Skip to Main Content
WeatherGenerator

About the Project

Weather Generator is a pan-European initiative that fuses state-of-the-art machine-learning architectures with high-performance computing to build an open, kilometer-scale foundation model of the coupled Earth system.

The main idea behind the WeatherGenerator is the use streams of various different data products – from observations via reanalysis data to model output – and the use of masked token learning in a self-supervised training approach to learn how to translate between the different data streams.

Once this is done, a timestepping scheme in latent space is established to allow for the progression of the model state in time to perform forecasts of the physical fields.

Users can than bring their own data to the tool and provide the WeatherGenerator with whatever data they have available to generate meaningful outputs for their application.

Work Structure

The overall project is organised in four Themes. The first Theme puts the datasets that are needed together and makes them useable for machine learning in an efficient and scalable manner. The second theme is building the core WeatherGenerator model. The third Theme is working on twenty-two applications that are developed by the project partners. The fourth Theme coordinates the project, organises hackathons and dissemination workshops, and provides services to externals to get started using the WeatherGenerator for their application.
Explore Themes
Video thumbnail
Data Model Application Service

Applications

We define the 22 applications that will test the WeatherGenerator. These applications span a range of use-cases, covering weather and climate, renewable energy, water, and the biosphere. Applications have global and regional coverage, cover timescales from nowcasting through seasonal forecasting, and include forecasting, downscaling, and data-fusion tasks.

  • Weather Prediction

  • Global medium-range forecasting

    AP1

    Global medium-range forecasting

    AP1
    About
    In this application, ECMWF will use the WeatherGenerator to generate global medium-range weather forecasts. Forecasts will be produced either directly in a zero-shot setting or by fine-tuning the forecasting decoder for medium-range predictions. The quality of predictions will be evaluated against IFS, AIFS, and other machine-learned forecasting systems using standard diagnostics and forecast scores. Experimental daily forecasts will be made publicly available via ecCharts, allowing transparent comparison between the WeatherGenerator, conventional IFS forecasts, and AIFS, and providing an important check on the model’s robustness.

    Key features:

    • Main developer: ECMWF
    • Time horizon: 3-10 days ahead
    • Coverage: Global

  • Global extended-range forecasting

    AP2

    Global extended-range forecasting

    AP2
    About
    In this application, ECMWF will use the WeatherGenerator to produce extended-range forecasts using a fine-tuned forecasting decoder. As an exploratory study, the model will also be run for several years following the AMIP protocol to assess its potential in climate applications. Forecasts will be coupled with sea surface temperature fields from the DestinE Climate Twin to test the robustness of the model when extrapolating into previously unseen climate states.

    Key features:

    • Main developer: ECMWF
    • Time horizon: Up to 7 weeks ahead
    • Coverage: Global

  • Global subseasonal to seasonal probabilistic forecasts

    AP3

    Global subseasonal to seasonal probabilistic forecasts

    AP3
    About
    ECMWF will use the WeatherGenerator to produce probabilistic S2S forecasts globally. These forecasts aim to improve skill on multi-week timescales, particularly for extreme events, by leveraging the multi-resolution core of the WeatherGenerator for multi-week rollouts. The application addresses the generally low skill of current S2S predictions and provides information relevant for climate-sensitive decision-making worldwide.

    Key features:

    • Main developer: ECMWF
    • Time horizon: weeks to months ahead
    • Coverage: Global

  • Seasonal forecasts for Western Europe

    AP4

    Seasonal forecasts for Western Europe

    AP4
    About
    KNMI will focus on subseasonal to seasonal forecasts for Western Europe, using a WeatherGenerator tail network to improve skill for extreme events such as cold spells, heat waves, and droughts. By providing early-stage predictions of high-impact events, the forecasts will support decision-making for multiple societal sectors. The multi-resolution structure of the WeatherGenerator will allow multi-week rollout training for enhanced predictability.

