BIGGER PICTURE Disaster preparedness
A weather forecast learned from decades of weather
Europe's weather centre added a different kind of forecaster to its daily operation: one trained on historical atmospheric states, running alongside the established physics-based system.
Source illustration- Operational introduction
- 25 February 2025
- Selected launch benchmark
- Up to 20% cyclone-track gains
- Open operational product
- Forecast data available
How the idea works
Read atmospheric fields
Analysed weather provides the starting state.
Evolve a learned representation
The model estimates how the state changes.
Deliver forecast grids
Weather services receive numerical fields for further use.
Past atmospheres become training examples
AIFS learns statistical representations of how the atmosphere changes. ECMWF describes an encoder that compresses weather fields, a processor that evolves the representation, and a decoder that maps it back to a forecast grid. The historical training data are not merely old newspaper weather reports: they include ECMWF's carefully assembled atmospheric reanalyses and operational analyses.
Sources: [2]
The important change was entering daily service
On 25 February 2025, ECMWF made AIFS operational alongside its physics-based Integrated Forecasting System. At launch it reported improvements on many measures, including gains up to 20% for tropical cyclone tracks. This is a specific forecast-performance comparison, not a promise that every weather variable or every storm improves by that amount.
Sources: [1]
From one model run to many practical decisions
Weather services can combine this additional forecast with other information when preparing guidance. Wind, rainfall and temperature affect decisions about travel, infrastructure and energy. The value is not a fluent explanation of tomorrow: it is numerical information people and operational systems can inspect. Keeping the physics-based forecast beside the learned model also preserves a useful comparison between approaches.
An open product, with versions that matter
ECMWF's current data page documents deterministic and ensemble AIFS products and their May 2026 upgrade to version 2. Forecast output is available under an attribution licence. These later versions should not inherit every numerical claim from the 2025 launch automatically. For an operational tool, identifiable versions and accessible data make improvements easier for other specialists to examine.
Sources: [3]