User Documentation
AIFS
OPERATIONAL
The AIFS is ECMWF's Artificial Intelligence Forecasting System. First launched in October 2023, the AIFS generates deterministic and ensemble weather forecasts from machine learning models.
- The AIFS single model has been operationally supported at ECMWF since 25 February 2025.
- The AIFS ensemble model has been operationally supported at ECMWF since 1 July 2025.
Details about the current operational versions of AIFS models, and a log of past model versions, can be found on AIFS Version History.
Meteorological information about the AIFS, including a guide for forecasters, is documented in the Forecast User Guide: Section 2.1.6 Machine Learning models.
How-to guides
- How to: Access AIFS model output data
- How to: Check which version of the AIFS is being used
- How to: Generate a forecast with the AIFS
FAQs
- FAQ: How does the AIFS work?
- FAQ: How does the AIFS compare to the IFS?
- FAQ: How does the AIFS compare to other data-driven models?
AIFS News
- August 2025 - "Representing ocean wind waves in ECMWF's AIFS"
- July 2025 - "ECMWF’s ensemble AI forecasts become operational"
- June 2025 - "Introduction to AIFS ENS v1" (video)
- May 2025 - "Verifying 2 m temperature forecasts in wintertime anticyclonic conditions"
- April 2025 - "Operational release of AIFS Single 1.0"
- February 2025 - "Introduction to AIFS Single v1" (video)
- February 2025 - "ECMWF's AI forecasts become operational"
- December 2024 - "First AIFS model weights are now open"
- December 2024 - "Accuracy versus activity"
- October 2024 - "Data-driven ensemble forecasting with the AIFS"
- June 2024 - "Enter the ensembles"
- March 2024 - "It's raining data"
- January 2024 - "First update to the AIFS"
- October 2023 - "ECMWF unveils alpha version of new ML model"
Anemoi
INCUBATING (see ECMWF's guidelines on software maturity)
Anemoi is an award-winning open-source, Python-based framework developed collaboratively by ECMWF and several European national meteorological services. It is designed to facilitate the development, training, and deployment of machine learning models for weather forecasting. As an 'end to end' framework, it provides a comprehensive toolkit that spans data preparation, model training, and inference, enabling meteorological organizations to leverage their own data for ML-based weather prediction. Anemoi packages are listed below, together with their dedicated user documentation.
In January 2025, a webinar series took place exploring various components of the Anemoi framework. Webinar recordings and presentation materials are freely available on the ECMWF website.
| Anemoi component | Description | Documentation | Webinar link |
|---|---|---|---|
| anemoi-core | A mono-repo containing core training and modelling functionality for Anemoi. Packages include: anemoi-graphs: Provides the functionality to create complex global or local area graphs. anemoi-models: Provides implementations for various type of models. These models are based on a graph encoder-processor-decoder approach and are implemented using the PyTorch library. anemoi-training: Provides the functionality to train machine learning models, using pytorch-lightning and Hydra. The training is highly configurable and fully defined through configuration files (utilising anemoi-models to achieve this). The package also includes profiling evaluation, plotting and logging of defined model and system metrics. anemoi-training is designed to work with datasets created using anemoi-datasets. | https://anemoi.readthedocs.io/projects/graphs/en/latest/ https://anemoi.readthedocs.io/projects/models/en/latest/ | |
| anemoi-datasets | Provides the tools to build datasets which are optimised for machine learning training, with appropriate chunking and precomputed statistics for normalisation. These datasets can be built from a range of input sources, including MARS, Grib, NetCDF, Zarr and more. | https://anemoi.readthedocs.io/projects/datasets/en/latest/ | 15 January 2025 |
| anemoi-inference | Provides the tools to take a trained model and perform the inference/rollout given some initial conditions. Inference makes full use of the metadata stored in a checkpoint to facilitate simple execution without requiring large amounts of boilerplate code. | https://anemoi.readthedocs.io/projects/inference/en/latest/ | 28 January 2025 |
| anemoi-registry | Provides the tools to save a dataset, a model or an experiment to the Anemoi catalogue so that it can be easily shared with others. The catalogue is accessible for permitted users only. | https://anemoi.readthedocs.io/projects/registry/en/latest/ | |
| anemoi-transform | Contains data transformation functions which can be applied to datasets (via filters). | https://anemoi.readthedocs.io/projects/transform/en/latest/ | |
| anemoi-utils | Contains miscellaneous utility functions which are used across the other packages. | https://anemoi.readthedocs.io/projects/utils/en/latest/ |
FAQs
Anemoi News
- October 2025 - "Introducing the Anemoi training-ready version of ERA5"
- May 2025 - "Florence Rabier and Anemoi to receive European Meteorological Society awards"
- February 2025 - "Discover Anemoi: Kicking off our 2025 machine learning training"
- October 2024 - "Anemoi: a new framework for weather forecasting based on machine learning"
- April 2024 - "Data-driven regional modelling"
ai-models
SANDBOX (see ECMWF's guidelines on software maturity)
ECMWF runs several third-party data-driven weather forecasting models alongside the AIFS. These state-of-the-art models are developed externally and are not maintained by ECMWF. The models are run daily and graphical output is publicly available on the OpenCharts platform. Raw data are not available to users.
Users can also generate their own forecasts from these models using the experimental ai-models package. Please note that support for third-party models accessed through ai-models is provided on a best-efforts basis, and may not always be possible. For model-specific functionality, we recommend contacting the original model developers.
Please note that, as ECMWF is co-developing the Anemoi framework for operational deployment of machine learning models, the ai-models interface will remain experimental, with no current plans for expanding its features.
How-to guides
- How to: Generate forecasts with ai-models
- How to: Create a plugin for your model with ai-models
- How to: Propose your model for inclusion in ECMWF's OpenCharts catalogue
FAQs
ai-models News
- September 2024 - "Run AI models yourself from ECMWF open data"
- December 2023 - "A new ML model in the ECMWF web charts"
GPUs
GPUs are available through the ATOS high performance computing facility (HPCF) and the European Weather Cloud. GPU usage is permitted only got approved users affiliated with National Meteorological and Hydrological Services in the Member and Co-operating States of ECMWF. GPU access may also be made available for research in the context of ECMWF Special Projects.
See Access to Computing Facilities for further details on how to request access to GPU resources.
Technical GPU documentation can be accessed via the following links:
- GPUs on the European Weather Cloud (including ready-made environments for the AIFS and ai-models)
- GPUs on the ATOS HPCF
GPUs News
- July 2023 - "All eyes on high-performance computing in meteorology"