ECMWF Newsletter #178

AIFS: a new ECMWF forecasting system

Simon Lang
Mihai Alexe
Matthew Chantry
Jesper Dramsch
Florian Pinault
Baudouin Raoult
Zied Ben Bouallègue
Mariana Clare
Christian Lessig
Linus Magnusson
Ana Prieto Nemesio


There has been substantial progress recently in the realm of data-driven weather forecasting. Big technological companies like Google, Huawei and Nvidia have built purely data-driven weather forecasting models. These models outperform leading physics-based global numerical weather prediction (NWP) models in many of the standard forecast scores, such as root-mean-square error (RMSE) and Anomaly Correlation Coefficient (ACC) for geopotential height at 500 hPa. They are trained on historical weather data, usually a subset of ECMWF’s ERA5 reanalysis dataset, and they rely on traditional NWP analyses as initial conditions when producing a forecast.

To better understand and explore the machine learning (ML) technologies underpinning these models, ECMWF decided to implement a data-driven forecast model, the Artificial Intelligence/Integrated Forecasting System (AIFS), a homage to ECMWF’s Integrated Forecasting System, the IFS. The AIFS is underpinned by a toolkit (Anemoi, Greek: 'Winds'), which provides high-performance building blocks and pipelines to create and train massively parallel AI‑driven forecast models.

AIFS products are now available on ECMWF’s OpenCharts alongside charts of the IFS and other machine learning models (

AIFS forecast skill.
AIFS forecast skill. We show the northern hemisphere Anomaly Correlation Coefficient (ACC) for geopotential height at 500 hPa of IFS forecasts (red, dashed) and AIFS forecasts (blue) for 2022. Higher values indicate better skill.

ECMWF’s fully data-driven weather forecast model

There is currently a diverse set of deep-learning architectures being employed in the context of data-driven weather forecasts, e.g. vision transformers, neural operators, and graph neural networks (GNNs). We have decided to follow the approach of Ryan Keisler and Google DeepMind’s GraphCast and implement a forecast system based on GNNs. One attractive property of GNNs is that they can learn from data on arbitrary grids, and this allows the AIFS to work directly with the native IFS reduced Gaussian grids.

Our toolkit has a modular design that allows for flexibility, such as future extensions or architectural changes to the AIFS. It relies on the Pytorch machine learning framework, and Pytorch Geometric is used to implement the GNN architecture.

The version of the AIFS available at the time of writing is trained on a subset of the ERA5 reanalysis for 1979–2018 and fine-tuned on operational IFS data from 2019 to 2020. Pressure level fields and surface fields are used for training, together with forcing data, like time of the year and insolation. A complete list of AIFS inputs and outputs is given in the table. The input and output are currently at about one degree resolution.

Although they are still produced at relatively coarse resolution, AIFS forecasts show higher skill than IFS forecasts when measured by a range of standard forecast scores (an example is shown in the image). This is already in line with the current leading data-driven external models.

One area where resolution has a substantial impact is for surface parameters, such as 2 m temperature. As expected, for these variables the AIFS at a resolution of one degree is still behind the operational IFS configuration. First tests with higher-resolution variants of the AIFS show that large improvements are possible in this area.

AIFS training data.
AIFS training data. The table provides an overview of AIFS inputs and outputs during training.


The first incarnation of the AIFS shows very promising results, replicating the rapid progress that has been made in the realm of data-driven AI weather models. We are now training a higher-resolution version of AIFS, which once tested will replace the current version on OpenCharts. An important next step will be to extend AIFS to create ensemble forecasts. Here, different avenues are possible, such as probabilistic training or exploiting generative AI architectures, like diffusion models. Work has already started to explore these options. Another interesting area of ongoing research is the coupling of data-driven models trained on NWP analysis to observational data, or even the training solely involving observations (see the feature article in this Newsletter).

Moving forward, a new ECMWF Member State pilot project on machine learning was recently approved. In this, Member States and ECMWF will work together to further develop data-driven forecasting methods and models, applied across a range of scales from global to local.

There has been rapid progress in the development of highly skilful data-driven weather forecast models, and we expect further advances soon.