AIFS ENS-v2 (operational)

Model description

Version 2 of AIFS ENS is a probabilistic weather forecasting system developed by ECMWF that uses machine learning to generate ensemble forecasts. It is implemented using the Anemoi framework, which consists of open source components, enabling the user community to run AIFS themselves. The model and an example notebook are freely available to the user community, demonstrating how to run AIFS ENS from ECMWF open data and the anemoi-inference package. 

The model is based on a graph neural network (GNN) encoder and decoder, and a sliding window transformer processor. The upgrade to v2 introduces several changes to the model:

• Replacing the afCRPS loss function used in AIFS ENS v1 with a multi-scale loss.
• 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 year of additional training data compared to AIFS ENS v1.

Model performance

Interactive scorecards presenting model performance across the period January - March 2026 are available: 

• AIFS ENS v2 (initialised from 50r1) compared with AIFS ENS v1 (initialised from 49r1)

  
  

• AIFS ENS v2 compared with AIFS ENS v1 (both initialised from 50r1)

Model resolution

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

Model output parameters

The release of AIFS ENS v2 introduces new parameters:

• A new wave component, including 11 wave variables, marking ECMWF's first operational data-driven wave forecasts.
• The addition of a new snow variable to the existing land component.
• The addition of tropical cyclone track forecasts.
• The addition of two variables already present in AIFS Single (vsw and cp).
• The addition of pressure level fields at 10hPa.

A list of parameters output by AIFS-ENS is provided on the implementation page. All parameters are output in GRIB2 format and available via dissemination, the MARS archive, open data platforms and OpenCharts. See https://www.ecmwf.int/en/forecasts/dataset/aifs-machine-learning-data for more information on how to access AIFS model output data.

AIFS ENS-CRPS v1 

Model description

Version 1 of the AIFS ensemble CRPS model was a probabilistic weather forecasting system developed by ECMWF that used machine learning to generate ensemble forecasts. It was implemented using the Anemoi framework, which consists of open source components, enabling the user community to run AIFS themselves. The model and an example notebook are freely available to the user community, demonstrating how to run AIFS ENS from ECMWF open data and the anemoi-inference package. 

The model was trained via a probabilistic training approach based on the Continuous Ranked Probability Score (CRPS), a loss function that helps ensure the forecasts are both accurate and well-calibrated.

Model performance

Interactive scorecards presenting model performance across the period March 2024 to March 2025 are available: 

• Scorecard of AIFS ENS compared against IFS 

Model resolution

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

Model output parameters

A list of parameters output by AIFS-ENS v1 is provided on the implementation page. All parameters were output in GRIB2 format and available via dissemination, the MARS archive, open data platforms and OpenCharts. See https://www.ecmwf.int/en/forecasts/dataset/aifs-machine-learning-data for more information on how to access AIFS model output data.

AIFS ENS-DIFF v0.1

This blog post introduces version 0.1 of the AIFS ensemble diffusion model. It comprised 51 members, using the same initial conditions as the IFS ensemble system.

Model resolution

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

Model output parameters

The parameters listed in the table below were output by AIFS ensemble version 0.1. 

AIFS Single v2 (operational)

Model description

Version 2 of the AIFS Single model is a determinsitc weather forecasting system developed by ECMWF that uses machine learning to generates a single forecast. It is implemented using the Anemoi framework, which consists of open source components, enabling the user community to run AIFS themselves. The model and an example notebook are freely available to the user community, demonstrating how to run AIFS ENS from ECMWF open data and the anemoi-inference package. 

The model is based on a graph neural network (GNN) encoder and decoder, and a sliding window transformer processor. No architectural changes have been introduced with the upgrade to v2. However, the data used for fine-tuning AIFS Single v2 has been updated:

• Fine-tuning is performed on 2 years of additional data compared to AIFS Single v1.1.
• In particular, fine-tuning was performed on ECMWF operational analysis data and IFS 50r1 esuite analysis data, covering the years 2018-2024 over 7,900 training steps.

Model performance

Interactive scorecards presenting model performance across the period January - March 2026 are available: 

• AIFS Single v2 (initialised from 50r1) compared with AIFS Single v1.1 (initialised from 49r1)

• AIFS Single v2 compared with AIFS Single v1.1 (both initialised from 50r1)

Model resolution

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

Model output parameters

The release of AIFS Single v2 introduces new parameters:

• A new wave component, including 11 wave variables, marking ECMWF's first operational data-driven wave forecasts.
• The addition of a new snow variable to the existing land component.
• The addition of pressure level fields at 10hPa.

A list of parameters output by AIFS-Single is provided on the implementation page. All parameters are output in GRIB2 format and available via dissemination, the MARS archive, open data platforms and OpenCharts. See https://www.ecmwf.int/en/forecasts/dataset/aifs-machine-learning-data for more information on how to access AIFS model output data.

AIFS Single v1.1

The upgrade of AIFS Single to version 1.1 represented a slight modification to the AIFS Single model. Version 1.1 addressed point rain artefacts identified in AIFS Single v1 (see Known AIFS Single Issues P3).

The underlying cause was identified as an unphysical relationship between soil moisture and precipitation, caused by an overemphasis of soil moisture in the training.

Version 1.1 of AIFS Single minimally changed the training configuration to a lower (by a factor 100) contribution from soil moisture. The outcome was a model with equivalent (but not identical) skill and bias. 

No other architectural changes were introduced; further details about the model are provided in the description of AIFS Single v1 (below).

AIFS Single v1

Model description

Version 1 of the AIFS deterministic model was implemented using the Anemoi framework. Anemoi consists of open source components, enabling the user community to run AIFS themselves. The model and an example notebook are freely available to the user community, demonstrating how to run the AIFS from ECMWF open data and the anemoi-inference package. 

The model architecture is identical to the previous experimental versions (v0.2 and v0.2.1). However, the training regime and pressure level scaling is revised, and a new way to handle bounded variables is introduced. In addition, more data are used for the fine-tuning step towards operational IFS analyses. The changes result in improved forecast skill, particularly for precipitation.

Model performance

Scorecards presenting performance of AIFS Single v1 across 2024 are available on Hugging Face. 

Model resolution

There were no changes to the resolution of version 1.

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

Model output parameters

AIFS Single v1 introduces new parameters, detailed below. 

Two existing parameters already introduced with AIFS Single v0.2.1 see changes due to GRIB 2 formatting with the implementation of AIFS Single v1.

All other existing parameters, listed on the implementation page, remain unchanged.

All parameters are output in GRIB2 format and available via dissemination, the MARS archive, open data platforms and OpenCharts. See https://www.ecmwf.int/en/forecasts/dataset/aifs-machine-learning-data for more information on how to access AIFS model output data.

AIFS Single v0.1

This blog post and newsletter article introduce the first version of the AIFS deterministic model. 

Model description

The first implementation of the AIFS was based on message-passing graph neural networks and with an internal icosahedral grid with multi-scale edges, similar to Deepmind’s GraphCast. 

Graph Neural Networks allow us to move away from lat-lon grids, which have many points near the poles, and use reduced Gaussian grids which have near equal distance between grid-points no matter where you are on the globe. These are also the grids which are used by the IFS to create ERA5 and operational initial conditions.

Version 0.1 was trained on a subset of the ERA5 reanalysis for 1979–2018 and fine-tuned on operational IFS data from 2019 to 2020.

Model resolution

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

Model output parameters