2024 is almost over, and it hasn’t been a quiet or slow one. Machine learning (ML) for weather and climate has continued to evolve at breathtaking pace, even staying up to date with news and developments around the world is looking like a full-time job.
This year at ECMWF, we began by moving to 0.25 degree resolution for the AIFS deterministic model. The AIFS deterministic paper was submitted to the arXiv, documenting the full implementation of the AIFS. In the summer, we began running our first AIFS ensemble system, a diffusion-based solution. On the arXiv, we released a preprint framing how one could train a model from observation datasets alone. In the autumn we made it easier for everyone to run open-source ML models themselves from ECMWF open data, and then made the AIFS v0.2.1 weights freely available under a permissive licence.
For me, the most exciting development has been working with brilliant scientists across Europe. A collaborative project on machine learning modelling was started featuring 12 of ECMWF’s Member States and ECMWF, all pooling resources to move faster together. Another example of this collaboration was the work led by colleagues at MET Norway, adapting the AIFS architecture to deliver their model, Bris, a model with high resolution over the Nordics. Full details can be found in this preprint. This concept is being further developed in the Member State collaborative project.
In the summer, we introduced Anemoi, the open-source ML framework being developed by ECMWF and collaborators across Europe, from which it’s possible to train the AIFS, Bris and more models. This has already featured contributions from colleagues at the German National Meteorological Service (DWD), the Finnish Meteorological Institute (FMI), the Royal Netherlands Meteorological Institute (KNMI), MET Norway, MeteoSwiss and Belgium’s Royal Meteorological Institute (RMI), and we look forward to this list growing next year.
The year isn’t quite over, and if you keep your eye on the arXiv you may see some preprints from ECMWF exploring ensembles and observation-based training.
There’s been lots of amazing work happening elsewhere this year. Everyone will have their own favourite papers, but a few of mine were Aurora, particularly for its work on atmospheric composition, LUCIE, for pushing the limits of how little data and compute you can use, and the work of Vonich and Hakim using ML models to explore the question of predictability.
Next year promises to be even bigger
2025 will be the year that the AIFS goes operational, meaning that users can rely on timely and reliable AIFS forecasts no matter the weather. This is scheduled for the second half of 2025.
We’ve also been cooking up a new version of the deterministic AIFS, AIFS-single v0.3, which we expect to release early in 2025. This version will improve scores and expand the range of products, particularly targeting the energy sector. On top of this, the AIFS ensemble will see significant progress in the first half of 2025, with moves to increase resolution, expand web chart offerings and provide direct data access.
Within Destination Earth, we’ve been developing ML models for the wider Earth system, encompassing ocean, sea-ice, wave, land and hydrological processes, with some exciting new results coming. Sub-seasonal (up to 46 days) and seasonal (up to 7 months) forecasting are also very promising for machine learning, with some very interesting preprints and published articles on the topic. We’ve been doing some work ourselves and will be showcasing it next year.
ECMWF will continue to contribute to Anemoi, including bringing in ensemble techniques for more people to use. We’re excited to see more people giving Anemoi a test drive.
Merry Christmas and a Happy New Year from the AIFS team.
Top banner image: Graph used in graph neural network processors in Anemoi.