In April 2026 the operational Artificial Intelligence Forecasting System (AIFS) will be upgraded.
- AIFS Single, the deterministic model, will be upgraded from v1.1 to v2.0. This page provides details of the implementation of AIFS Single v2.0.
- AIFS ENS, the ensemble model, will be upgraded from v1.0 to v2.0. Please refer to <link page> for details of the implementation of AIFS ENS v2.0.
IMPORTANT: The AIFS will be upgraded on the same day as IFS Cycle 50r1.
Subscribe to forecast_changes-request@lists.ecmwf.int to receive notifications about updates, including when the test data will be available.
Description of the upgrade
The operational release of AIFS Single v2.0 will introduce scientific changes to the AIFS deterministic model.
Changes include:
The introduction of a new wave component, including 10 wave variables, marking ECMWF's first operational data-driven wave forecasts.
The addition of 2 new snow variables to the existing land component.
- Improved vertical velocities, by changing parameter W from a prognostic to a diagnostic field.
- An improved representation of the stratosphere, with the addition of pressure level fields at 10hPa.
- An improved training regime, using 2 years of additional training data compared to AIFS Single v1.1.
Description of the model
The upgrade of AIFS Single to version 2.0 introduces changes to the training regime.
The full training procedure for AIFS Single v2.0 consists of:
- Pre-training, which remains the same as for AIFS Single v1.1.
- Pre-training is performed on ERA5 data covering the years 1979–2022 over 260,000 training steps.
- Fine-tuning, which has been updated for AIFS Single v2.0.
- Fine-tuning is performed on 2 years of additional data compared to AIFS Single v1.1. In particular, fine-tuning is performed on ECMWF operational analysis data and available IFS 50r1 esuite analysis data, covering the years 2018-2024 over 7,900 training steps.
- As in previous versions of AIFS Single, IFS fields are interpolated from their native O1280 resolution (approximately 0.1°) using MARS default interpolation tools down to N320 (approximately 0.25°) for fine-tuning and initialisation of the model during inference.
- Fine-tuning is performed on 2 years of additional data compared to AIFS Single v1.1. In particular, fine-tuning is performed on ECMWF operational analysis data and available IFS 50r1 esuite analysis data, covering the years 2018-2024 over 7,900 training steps.
For both phases of training, a cosine learning rate schedule and a batch size of 16 is used.
This upgrade does not introduce any changes to the model architecture; the model remains the same as the existing operational model (AIFS Single v1.1). Technical details about the model architecture are detailed in the arXiv preprints here and here.
AIFS Single v2.0 is implemented using the Anemoi framework. Anemoi consists of open source components, enabling the user community to run AIFS themselves.
Want to get updates about future AIFS cycle upgrades?
Join the mailing list
To subscribe or unsubscribe, please send an email to
forecast_changes-request@lists.ecmwf.int
with either subscribe or unsubscribe as Subject.
Join our FORUM: https://forum.ecmwf.int/ and 'watch' the announcements.
Follow the LinkedIn channel for users: https://www.linkedin.com/showcase/ecmwf-users/posts
Timeline of the implementation
Why is AIFS Single v2.0 being implemented on the same day as IFS Cycle 50r1?
Degraded performance is observed when initialising AIFS Single v1.1 from esuite data from the new IFS Cycle, 50r1. Therefore, the AIFS model needs to be upgraded at the same time as the IFS model. The new AIFS Single version, v2.0, is fine-tuned on esuite data from the new IFS Cycle, 50r1.