MeteoModels

Explainer

AI weather models, explained

Last updated: 14 November 2026

For seventy years, weather forecasts came from supercomputers solving the equations of physics. Since 2022 a second approach has gone from research papers to daily operations: machine-learning models that learn how the atmosphere evolves from decades of past data. This guide explains how they work, how good they are and what they still cannot do.

The main AI weather models

ModelDeveloperNotes
FourCastNetNVIDIA2022. Early demonstration that AI forecasts could approach traditional models.
Pangu-WeatherHuawei2023. Published in Nature; competitive with ECMWF on many scores.
GraphCastGoogle DeepMind2023. Published in Science; 10-day forecast in under a minute.
GenCastGoogle DeepMind2024. Probabilistic ensemble forecasts to 15 days.
AIFSECMWFOperational since February 2025; ensemble version since July 2025.
AIGFSNOAAOperational since December 2025; based on GraphCast.

How a traditional model works

A conventional numerical weather prediction model, such as ECMWF’s IFS or the GFS, divides the atmosphere into millions of grid boxes and applies the laws of fluid motion and thermodynamics to step each box forward a few minutes at a time. Processes too small for the grid, such as individual clouds and turbulence, are represented by approximations.

It works, and it has improved by about a day of useful forecast per decade. It is also enormously expensive: a single 10-day forecast takes around an hour on one of the world’s largest supercomputers.

How an AI model works

An AI weather model contains no equations of physics. It is a neural network trained on a long record of what the atmosphere actually did — typically ERA5, ECMWF’s reconstruction of global weather every hour since 1940. During training, the network is shown the state of the atmosphere at one time and learns to predict the state six hours later, millions of times over.

To make a forecast, the trained model is given the current state of the atmosphere and predicts six hours ahead. That output is fed back in to predict the next six hours, and so on, out to 10 or 15 days.

Training takes weeks on large clusters of graphics processors. Once trained, the model produces a full global forecast in about a minute on a single machine — thousands of times less energy than a physics-based run.

How accurate they are

Better than almost anyone expected. When GraphCast was published in 2023, it outperformed ECMWF’s high-resolution forecast on about 90% of the measures tested. GenCast did the same against ECMWF’s ensemble a year later. ECMWF’s own AIFS now matches or beats its physics-based sibling on many headline scores.

Their clearest strength is the large-scale pattern several days ahead: where the highs, lows and jet stream will be. They have also proved very good at the tracks of hurricanes and typhoons, often seeing a turn or a landfall location earlier than traditional models.

Where they fall short

Smoothness. Many AI models are trained to minimise average error, and the safest way to do that at long lead times is to hedge. The result can be fields that look increasingly blurred, with weaker peak winds and lighter rainfall maxima than reality. Newer probabilistic models such as GenCast and the ensemble version of AIFS address this, but it has not disappeared.

Intensity of extremes. A related weakness: the strongest tropical cyclones, the heaviest rain and the sharpest fronts tend to be underestimated. A model that has learned from the past also has less to go on for events more extreme than anything in its training record.

Resolution. Most global AI models work on a 0.25° grid, about 28 km. That is coarser than the 9 to 13 km of the leading physics-based models and far coarser than regional models, so local detail in mountains and along coasts is limited.

Fewer variables. AI models predict a set list of quantities. Traditional models produce hundreds of fields, from cloud layers to soil moisture, that forecasters and downstream users rely on.

They still need the physics

Every AI forecast starts from an analysis of the current atmosphere, and that analysis is produced by a traditional system that blends observations with a physics-based model. The training data, ERA5, is itself a product of the same kind of system. AI models are, for now, built on top of conventional forecasting, not a replacement for it.

That is why ECMWF and NOAA run their AI models next to their physics-based ones. Research into AI systems that work directly from raw observations is active, but it is not yet operational practice.

How forecasters use them

As additional, largely independent opinions. An AI model makes different kinds of errors from a physics-based one, so agreement between the two families is a strong signal, and disagreement is a useful warning.

A sensible way to read them: trust AI models for the broad pattern and for storm tracks in the medium range; rely on physics-based and high-resolution models for peak intensity, rainfall totals and local detail.

AI models on MeteoModels

MeteoModels includes two operational AI models, ECMWF’s AIFS and NOAA’s AIGFS, as part of the optional Premium subscription. They can be overlaid with the physics-based models and included in the consensus, so you can see directly where the AI and traditional forecasts agree. The eight physics-based global models are free.

Frequently asked questions

What is an AI weather model?

A machine-learning model that forecasts the weather by learning patterns from decades of past atmospheric data, instead of solving the equations of physics. Examples include GraphCast, GenCast, ECMWF’s AIFS and NOAA’s AIGFS.

Are AI weather models more accurate than ECMWF or GFS?

On many large-scale accuracy scores, the best AI models now match or beat the leading physics-based models. They are weaker on fine detail and on the intensity of extreme events.

What is the ECMWF AIFS?

The Artificial Intelligence Forecasting System, ECMWF’s machine-learning weather model. It became operational in 2025 and runs alongside the physics-based IFS.

What is GraphCast?

An AI weather model from Google DeepMind, published in 2023. It produces a 10-day global forecast in under a minute and outperformed ECMWF’s high-resolution model on most of the measures tested.

Will AI replace traditional weather models?

Not in the near term. AI models still depend on physics-based systems for their starting conditions and training data, so weather centres run both side by side.

Keep reading

Open the live model map All guides