AI weather forecasting

Also known as: machine learning weather prediction

On this page 7

AI weather forecasting is the use of machine learning models trained on historical atmospheric data to predict future weather states, such as temperature, wind, pressure and precipitation, without explicitly solving the equations of fluid dynamics at run time. Among AI applications, it is one of the clearest cases where a learned model matches or exceeds an established scientific method on many standard scores.

The approach is also called machine learning weather prediction. Its central promise is speed: a global ten-day forecast that takes hours on a supercomputer with numerical weather prediction can be produced by a trained network in about a minute on a single accelerator.

How AI weather models produce a forecast

An AI weather model learns a mapping from the current atmospheric state to the state a few hours later, then applies that mapping repeatedly to reach longer lead times. Most global models step forward in 6-hour increments, so a ten-day forecast requires about 40 consecutive steps, a process known as autoregressive rollout.

Training relies on reanalysis data, mainly the ERA5 dataset, which reconstructs global weather from 1979 onward on a grid of about 0.25 degrees, roughly 31 kilometers. The model sees millions of pairs of consecutive states and adjusts its weights to minimize the error of the predicted next state.

The table below lists the stages of a typical AI forecasting pipeline and what each stage contributes.

StageInputOutput
Data assimilationSatellite, radar, balloon and station observationsBest estimate of the current atmosphere
EncodingGridded analysis with dozens of variables and levelsCompact latent representation
Learned time stepLatent state at time tPredicted state at t plus 6 hours
RolloutRepeated time stepsForecast up to 10 or 15 days
PostprocessingRaw model fieldsLocal forecasts, warnings and products
Stages of an AI forecasting pipeline

A crucial detail is that most current systems still depend on physics-based data assimilation for their starting conditions. The learned model replaces the forecast step, not the full observing and analysis chain.

Main model families and architectures

AI weather models differ mainly in the neural architecture they use to represent the spherical atmosphere and in whether they produce one forecast or an ensemble. The best known systems were published between 2022 and 2025 by technology companies and meteorological centers.

The following table compares representative models by architecture and output type.

ModelDeveloperArchitectureOutput type
FourCastNetNVIDIA and partnersFourier neural operatorDeterministic
Pangu-WeatherHuawei3D vision transformerDeterministic
GraphCastGoogle DeepMindGraph neural network on a meshDeterministic
GenCastGoogle DeepMindDiffusion modelProbabilistic ensemble
AIFSECMWFGraph and transformer hybridDeterministic and ensemble
Representative AI weather models

Deterministic models give one most likely future, which tends to become smooth at long lead times because the training loss rewards average outcomes. Generative ensemble models sample many plausible futures and therefore describe uncertainty, which matters for risk decisions.

AI forecasting compared with numerical weather prediction

AI forecasting and numerical weather prediction (NWP) solve the same problem from opposite directions: NWP simulates physical laws, while AI models learn statistical regularities from past weather. In practice they are complementary, and operational centers now run both.

AspectNumerical weather predictionAI weather forecasting
Core methodDiscretized equations of motion and physicsLearned function fitted to historical data
Compute per forecastHours on a supercomputerMinutes on one GPU or TPU
Training costNone, but years of model developmentWeeks of accelerator time
Extreme eventsPhysically consistent intensityTendency to underestimate peaks
Unseen climatesGoverned by physical lawsUncertain outside the training range
InterpretabilityProcess-level diagnostics availableLimited insight into internal reasoning
AI models and physics-based models side by side

Published verification shows leading AI models outperforming top deterministic NWP on most headline variables at medium range, including tropical cyclone track errors. Scores for precipitation intensity, local extremes and fine-scale structure remain more mixed and depend on resolution.

Practical uses across forecast ranges

AI weather forecasting is applied at several time scales, from nowcasting the next few hours to subseasonal outlooks weeks ahead. Each range uses different data and offers a different kind of value.

RangeTypical dataCommon uses
Nowcasting, 0 to 6 hoursRadar and satellite imageryRain alerts, aviation, outdoor events
Short range, 1 to 3 daysAnalyses and regional observationsEnergy trading, logistics, road safety
Medium range, 3 to 15 daysGlobal reanalysis and analysesCyclone tracking, heat waves, planning
Subseasonal, 2 to 6 weeksOcean and land surface statesAgriculture and water management
Applications by forecast range

Because inference is cheap, forecasters can run large ensembles of hundreds of members and update them more often. Utilities use this speed to refresh wind and solar power forecasts several times a day, and insurers use it to model many storm scenarios quickly.

Limitations and open questions

The most serious limitation is reliance on historical patterns, which raises doubts about performance in record-breaking events and in a warming climate. Several further issues shape how responsibly the technology can be deployed.

  • Blurring at long lead times: deterministic models smooth fields, so extremes look weaker than they will be.
  • Physical inconsistency: outputs may break conservation of mass or energy unless constraints are added.
  • Dependence on NWP analyses: end-to-end models that start from raw observations are still experimental.
  • Limited variable sets: many models omit cloud cover, visibility or detailed surface variables.
  • Verification pitfalls: comparing models on smoothed scores can favor blurry forecasts over useful ones.

For these reasons, national weather services treat AI output as an additional guidance product reviewed by human forecasters. Official warnings still combine several models with expert judgment and local knowledge.

The short version

AI weather forecasting teaches a computer to predict the weather by showing it decades of past weather maps. The computer learns how today’s weather usually turns into tomorrow’s and then repeats that step many times.

Traditional forecasts calculate the physics of air and water, which needs huge supercomputers. The AI shortcut is much faster and often just as accurate, but it can miss the full strength of rare, extreme storms, so human forecasters still check the results.

Where it makes a difference

A national weather service tracks a tropical cyclone forming over the ocean seven days before possible landfall. Its supercomputer runs the regular physics-based ensemble twice a day, while a graph-based AI model produces a new global forecast in about a minute every time fresh analyses arrive.

The AI runs place landfall on a stretch of coast about 150 kilometers north of the earlier physics-based guidance, and later physics runs move toward the same area. Forecasters weigh both sources, note that the AI model underestimates peak wind speed, and take intensity from the physics-based ensemble.

The outcome is an earlier and better targeted evacuation warning. Emergency managers gain roughly a day of preparation time for the correct region, while the warning text still reflects realistic storm strength.