Description of the upgrade
The operational release of AIFS Single v1 (aifs-single) marks the first operationally supported AIFS model. Version 1 will supersede the existing experimental version, v0.2.1. The new version will bring changes to the deterministic AIFS model, including among many others:
- Improved performance for upper-level atmospheric variables (AIFS Single v1 still uses 13 pressure-levels, so this improvement mainly refers to 50 hPa).
- Improved scores for total precipitation.
- Additional output variables, including 100 meter wind components, snowfall, solar radiation and land variables such as soil moisture and soil temperature.
Description of the model
Version 1 is the newest version of the AIFS deterministic model. For the first time, it has been implemented using the Anemoi framework. Anemoi consists of open source components, enabling the user community to run AIFS themselves.
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 is used for the fine-tuning step towards operational IFS analyses. The changes result in improved forecast skill, especially for precipitation.
The training procedure for AIFS Single v1 consists of:
- Pre-training on the ERA5 dataset for the years 1979 to 2022. A cosine learning rate (LR) schedule and a total of 260,000 steps are used. The LR is increased from 0 to 10−4 during the first 1000 steps, then it is annealed to a minimum of 3×10−7. The local learning rate used for this stage is 3.125x10-5.
- Pre-training is then followed by rollout on operational real-time IFS analyses for the years 2016 to 2022. A local learning rate of 8×10−7 is used, which is subsequently decreased to 3×10−7. Rollout steps are increased with each epoch. In this second, shorter stage, the warm up period of the optimiser is set to 100 steps. Optimizer steps are equal to 7900 (12 epochs with ~630 steps per epoch).
As in previous versions of AIFS Single, IFS fields are interpolated from their native O1280 resolution (approximately 0.1°) down to N320 (approximately 0.25°) for fine-tuning and initialisation of the model during inference. The optimizer used is AdamW (Loshchilov and Hutter (2019)) with the β-coefficients set to 0.9 and 0.95.
Open source distribution of AIFS Single v1
The experimental AIFS Single v0.2.1 model is already available on Hugging Face, enhancing accessibility and transparency for ECMWF’s advanced AI-driven forecasting system. An example notebook is available to the user community, demonstrating how to run the AIFS from ECMWF open data and the anemoi-inference package. Additionally, the new operational model version, AIFS Single v1, will also be published on the Hugging Face platform, further supporting the community’s exploration and utilisation of these cutting-edge models.
Timeline of the implementation


Datasets affected
This upgrade impacts the deterministic AIFS forecast dataset (Set IX - AIFS).
Resolution
There are no changes in resolution compared to previous version AIFS Single v0.2.1.
| Component | Horizontal resolution | Vertical resolution [pressure levels] |
Atmosphere | AIFS Single v1 | N320 | ~36 km | 13* |
*Levels (hPa) 50, 100, 150, 200, 250, 300, 400, 500, 600, 700, 850, 925, 1000
Dissemination schedule
With the operationalisation of the AIFS, output products will adhere to a new dissemination schedule from 25 February 2025 06 UTC run.
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 be removed for AIFS data, making the data available as soon as they are produced.
Meteorological content
Assimilation
- AIFS Single v1 uses the operational IFS control initial condition, regridded to an N320 grid.
Observations
- No observations are used to train AIFS Single v1.
Model
In terms of model changes, the following modifications have been introduced for AIFS Single v1:
- Include new variables
- prognostic: soil moisture (vsw), soil temperature (sot)
- diagnostic: 100 metre wind components (100u, 100v), solar radiation (ssrd, strd), cloud variables (tcc, hcc, mcc, lcc), runoff (rowe), snow fall (sf)
- fixed: standard deviation of sub-gridscale orography (sdor), slope of sub-gridscale orography (slor)
- For the upper-air variables, the loss weight depends on the pressure level (linear in pressure level). In this new release, we introduce a minimum value of this pressure-level dependent weight to 0.2, this increases the skill at 50 & 100hPa.
- Bounding strategies are applied to specific variables (tp, cp, tcc, lcc, mcc, hcc) achieved via activation functions. This ensures that variables like total precipitation and convective precipitation are constrained to be positive definite. Cloud cover components (tcc, lcc, mcc, hcc) are each bounded between 0 and 1, and ensure than the sum of lcc, mcc & hcc cannot exceed tcc.
Additionally, this new release has been implemented using the Anemoi framework.
Fields at step 0
A new feature of AIFS Single v1 is the introduction of step 0 for all the parameters, which should streamline user workflows.
