AI weather forecasts replacing physics-based models

How far are AI weather forecasts from replacing physics-based forecasting models?

AI forecast models now run in operational services. ECMWF has run its AIFS model alongside its physics-based system since 25 February 2025. NOAA announced a suite of AI and hybrid AI-physics models on 17 December 2025. In both cases the physics-based systems still run alongside the AI models. The main gaps are rare and extreme events, tropical cyclone intensity, and dependence on reanalysis records of past weather. Most headline performance figures come from the developers' own tests.

AI illustration · Illustrative scene, not a photograph of the specific device or experiment.
WHERE WE AREPilots

Close. Early pilots are running in the real world.

AI forecast models run in two operational services, ECMWF and NOAA, so the approach has moved past prototypes. It is still a pilot because the AI models run alongside physics-based systems rather than replacing them. ECMWF describes combined AI and physics forecasting as an area of research, and the Met Office's FastNet model is not yet operational. Neither source says that either service will stop running its physics-based system.

The tech tree

Everything recorded on Vrulum that this future needs, oldest foundations at the top and the goal at the bottom. Lines show what each step requires.

What it takes: AI weather forecasts replacing physics-based models Unlocked Being researched Locked
  1. 2023AI weather forecastingPilot and early commercial Unlocked.
  1. 2023Learned model produces a global medium-range forecast in research testsDone Unlocked.
  2. 2024Learned ensemble forecast shown competitive in research testsDone Unlocked.
  3. 2025Deterministic AI model run in an operational forecasting serviceDone Unlocked.
  4. 2025National AI forecast suite run in operationsDone Unlocked.
  5. 2026Met Office AI model made operationalIn progress Being researched.
  6. Date not recordedAI ensemble forecast in ECMWF operationsNot started Locked.
  7. Date not recordedAI and physics forecasting combined as the standard approachNot started Locked.
  1. GoalAI weather forecasts replacing physics-based modelsGoalRequires: AI weather forecasting, Learned model produces a global medium-range forecast in research tests, Learned ensemble forecast shown competitive in research tests, Deterministic AI model run in an operational forecasting service, National AI forecast suite run in operations, AI ensemble forecast in ECMWF operations, Met Office AI model made operational, AI and physics forecasting combined as the standard approach. Locked.

The road so far

4 of 7 steps done, 1 in progress.

  1. Done · 14 Nov 2023Learned model produces a global medium-range forecast in research tests

    The GraphCast authors report a 10-day global forecast at 0.25 degree resolution in under one minute. They report that it outperformed the most accurate operational deterministic systems on 90% of 1,380 verification targets, in their own evaluation. [1]

  2. Done · 4 Dec 2024Learned ensemble forecast shown competitive in research tests

    The GenCast authors report a probabilistic ensemble forecast with greater skill than ECMWF's ENS ensemble on most of 1,320 targets in their evaluation. The preprint abstract gives 97.4% of targets, and says GenCast better predicts extreme weather, tropical cyclones and wind power production. The published Nature version was not read, so its figures may differ. [2][3]

  3. Done · 25 Feb 2025Deterministic AI model run in an operational forecasting service

    ECMWF put AIFS Single into operations on 25 February 2025, running it side by side with its physics-based Integrated Forecasting System (IFS). ECMWF describes AIFS as the first fully operational open weather prediction model using machine learning. [4]

  4. Done · 17 Dec 2025National AI forecast suite run in operations

    NOAA's release of 17 December 2025 says it has launched AIGFS, a single AI forecast model, AIGEFS, a 31-member AI ensemble, and HGEFS, a 62-member hybrid ensemble that combines AI and physics-based members. NOAA says AIGFS uses up to 99.7% less computing resource than its traditional counterpart, and that HGEFS outperformed both the AI-only and physics-only ensembles in initial testing. These are NOAA's own figures. [5]

  5. Not startedAI ensemble forecast in ECMWF operations

    ECMWF's announcement describes AIFS Single as deterministic and says that an ensemble version is a future step, with a 50-member ensemble as the aim. It gives no date for that ensemble. [4]

  6. In progress · 8 Jul 2026Met Office AI model made operational

    The Met Office says its FastNet model, co-developed with the Alan Turing Institute, reaches performance comparable to its Global Model on metrics such as RMSE, and exceeds it on some metrics. FastNet is not yet operational, and the press release gives no timeline. [6]

  7. Not startedAI and physics forecasting combined as the standard approach

    ECMWF describes hybrid machine-learning and physics-based forecasting as an area of research. The Met Office expects the future to involve a blend of AI and physics-based methods. No service has yet set out a date for when AI will be the main approach. [4][6]

Why it matters

AI forecasts can be produced with far less computing, and ECMWF and NOAA both report large savings on their own figures. If those savings hold, high-quality weather information could become cheaper and easier to produce. The gap that matters is accuracy on the events where forecasts matter most, such as storms, heat and floods. Whether AI forecasts can be trusted there will decide how far national services rely on them.

What stands in the way

  • The Met Office's evidence for FastNet on extreme events rests on two case studies, Hurricane Ian and Storm Ciarán. The press release describes qualitative improvements and gives no numerical results.
  • NOAA reports that AIGFS v1.0 shows degraded tropical cyclone intensity forecasts, which it says future versions will address. Its tropical cyclone track errors fell at longer lead times.
  • ECMWF's energy and accuracy figures for AIFS come with no stated baseline, period or verification method, so they cannot yet be checked independently.
  • The models learn from reanalysis records of past weather. The GraphCast abstract says it can be trained directly from reanalysis data, and the GenCast preprint abstract says it was trained on decades of reanalysis data.
  • Operational services still run physics-based systems beside the AI models. NOAA's hybrid ensemble, which combines AI and physics-based members, outperformed both its AI-only and physics-only ensembles in initial testing.

Signals to watch

  • ECMWF releasing an AI ensemble forecast with a stated operational date.
  • Published baselines, periods and verification methods for ECMWF's AIFS performance and energy figures.
  • A NOAA or ECMWF verification showing better tropical cyclone intensity forecasts from a later AI model version.
  • The Met Office deciding to make FastNet operational.
  • A national or regional service reducing or retiring its physics-based medium-range system.

Sources

  1. peer reviewed paper
    Learning skillful medium-range global weather forecasting
    Remi Lam et al.. Science, 14 Nov 2023.
  2. peer reviewed paper
    Probabilistic weather forecasting with machine learning
    Ilan Price et al.. Nature, 4 Dec 2024.
  3. preprint
    GenCast: Diffusion-based ensemble forecasting for medium-range weather
    Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet et al.. arXiv, 25 Dec 2023.
  4. other
    ECMWF's AI forecasts become operational
    ECMWF, 25 Feb 2025.
  5. government report
    NOAA deploys new generation of AI-driven global weather models
    NOAA, 17 Dec 2025.
  6. other
    New research shows AI can produce trustworthy, physically realistic weather forecasts
    Met Office, 8 Jul 2026.

Last reviewed 21 Oct 2026.