Description of the model

The AIFS ENS v1 model is a probabilistic weather forecasting system developed by ECMWF that uses machine learning to generate ensemble forecasts. It produces multiple forecast members by sampling from a learned distribution, capturing the uncertainty in future weather conditions. The model is trained using a version of the Continuous Ranked Probability Score (CRPS), a loss function that helps ensure the forecasts are both accurate and well-calibrated. This training approach accounts for the limitations of using a finite number of ensemble members. AIFS ENS outperforms ECMWF’s traditional physics-based ensemble system in medium-range forecasts and performs competitively for subseasonal forecasts when evaluated as anomalies.

New Methodology

The experimental AIFS ENS model available on ecCharts used a different methodology to AIFS ENS v1 (the experimental version uses diffusion, whereas the operational model optimises the CRPS). Therefore this implementation should not be seen as an update that can be compared with the previous experimental version, but instead it should be seen as the first version of a new model.

Control member

While the AIFS ENS includes a dedicated control member, this has a different interpretation to that of a physics-based ensemble (i.e. the IFS ENS). In traditional physics-based ensembles (like the IFS ensemble), a control member is a deterministic forecast run at high resolution without perturbations. It's often used as a baseline for comparison.

In contrast, AIFS ENS is a stochastic machine learning model that generates ensemble members by sampling from its learned internal distribution. The uncertainty within the AIFS ENS cannot be switched off to run a control run.

In AIFS ensemble the control member is a member of the ensemble that has been initialised using the same initial condition as the IFS control. Therefore it has unperturbed initial conditions but is not a true control member as the model is perturbed.

Timeline of the implementation 

 AIFS ENS will be implemented on the 1 July 2025  06 UTC run.

A short Release Candidate testing Phase (RCP) will start on 23 June 2025.

Datasets affected

This upgrade impacts the ensemble AIFS forecast data set.

Resolution

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

Dissemination schedule 

Schedule for dissemination

• AIFS data will be available according to the same schedule as IFS data. See Dissemination schedule for further details.
• Pre-schedule delivery will be available for eligible users, making the data available as soon as they are produced. 

Schedule for open data

• The delay on open data platforms will not be implemented for AIFS ENS data, making the data available as soon as they are produced. 

Meteorological content 

The variables provided by AIFS ENS are the same as those in AIFS-Single in the atmosphere. At the surface they are largely the same, but AIFS ENS will not provide some precipitation and land variables (see table below for full list).

Tropical cyclone tracks will be available only as graphical products in the first period. Raw BUFR files will be available at the later stage. 

Assimilation

• AIFS ENS uses the operational IFS ensemble initial conditions, regridded to an N320 grid. In other words, member n of AIFS ENS uses the same initial condition as member n of the IFS ensemble and the control of AIFS ENS uses the same initial condition as the control of the IFS ensemble.

Observations

• No observations are used to train AIFS ENS.

Model

Fields at step 0

There will be step 0 for all the parameters, which should streamline user workflows.

• For accumulated fields time step 0 will have a value of 0,
• For instantaneous fields, these come from the IFS analysis, regridded to an N320 grid. 

Evaluation 

Interactive scorecards presenting the new model performance across period February 2024 to February 2025 have been updated: 

• Scorecard of AIFS ENS compared against IFS 

Known issues

Known issues with the AIFS ENS forecast can be found on the following page: Known AIFS ENS Forecasting Issues.
This page will be updated as issues are investigated and resolved.

Key configuration values

A new MARS keyword - model - has been introduced to support AIFS Single v1, and is mandatory to use with AIFS ENS v1 as well. This keyword distinguishes between different AIFS models and will need to be specified in MARS requests for historical data and dissemination requests for real-time data.