- For instantaneous fields, these come from the IFS analysis, regridded to an N320 grid.
Meteorological impact
Medium range
Upper-air parameters
Compared to the previous version of AIFS, the performance of upper-air parameters in the troposphere is generally improved (see scorecards below). For some parameters (e.g 500hPa temperature) in the extra-tropics, we see a decreased smoothing, while the same parameter comes out smoother in the tropics. The drift towards colder temperatures at 500hPa is also reduced in the extra-tropics.
Tropical cyclones
The considerable improvements in tropical cyclone track predictions already found in previous AIFS versions compared to the IFS remain in the new version 1. In fact, the new AIFS version reduces TC position errors even further, by a small amount, which is due to an enhanced environmental steering flow and thus a reduced along-track bias. In terms of tropical cyclone intensity, this model change has an overall neutral effect on minimum central pressure and maximum wind speed, which means that the strong underestimation of intensity continues to exist. The resulting deviation in the pressure-wind relationship compared to observed values is inherited from the ERA5 training data.
Weather parameters
Precipitation
With AIFS version 1, a bounding is applied to precipitation during the training of the model to constrain the model to predict positive precipitation amounts. In the previous version, a simple clipping to zero was applied during inference. The new version sees a clear and positive impact on forecast performance throughout the 10 days of lead times. In particular, the bias related to small precipitation amounts is significantly reduced.
Snowfall
The new version includes snowfall as an accumulated parameter. The snowfall is a part of the total precipitation and can be either convective or large-scale, but this aspect is currently not included as output. As the snowfall training data is coming from ERA5 it relies on the diagnosed precipitation type in ERA5, which is not constrained by observations in the data assimilation in the reanalysis. Snowfall does not equate to either snow depth on the ground, or to additional snow depth on the ground. There is no snow depth parameter in this AIFS version.
Clouds and radiation
Cloud cover variables and solar radiation components are also new parameters included for the first time in this AIFS version. An assessment of total cloud cover and surface short-wave downwards radiation against independent observations reveals substantial lower mean squared error with respect to IFS forecasts at the cost of forecasts that are less well-calibrated. The cloud averaged distribution in AIFS tends towards more intermediate cloudy situation than IFS or the observations. This can be seen in the cloud forecast product as less sharp images due to the lack of cloud free and fully cloudy pixels. Similarly, solar radiation forecasts are less active in AIFS than IFS.
Soil moisture and temperature
AIFS Single v1 includes the soil temperature and moisture on soil layers 1 and 2, which represent 0-7 cm and cm respectively. While soil moisture is very difficult to verify due to lack of observations, the AIFS output was checked for a few points and the evolution of the moisture in the forecast looks reasonable.
100-metre wind components
100 metre U and V components are now available as a diagnostic output of AIFS. The relative to IFS analyses is overall close to IFS predictive skill. In general, AIFS shows a lower mean squared error than IFS but a higher mean error (a negative bias) that tends to increase with lead time before plateauing.
Sub-seasonal range
- AIFS Single v1 provides medium-range forecasts up to 15 days ahead, therefore no validation or impact is assessed in sub-seasonal scales.
Evaluation
Interactive presenting the new cycle performance across 2024 have been updated:
Known issues
Known issues with the AIFS forecast can be found on the following page: Known AIFS Forecasting Issues. This page will be updated as issues are investigated and resolved.
Key configuration values
A new MARS keyword has been introduced to support AIFS Single v1, highlighted in blue in the table below. 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.
There will be no changes to the configuration of the atmospheric model.
|
| |
---|
|
| Current (v0.2.1) | Upgrade (v1) |
---|
Basetime & frequency | 00/06/12/18 daily | 00/06/12/18 daily |
---|
Forecast range | 15-days | 15-days |
---|
15-days | 15-days |
| | ai | ai |
---|
| oper | oper |
---|
| N/A | aifs-single |
---|
Spectral | N/A | N/A |
---|
Gaussian grid | n320 | n320 |
---|
Horizontal grid resolution | ~36 km | ~36 km |
---|
Dissemination (LL) | 0.25° | 0.25° |
---|
Model Level vertical resolution | 13 | 13 |
---|
3 Comments
Alan Hally
Congratulations to all in ECMWF on this fantastic achievement!
When trying to access the scorecards linked under the "Evaluation" heading, I get a "Sorry, you're not authorized to access this page". Just wondering is this my issue or a permissions setting on the embedded link?
Thanks in advance.
Milana Vuckovic
Hi Alan,
Sorry about that! Try now.
Alan Hally
Perfect, thanks for the rapid response Milana!