gains up to 20% on measures such as surface temperature.23 The limits are stated too: AIFS ENS works at 31 km against 9 km for the physics-based ensemble, which ECMWF says remains indispensable for high-resolution fields and coupled Earth-system processes, and NOAA says its AIGFS v1.0 shows a degradation in tropical-cyclone intensity forecasts even though track forecasts are better.35 Google DeepMind's papers report that GraphCast beat the best operational deterministic system on 90% of 1,380 verification targets and that GenCast beat ECMWF's ensemble on 97.4% of 1,320 targets; those are the authors' own tests of research models, not operational results.89
What forecasters still do
The AI models are guidance, not forecasters. Both centers run them beside their physics-based systems rather than replacing them, ECMWF calls the two complementary and says its users decide which suits their needs, and NOAA says the models give meteorologists faster delivery of guidance.35 The step that turns observations into a starting state is still physics-based, and ECMWF is exploring hybrids of the two approaches.3 ECMWF's own February 2026 presentation introduces "AI-DOP", forecasting directly from observations with machine learning, and we found no statement that it is operational.6 We found no source in which an AI model issues a public forecast or severe-weather warning without a person; how national services use the guidance is for each service to say.
For how MetaRoles grades cases like these, see how we measure how much of a role AI does.
Sources
- O*NET OnLine (U.S. Department of Labor), Summary Report for 19-2021.00, Atmospheric and Space Scientists (O*NET 31.0 tasks), retrieved 2026-10-22.
- European Centre for Medium-Range Weather Forecasts, ECMWF's AI forecasts become operational, February 25, 2025.
- European Centre for Medium-Range Weather Forecasts, ECMWF's ensemble AI forecasts become operational, July 1, 2025.
- European Centre for Medium-Range Weather Forecasts, AIFS Machine Learning data (AIFS v2 pages), retrieved 2026-10-22.
- WeatherNation (reproducing the NOAA release), NOAA Deploys New Generation Of AI-Driven Global Weather Models, December 18, 2025 (NOAA's December 17, 2025 release, a secondary copy because noaa.gov was not reachable; the World Meteorological Organization carries the same release and quote at wmo.int).
- E. Pinnington, ECMWF, Adding Land Variables to the AIFS and introduction to Direct Observation Prediction (AI-DOP), presentation, February 12, 2026.
- ECMWF, AIFS – ECMWF's data-driven forecasting system, arXiv:2406.01465.
- Lam et al. (Google DeepMind), GraphCast: Learning skillful medium-range global weather forecasting, arXiv:2212.12794 (v2, August 2023).
- Price et al. (Google DeepMind), GenCast: Diffusion-based ensemble forecasting for medium-range weather, arXiv:2312.15796 (v2, May 2024).