Skip to content
ContentLora

    Tip: press / anywhere to search.

    Explainer

    How AI weather models work

    AI weather models learn how the atmosphere evolves from decades of reanalysis data instead of solving physics equations at every step.[1] They forecast in minutes on a single chip, and since 2025 ECMWF and NOAA have run them operationally alongside physics-based models.[2][3][4]

    Editor reviewedUpdated AI for scienceArtificial intelligenceScience

    Weather forecasting was one of the first sciences where AI went from research paper to daily operations. Between 2023 and 2025, machine-learning models matched or beat the best physics-based forecasts on many measures, and the European and US weather agencies put them into service.[5][3][4]

    The traditional way

    For decades, forecasts have come from huge computer programs that apply the laws of physics to the atmosphere, step by step, on supercomputers.[6] They work well but are slow and expensive to run.

    Numerical weather prediction (NWP) discretises the governing equations and integrates them forward on supercomputers.[6] A 10-day deterministic run of ECMWF’s high-resolution HRES can take hours on hundreds of machines.[2]

    The AI way

    An AI model instead studies about 40 years of past weather and learns what usually happens next. Give it the weather now and six hours ago, and it predicts six hours ahead, then repeats that step to reach 10 days.[1][7] Once trained, it needs only a single chip and under a minute.[2]

    GraphCast, a graph neural network, was trained on four decades of ERA5 reanalysis and rolls out autoregressively in 6-hour steps from two input states.[1][7] It beat HRES on more than 90% of 1,380 verification targets.[5] Because ERA5 is produced by data assimilation with an NWP model, these systems still depend on physics-based infrastructure for training data.[1]

    From one forecast to many

    Weather is uncertain, so forecasters run ensembles: many slightly different forecasts that show the range of outcomes. Early ML models produced a single forecast and remained less reliable than physics-based ensembles.[8]

    Newer AI models make dozens or hundreds of forecasts at once, so forecasters can see the chance of a storm rather than one guess.[9]

    GenCast, a probabilistic ML model, produces 15-day ensembles at 0.25° in 8 minutes and beat ECMWF’s ENS on 97.2% of 1,320 targets.[10][11] WeatherNext 2 uses a Functional Generative Network that injects noise into the architecture, producing each ensemble member in under a minute on a TPU.[12][9]

    Into operations

    ECMWF made its Artificial Intelligence Forecasting System operational on 25 February 2025, alongside its physics-based IFS, and added an ensemble version on 1 July 2025.[3][13] ECMWF says the AIFS improves tropical cyclone tracks by up to 20% and uses about 1,000 times less energy per forecast.[14] NOAA followed on 17 December 2025 with AIGFS, AIGEFS and a hybrid AI-physics ensemble, HGEFS.[4]

    The AIFS pairs a graph neural network encoder and decoder with a sliding-window transformer processor, trained on ERA5 and ECMWF’s operational analyses.[15] It runs four times a day, its forecasts are open data, and the weights of AIFS Single v1.0 are published under CC BY 4.0.[16][17]

    Storms and extremes

    Hurricanes are hard to forecast because small differences in the atmosphere can change their path and strength.[18] Since June 2025 Google’s Weather Lab has run an experimental AI cyclone model that draws 50 possible storm paths up to 15 days ahead, and US National Hurricane Center forecasters can see its predictions next to traditional models.[19] Google says it is a research tool, not an official warning.[20]

    Google reports that internal testing put the cyclone model’s track and intensity errors on par with or better than physics-based methods.[21] Intensity is a known weak spot elsewhere: ECMWF’s model card for the AIFS lists blurred fields at longer lead times, weaker skill in the stratosphere and reduced intensity of some high-impact systems such as tropical cyclones.[22]

    What to keep in mind

    Most comparisons are published by the model developers and use their own scoring choices, and the agencies still run physics-based models in parallel.[3] NOAA’s system explicitly combines AI and physics-based approaches.[23] Details of each system are on the weathernext and ecmwf-aifs pages.

    Questions readers ask

    Do AI weather models replace physics?

    Not fully. They are trained on ERA5 reanalysis, which is built using a physics-based model, and ECMWF runs its AI system side by side with its physics-based IFS.[1][3]

    How fast are they?

    GraphCast makes a 10-day forecast in under a minute on one TPU v4 machine; a conventional HRES forecast can take hours on a supercomputer.[2]

    Can AI models give probabilities, not just one forecast?

    Yes. GenCast and WeatherNext 2 generate ensembles of possible outcomes, and ECMWF made its ensemble AI system operational in July 2025.[10][9][13]

    Which weather agencies use AI models operationally?

    ECMWF has run its AIFS operationally since February 2025, and NOAA deployed AI-driven global models in December 2025.[3][4]

    Sources

    Each numbered claim is a statement we checked against the sources listed with it. Status shows how well established it is.

