AI weather forecasting: what it can tell you
AI models can produce useful forecasts of large weather systems days ahead. Their value depends on the model, the variable and the forecast range; a global forecast does not tell you whether a tornado will strike your street.
Updated September 12, 2026. Sources and model versions are identified below.
How does AI weather forecasting work?
Models such as GraphCast learn relationships in historical atmospheric data, then use a recent atmospheric state to predict later conditions. Forecasts can be advanced in steps to cover several days. Numerical weather prediction instead calculates the evolution of the atmosphere through physical equations. Both depend on good starting information, and both must be checked against observations.
GraphCast: a global forecast, not a street forecast
Google DeepMind's original GraphCast work describes a 10-day forecast on a 0.25-degree global grid, advanced in six-hour steps. Its published comparison found better performance than ECMWF's HRES on more than 90% of 1,380 evaluated variable and lead-time combinations. That is a result for a defined experiment, not a claim that every local prediction is 90% correct. Grid spacing, training data, forecast date and the comparison system all matter. Read DeepMind's GraphCast explanation and linked paper.
Aurora: a model family with different tasks
Microsoft describes Aurora as an Earth-system foundation model that can be adapted to different forecasting tasks. Its project includes weather and other environmental applications. The current project documentation also distinguishes Aurora 1.5, with additional variables, hourly output and ensemble forecasting. Results from one version or task should not be assumed to describe every Aurora forecast. Check the version and output variable when comparing a chart with another service. Microsoft Research's Aurora project.
When did ECMWF AIFS become operational?
ECMWF announced that its Artificial Intelligence Forecasting System became operational on February 25, 2025. It runs alongside ECMWF's physics-based forecasting system, providing an additional forecasting approach. The earlier version of this article incorrectly gave 2024 as the operational start date. Operational status means a system is being produced for use; it does not mean its forecasts are certain. ECMWF's operational announcement.
How to compare AI and conventional forecasts
- Compare forecasts initialized at the same time and valid for the same hour.
- Check whether the map shows a single forecast or a probability from an ensemble.
- Compare the same variable: a pressure-pattern score does not establish local rainfall accuracy.
- Look for verification across many events, regions and seasons, rather than one successful storm track.
- Treat runtime claims as specific to the hardware and task. Training, preparing input data and producing a forecast are different costs.
Can AI predict a tornado at my address?
The global models discussed here should not be used as address-level tornado alerts. A regional environment favorable for severe storms is different from confirmation of an individual tornado. Use the severe-weather outlook to understand broader risk, and learn the difference between a watch and a warning. For immediate decisions, follow warnings and instructions from your official local weather service. National Weather Service tornado safety guidance.