Description of the upgrade

The operational release of AIFS ENS v2 will introduce scientific and architectural changes to the AIFS ensemble model.

New and revised output variables 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.
• The addition of 2 variables already present in AIFS Single (vsw and cp), to harmonise the two models.
• An improved representation of the stratosphere, with the addition of pressure level fields at 10hPa.

Architectural changes include:

• Replacing the afCRPS loss function used in AIFS ENS v1 with a multi-scale loss. See https://arxiv.org/pdf/2506.10868.
• Imposing the same variable bounding as is used by the AIFS Single model, improving physical consistency.
• Revised graph features, utilising more edges in the decoder.
• An improved training regime, using 1 years of additional training data compared to AIFS ENS v1.

This upgrade also includes a technical fix:

• In AIFS ENS v1.0, some parameters have incorrect encoding of the ecCodes key productDefinitionTemplateNumber, which is used to describe the type of data.
  • Parameters 100u, 100v, hcc, lcc, mcc, rowe, sf, ssrd, strd, tcc and tp are encoded with productDefinitionTemplateNumber=0 or 8, which correspond to single forecasts.
  • The correct encoding for ensemble forecasts is productDefinitionTemplateNumber=1 or 11, which will be applied to all AIFS ENS output products with this upgrade.
  • See https://codes.ecmwf.int/grib/format/grib2/ctables/4/0/ for a description of these template numbers.

Description of the model

The upgrade of AIFS ENS to version 2 introduces changes to the training regime and model architecture. 

The full training procedure for AIFS ENS v2 consists of:

• Pre-training, which remains the same as for AIFS ENS v1
  • Pre-training is performed on ERA5 data covering the years 1979–2022 over 300,000 training steps.
• Fine-tuning, which has been updated for AIFS ENS v.2 
  • Fine-tuning is performed on  years of additional data compared to AIFS ENS v1.0. 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 ENS, 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. 

For both phases of training, a cosine learning rate schedule and a batch size of 16 is used. 

Further model changes include introducing a multi-scale loss and removing reference field truncation, imposing the same variable bounds used by the AIFS Single model to improve physical consistency, and revising graph features by using more edges in the decoder together with new edge features.

AIFS ENS v2 is implemented using the Anemoi framework. Anemoi consists of open source components, enabling the user community to run AIFS themselves.

Timeline of the implementation 

Datasets affected

This upgrade impacts the ensemble AIFS forecast dataset (Set X - AIFS ENS).

Resolution

AIFS ENS v2 introduces a new pressure level to the stratospheric component.

Meteorological content 

Initial conditions

• AIFS ENS v2 uses the operational IFS control initial condition, regridded to an N320 grid.

Observations

• No observations are used to train AIFS ENS v2.

Meteorological impact

Coming soon!

Evaluation 

Coming soon!

Known issues

Coming soon!

Key configuration values

*The MARS keyword stream=enfo will be used for the new wave parameters introduced in this upgrade.

**Levels (hPa): 10 (new), 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000

Technical content

GRIB encoding

The GRIB model process identification numbers for AIFS ENS v2 will be as follows:

grib_get -p generatingProcessIdentifier <file.grib2>

Encoding of wave parameters

Wave parameters are newly introduced with the implementation of AIFS ENS v2.0.

Users should note that wave parameters in AIFS are encoded differently from those produced by the IFS model. This is to facilitate ECMWF’s migration towards GRIB2 as the standard output data format.

• AIFS wave parameters use MARS stream 'enfo' (stream=enfo).
• IFS wave parameters use a dedicated wave stream (stream=waef).

Availability of AIFS ENS v2 test data

Test data in MARS 

Coming soon!

Test data in dissemination

Coming soon!

Test data on ECMWF's Open Data platform

Coming soon!

Resources

Webinar

Coming soon!

References

• Let it snow! How machine learning will forecast snow in the AIFS, Raoult, N., Pinnington, E., Arduini, G., Rudiger, C., De Rosnay, P., Chantry M.

• Representing ocean wind waves in ECMWF's AIFS, Hahner, S., Bidlot, J., Kousal, J., Zampieri, L., Lessig, C., Chantry, M.
• Simon Lang, Martin Leutbecher, Pedro Maciel, “A multi-scale loss formulation for learning a probabilistic model with proper score optimisation” (2025), arXiv:2506.10868.