    1. [1]

      GraphCast is a graph neural network trained on four decades of ECMWF's ERA5 reanalysis data, which is itself built from observations using traditional numerical weather prediction. confirmedas of 2023-11-14

    2. [2]

      GraphCast makes a 10-day forecast in less than a minute on a single Google TPU v4 machine, while a conventional HRES forecast can take hours on a supercomputer with hundreds of machines. confirmedas of 2023-11-14

    3. [3]

      ECMWF took its Artificial Intelligence Forecasting System (AIFS) into operations on 25 February 2025, running side by side with its physics-based IFS. confirmedas of 2025-02-25

    4. [4]

      On 17 December 2025 NOAA deployed operational AI-driven global models, AIGFS, AIGEFS and the hybrid HGEFS, which NOAA says deliver more accurate guidance using a fraction of the computing resources. confirmedas of 2025-12-17

    5. [5]

      Google DeepMind's GraphCast, described in Science in November 2023, gave more accurate predictions than ECMWF's HRES forecast on more than 90% of 1,380 test variables and lead times. confirmedas of 2023-11-14

    6. [6]

      Traditional numerical weather prediction translates physics equations into algorithms run on supercomputers. confirmedas of 2023-11-14

    7. [7]

      GraphCast takes the weather state from six hours ago and the current state, predicts six hours ahead, and repeats the step to forecast up to 10 days. confirmedas of 2023-11-14

    8. [8]

      The GenCast paper notes that earlier ML weather models focused on single deterministic forecasts and remained less accurate and reliable than the best physics-based ensemble forecasts. confirmedas of 2024-12-04

    9. [9]

      WeatherNext 2 can produce hundreds of possible weather outcomes from one starting point, each in under a minute on a single TPU. confirmedas of 2025-11-17

    10. [10]

      GenCast generates an ensemble of 15-day global forecasts at 0.25° resolution for more than 80 variables in 8 minutes. confirmedas of 2024-12-04

    11. [11]

      GenCast, a probabilistic machine-learning weather model published in Nature in December 2024, had greater skill than ECMWF's ENS ensemble on 97.2% of 1,320 evaluated targets. confirmedas of 2024-12-04

    12. [12]

      Google DeepMind's WeatherNext 2, announced in November 2025, uses a Functional Generative Network and surpasses the previous WeatherNext model on 99.9% of variables and lead times from 0 to 15 days. confirmedas of 2025-11-17

    13. [13]

      ECMWF took the ensemble version of the AIFS into operations on 1 July 2025. confirmedas of 2025-07-01

    14. [14]

      ECMWF says the AIFS outperforms physics-based models on many measures, including tropical cyclone tracks with gains of up to 20%, and uses about 1,000 times less energy per forecast. confirmedas of 2025-02-25

    15. [15]

      ECMWF describes the AIFS as a graph neural network encoder and decoder with a sliding-window transformer processor, trained on ERA5 reanalysis and ECMWF's operational analyses. confirmedas of 2024-08-07

    16. [16]

      The AIFS runs four times a day alongside ECMWF's physics-based model, and its forecasts are public under ECMWF's open data policy. confirmedas of 2026-10-10

    17. [17]

      ECMWF publishes the weights of AIFS Single v1.0, the first operationally supported AIFS model, under a CC BY 4.0 licence. confirmedas of 2026-10-10

    18. [18]

      Google DeepMind notes that tropical cyclones are very sensitive to small differences in atmospheric conditions, which makes them notoriously difficult to forecast accurately. confirmedas of 2025-06-12

    19. [19]

      In June 2025 Google DeepMind and Google Research launched Weather Lab, with an experimental AI cyclone model that generates 50 scenarios up to 15 days ahead, and began sharing its predictions with the US National Hurricane Center. confirmedas of 2025-06-12

    20. [20]

      Google DeepMind says Weather Lab is a research tool and its live predictions are not official warnings. confirmedas of 2025-06-12

    21. [21]

      Google DeepMind said its internal testing found the experimental cyclone model's track and intensity predictions as accurate as, and often more accurate than, physics-based methods. confirmedas of 2025-06-12

    22. [22]

      ECMWF lists known AIFS limitations including blurred forecast fields at longer lead times, reduced skill in the stratosphere and reduced intensity of some high-impact systems such as tropical cyclones. confirmedas of 2026-10-10

    23. [23]

      NOAA's AI models stem from Project EAGLE, a multi-year collaboration across NOAA research, the National Weather Service, academia and industry, and combine AI and physics-based approaches. confirmedas of 2025-12-17

    Revision history (2)
    1. Page created.
    2. Added AI cyclone forecasting (Weather Lab and the National Hurricane Center), the AIFS architecture and open weights, and ECMWF's list of known AI-model weaknesses.

    Created Oct 10, 2026. Last reviewed by an editor on Oct 10, 2026. Next scheduled review: Jan 10, 2027.

    Cite this page

    "How AI weather models work." ContentLora, updated Oct 10, 2026. https://contentlora.com/explain/how-ai-weather-models-work

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.

    Keep exploring

    ☰ Contents