AI weather forecasting

Also known as Machine-learning weather prediction, Data-driven weather forecasting, ML weather models, GraphCast, GenCast, AIFS

Machine-learning weather models such as GraphCast, GenCast and ECMWF's AIFS produce global forecasts in minutes. In published and operational tests they match or beat leading physics-based systems on many headline measures, but they still struggle with small-scale detail, the early growth of forecast error and physical consistency.

Pilot and early commercialDemonstratedReviewed 21 Oct 2026
AI illustration · Illustrative scene, not a photograph of the specific device or experiment.

In plain language

Weather forecasts have long been made with physics-based models. These write down equations for air, water and heat, feed in the latest observations, and step the atmosphere forward in time on large supercomputers. A newer approach trains a neural network on past weather records so that it learns how the atmosphere usually changes from one moment to the next. Once trained, the network can produce a global forecast in a fraction of the time. The GraphCast authors report under one minute for a 10-day forecast.

Several research groups report that these models can match or beat leading physics-based forecasts on many tests, including tracking tropical cyclones and predicting extreme temperatures. They are not simply replacing physics, though. Critics point out that the models struggle to reproduce small-scale weather features and do not always behave like the real atmosphere. In February 2025 the European Centre for Medium-Range Weather Forecasts (ECMWF) put one of these models, AIFS, into operations alongside its physics-based system. That is a real-world test, not just a lab result.

Technical explanation

GraphCast is a machine-learning model trained on reanalysis data. It forecasts hundreds of weather variables globally for 10 days at 0.25 degree resolution, and the authors report that a forecast takes under one minute (Lam and colleagues, 2023). Pangu-Weather is a three-dimensional deep neural network for medium-range global forecasting (Bi and colleagues, 2023). GenCast is a probabilistic ensemble model. It generates 15-day forecasts in 12-hour steps, giving a range of possible outcomes rather than one answer (Price and colleagues, 2024). ECMWF's AIFS Single is the deterministic model it runs in operations, alongside the physics-based IFS (ECMWF, 2025).

WeatherBench 2 provides an open evaluation framework, publicly available training, ground-truth and baseline data, and a continuously updated results site, so that data-driven and physics-based models can be compared on the same basis for lead times of 1 to 14 days (Rasp and colleagues, 2024).

The physical limits are subtle. Bonavita (2024) examined Pangu-Weather, FourCastNet and GraphCast and concluded that they struggle with smaller-scale weather features and lack the fidelity and physical consistency of physics-based models. Selz and Craig (2023) tested whether an AI-based model reproduces the rapid early growth of very small initial errors. Physics-based models predict that growth, and it sets a fundamental limit on predictability. The AI model did not reproduce that early growth, which would wrongly imply unlimited predictability. With larger initial perturbations, comparable to current observational uncertainty, the AI model broadly agreed with physics-based simulations, though some deficits remain.

What the evidence says

Each statement is typed by the strength of its evidence and linked to the source that supports it.

Demonstrated Shown in a published experiment, test or shipping product.

GraphCast, trained on reanalysis data, predicts hundreds of weather variables globally for 10 days at 0.25 degree resolution, and the authors report that a forecast takes under one minute.

Sources: [1] · Abstract

In the authors' evaluation, GraphCast outperformed the most accurate operational deterministic systems on 90% of 1,380 verification targets, with better prediction of severe events including tropical cyclones.

Sources: [1] · Abstract · medium confidence

In the authors' evaluation, GenCast showed greater skill than ECMWF's ENS ensemble on 97.2% of 1,320 targets, and it better predicted extreme weather, tropical cyclone tracks and wind power production.

Sources: [3] · Abstract · medium confidence

Pangu-Weather produced stronger deterministic forecasts than ECMWF's operational IFS on reanalysis data in all tested variables, and it tracked tropical cyclones more accurately when initialised from reanalysis data.

Sources: [2] · Abstract · medium confidence

ECMWF put AIFS Single into operations on 25 February 2025, running alongside its physics-based IFS, and describes AIFS as the first fully operational open weather prediction model using machine learning.

Sources: [4] · Announcement text

Selz and Craig found that an AI-based weather model does not reproduce the rapid early growth of very small initial perturbations, which would wrongly imply unlimited predictability, and that with larger perturbations it broadly agrees with physics-based simulations, though some deficits remain.

Sources: [6] · Abstract · medium confidence

WeatherBench 2 provides an open evaluation framework, publicly available training, ground-truth and baseline data, and a continuously updated results site for comparing physical and data-driven global models at lead times of 1 to 14 days, and its authors discuss caveats in the current evaluation setup.

Sources: [7] · Abstract

Reported, unverified Stated by a company or team but not independently confirmed.

ECMWF reports that AIFS outperforms state-of-the-art physics-based models on many measures, including tropical cyclone tracks, with gains of up to 20%, and that it uses approximately 1,000 times less energy to make a forecast. These are ECMWF's own figures, and the announcement gives no baseline or conditions for them.

Sources: [4] · Performance claims · low confidence

Disputed Contested by other credible sources.

