Project Meeting in Oslo: WeatherGenerator After Year One
From 10–12 February 2026, around 60 representatives of the EU project WeatherGenerator met in Oslo for the General Assembly and a dissemination workshop. One year after the project’s launch, the consortium reviewed progress, discussed scientific developments and challenges, and deliberately opened the project to external stakeholders.
At the center of the initiative is the development of a data-driven foundation model for weather and climate forecasting, integrating a wide variety of heterogeneous datasets. The “Earth System Data Exploration” research group at Forschungszentrum Jülich leads the core model development. The meeting marked the transition from an intensive initial phase to a more focused approach with publication-oriented and application-driven activities.
A European Large-Scale Project for Weather and Climate
WeatherGenerator is a four-year EU project with 16 partners from academia, national meteorological services, and industry. It is coordinated by the European Centre for Medium-Range Weather Forecasts (ECMWF). The project’s goal is to develop a foundation model for weather and climate forecasting – a model that ingests large amounts of diverse observational and simulation data, represents it in a latent space, and generates forecasts across multiple temporal scales.
Partners include Forschungszentrum Jülich, the Norwegian Meteorological Institute, the Max Planck Institute for Biogeochemistry, KNMI – Royal Netherlands Meteorological Institute, Météo-France, the Met Office, the CMCC Foundation – Centro Euro Mediterraneo sui Cambiamenti Climatici, ETH Zürich, and industry partners such as Statkraft.
The Jülich group leads the core model development. The challenge is high: to develop new methods for data-driven modeling, scale them on high-performance computers, and make them applicable from nowcasting to climate timescales.
One Year In: Where Do We Stand?
Project coordinator Peter Dueben (ECMWF) started the meeting with an explicit invitation to openness: challenges and even mistakes should be discussed transparently.
The first day focused on the thematic work strands: data and model architecture, evaluation, applications, coordination, communication and dissemination.
After only one year, the results already show the potential of integrating heterogeneous datasets – supported by an exceptionally engaged scientific community.
The model integrates a wide range of meteorological and geophysical datasets. Self-supervised learning plays a central role. Masking strategies allow information gaps to be filled from other datasets, enabling cross-training effects.
A student-teacher approach is used, where a student model is trained against a teacher implemented as an exponential moving average (EMA) of the student. This ensures the teacher remains slightly more accurate, stabilising the training process.
The WeatherGenerator’s forecast skill is now stable and approaching state-of-the-art levels. The model can already generate realistic cyclones (e.g., in the Indian Ocean). Temporal interpolation has also improved: six-hour forecasts can now be refined to hourly predictions. An advanced SYNOP decoder additionally accounts for orographic effects.
There is not a single correct approach when dealing with climate and weather problems. Every idea must be challenged and carefully planned before taking it into development.
FastEval was developed to directly evaluate model outputs (e.g., using RMSE or histograms) and can be combined with existing evaluation packages. The aim is a transparent, reproducible, and fast assessment of model performance.
WeatherGenerator’s applications range from weather and climate forecasts to renewable energy, water and food security, and health and biosphere considerations. Covered timescales span nowcasting, short- and medium-range forecasts, and subseasonal to seasonal predictions, with longer-term multi-year applications under development.
Opening to External Stakeholders: Dissemination and Networking
The second day was dedicated to external stakeholders. In addition to presenting the concept, architecture, and applications, sister projects were invited, including TerraDT within the Destination Earth initiative and UrbanAir in air quality modeling.
MetNorway provided insights into successes and challenges in integrating data-driven models into operational weather forecasting. Longer-term extensions, such as AI-driven climate simulations like HClimRep, were also presented.
A key message: The WeatherGenerator model is designed to allow third parties to join and use it soon. Dissemination is thus not only about sharing information but also an invitation to participate.
Panel Discussion: Ambition, Reality and Perspective
A highlight of the meeting was the panel session with representatives from both science and applications. Below are some of the key points discussed by the panel:
- It’s a possibility to train a relatively small model, allowing you to get by with very little compute and to make better use of different observations.
- This particularly excites Joel Oskarsson (ETH Zürich) and application lead Thomas Nipen (Norwegian Meteorological Institute).
- From an industry perspective, Jesica Piñón Rodríguez (Statkraft) noted: “From the industry side, it [the WeatherGenerator] can simplify our workflows, but the challenge is to become operational and arrive in time while so much is still evolving.”
- Critical reflection was also part of the discussion. A central question was how to ensure that new data-driven methods truly outperform existing physical models rather than merely complement them.
- The rapid evolution of the field was also addressed. Since the project proposal, data-driven methods have advanced significantly. Progress appears less explosive than during the early transformer breakthroughs but remains continuous and high-level.
- Richard Turner (The Alan Turing Institute) put it succinctly: “Is training AI weather models an art or a science? And do we need to rethink what these models are actually doing?” Especially for climate-related applications, out-of-distribution generalisation poses a particular challenge, he said.
Scientific Culture and Collaboration
Beyond content progress, collaboration was a key focus. Breakout groups discussed questions of publication and authorship, operationalisation, and training strategies for high-resolution regional applications.
A recurring theme was coordination: when over one hundred developers work simultaneously on a single model, efficient communication, version control, and quality assurance become critical.
The assembly came at a very good time where the project transitions from the initial development that may sometimes have seemed to be overwhelming and many-things-at-once to a somewhat more directed effort with clear intermediate goals, both with respect to desired model features but also in terms of scientific publications. Witnessing the enthusiasm of many people and their dedication to the WeatherGenerator is indeed a wonderful experience.
The Oslo meeting marks a transition: from an intensive build-up phase to more focused, publication- and application-oriented work.
WeatherGenerator is scientifically “well on track and well organised.” At the same time, the challenge remains high: to develop new methods, scale them, critically evaluate them, and ultimately make them operational.
For the Jülich research group, the meeting represents not only scientific progress but also an expansion of their engagement portfolio – in international collaboration, in nurturing early-career scientists, and in actively shaping the next generation of data-driven weather and climate models.