Bonavita argues that Pangu-Weather, FourCastNet and GraphCast struggle with smaller-scale weather features and fall short of physics-based models on faithfulness and physical coherence, so their forecasts need careful interpretation.

Sources: [5] · Abstract · medium confidence

Limitations and open problems

  • The headline comparisons cited here are the developers' own evaluations on chosen verification targets. WeatherBench 2 was built to make comparisons more open, and its authors discuss caveats in the current evaluation setup (Rasp and colleagues, 2024).
  • Bonavita (2024) reports that leading models struggle with smaller-scale weather features and fall short of physics-based models on faithfulness and physical coherence.
  • Selz and Craig (2023) found that an AI-based model does not reproduce the rapid early growth of very small initial errors, so it can overstate how predictable the atmosphere is.
  • ECMWF's energy and accuracy figures for AIFS are its own. The announcement gives an approximately 1,000-fold reduction in energy use for making a forecast and gains of up to 20% on many measures, with no baseline, time period or verification method given.

Possible future uses

Speculative

Ideas, not predictions. None of these has been demonstrated.

  • Probabilistic machine-learning forecasts feeding planning tools for wind and solar power output, where probabilistic forecasts of wind power production matter.
  • Forecasts for regions with thin observation networks, where a fast, low-cost model could fill gaps that physics-based systems struggle to cover.
  • Hybrid systems that combine a learned model with a physics-based model, which ECMWF describes as a field of research over the coming years.

Timeline

  1. 5 Jul 2023Pangu-Weather is published in Nature, reporting stronger deterministic results than ECMWF's operational IFS on reanalysis data for all tested variables. Sources: [2]
  2. 26 Oct 2023Selz and Craig publish a test showing that an AI-based model does not reproduce the rapid early growth of very small initial errors. Sources: [6]
  3. 14 Nov 2023GraphCast is published online in Science, reporting better results than the most accurate operational deterministic systems on 90% of 1,380 verification targets in the authors' evaluation. Sources: [1]
  4. 14 Jun 2024Bonavita publishes a critique in Geophysical Research Letters of the faithfulness and physical coherence of Pangu-Weather, FourCastNet and GraphCast. Sources: [5]
  5. 4 Dec 2024GenCast is published online in Nature, reporting greater skill than ECMWF's ENS ensemble on 97.2% of 1,320 targets in the authors' evaluation. Sources: [3]
  6. 25 Feb 2025ECMWF puts AIFS Single into operations, running alongside its physics-based IFS. Sources: [4]

Futures this could help bring about

AI illustration
How far away?

A full digital twin of the Earth

Pilots · 2 of 8 steps done

AI illustration
How far away?

AI weather forecasts replacing physics-based models

Pilots · 4 of 7 steps done

Related breakthroughs

Glossary

Numerical weather prediction
Forecasting that solves the equations of physics for the atmosphere on a computer grid, starting from the latest observations.
Reanalysis
A long, consistent record of past weather built by combining a physics-based model with historical observations, used to train and test forecast models.
Ensemble forecast
A set of forecasts run from slightly different starting conditions or with different model settings, giving a range of possible outcomes and their likelihoods.
Lead time
How far ahead a forecast predicts, for example one day or 10 days.

Sources

  1. peer reviewed paper
    Learning skillful medium-range global weather forecasting
    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson et al.. Science, 14 Nov 2023. DOI: 10.1126/science.adi2336. Accessed 21 Oct 2026.
  2. peer reviewed papercorrected
    Accurate medium-range global weather forecasting with 3D neural networks
    Kaifeng Bi, Ling-Xi Xie, Heng-Heng Zhang et al.. Nature, 5 Jul 2023. DOI: 10.1038/s41586-023-06185-3. Accessed 21 Oct 2026.
  3. peer reviewed paper
    Probabilistic weather forecasting with machine learning
    Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet et al.. Nature, 4 Dec 2024. DOI: 10.1038/s41586-024-08252-9. Accessed 21 Oct 2026.
  4. other
    ECMWF's AI forecasts become operational
    ECMWF, 25 Feb 2025. Accessed 21 Oct 2026.
  5. peer reviewed paper
    On Some Limitations of Current Machine Learning Weather Prediction Models
    Massimo Bonavita. Geophysical Research Letters, 14 Jun 2024. DOI: 10.1029/2023GL107377. Accessed 21 Oct 2026.
  6. peer reviewed paper
    Can Artificial Intelligence-Based Weather Prediction Models Simulate the Butterfly Effect?
    T. Selz, G. C. Craig. Geophysical Research Letters, 26 Oct 2023. DOI: 10.1029/2023GL105747. Accessed 21 Oct 2026.
  7. peer reviewed paper
    WeatherBench 2: A Benchmark for the Next Generation of Data-Driven Global Weather Models
    S. Rasp, Stephan Hoyer, Alexander Merose et al.. Journal of Advances in Modeling Earth Systems, 18 Jun 2024. DOI: 10.1029/2023MS004019. Accessed 21 Oct 2026.

Revision history

3 changes since this profile was created.

Changes on 21 Oct 2026
Published on owner instruction (content loop iteration 6)2124aeb
Sources and claims re-checked against originalsfbb434c
Initial profile drafted from primary sources7e66249