    Key features:

    • Main developer: KNMI
    • Time horizon: weeks to months ahead
    • Coverage: Western Europe

  • High-resolution ensemble weather forecasts for Western Europe

    AP5

    High-resolution ensemble weather forecasts for Western Europe

    AP5
    About
    National Meteorological Services translate complex weather and climate information into actionable insights for stakeholders, a mission that critically depends on accurate forecasts of extreme events. Extreme events are inherently rare, making them difficult to capture with conventional machine learning, but the WeatherGenerator can leverage global data to improve modelling of such events. In this application, KNMI will use the WeatherGenerator to generate high-resolution probabilistic forecasts of extreme weather for Western Europe, by training on regional NWP data, in-situ observations, and gauge-adjusted radar precipitation. The application will monitor forecast skill for extremes and apply corrective measures as necessary, including importance sampling of extreme events in other datasets, such as output from hectometric simulations from DestinE’s On-Demand Extremes Digital Twin. Different tail-network architectures and loss functions will be explored to maximize predictive skill.

    Key features:

    • Main developer: KNMI
    • Time horizon: Up to 2 days ahead
    • Coverage: Western Europe
    • Spatial resolution: 2 km

  • Extreme weather forecasting for France

    AP6

    Extreme weather forecasting for France

    AP6
    About
    Accurate forecasts of high-impact weather events, such as heavy rainfall, thunderstorms, and tropical cyclones, are critical to safeguard lives and property. In this application, Météo-France will fine-tune the WeatherGenerator to produce high-resolution, short-range ensemble forecasts over Western Europe, including French overseas territories, with a focus on extreme events. Both full emulation from initial conditions and downscaling from coarse models will be explored, with attention to lateral boundary conditions and the role of physical constraints in improving forecast consistency. Forecasts will be evaluated against the operational Arome model and a task-specific regional machine learning emulator, using innovative metrics to assess realism and physical consistency.

    Key features:

    • Main developer: Météo-France
    • Time horizon: Up to 2 days ahead
    • Coverage: Western Europe and French overseas territories
    • Spatial resolution: 2.5 km and 1 km

  • 21-day forecasts for the Nordics

    AP7

    21-day forecasts for the Nordics

    AP7
    About
    Accurately forecasting weather in the Nordics is challenging due to wintertime inversions, complex topography, and intricate coastlines. This application will combine a high-density network of crowdsourced weather stations with satellite and radar data to improve existing operational forecasts, producing hourly surface parameter predictions at 1 km resolution. Forecasts will be evaluated against measurements from high-quality weather stations. The goal is to provide more accurate and reliable weather forecasts for the general public.

    Key features:

    • Main developer: MetNor
    • Time horizon: Up to 21 days
    • Coverage: The Nordic region
    • Spatial resolution: 1 km

  • High-resolution forecasts for the Alps

    AP8

    High-resolution forecasts for the Alps

    AP8
    About
    Accurate weather forecasting in mountainous regions is challenging due to complex terrain, wintertime inversions, and systematic model errors in cloud cover. In this application, Meteo Swiss will fine-tune the WeatherGenerator to produce high-resolution (up to 1 km) forecasts for the Alps from full initial conditions. The resulting forecasts will be compared with a task-specific regional machine learning emulator, using in-situ observations to validate inversions and satellite data to evaluate cloud predictions. This approach aims to improve forecast accuracy in complex alpine terrain.

    Key features:

    • Main developer: Meteo Swiss
    • Time horizon: Up to 10 days
    • Coverage: Central Europe, centred on Switzerland
    • Spatial resolution: 1 km

  • Nowcasts and short forecasts of cloud and precipitation using satellite and radar data

    AP9

    Nowcasts and short forecasts of cloud and precipitation using satellite and radar data

    AP9
    About
    High-resolution real-time observations are essential for nowcasting high-impact weather events, but radar, satellite, and ground data are often incomplete or inconsistent. In this application, SMHI will use the WeatherGenerator to produce fused nowcasts and short-term forecasts of cloud and precipitation, combining radar (OPERA) and satellite (NWCSAF) data. Tail networks will generate outputs at the same spatial and temporal resolution as the original products, creating seamlessly blended datasets. The application will assess whether the WeatherGenerator can replace current NWCSAF and OPERA production workflows, potentially transforming operational processing of radar and satellite observations.

    Key features:

    • Main developer: SMHI
    • Time horizon: Current conditions and 12-hour forecasts
    • Coverage: Europe

  • Extreme precipitation nowcasts in sub-Saharan Africa

    AP10

    Extreme precipitation nowcasts in sub-Saharan Africa

    AP10
    About
    Accurate nowcasting of extreme convective rainfall is crucial in regions with limited radar and ground station coverage. In this application, eScience will use the WeatherGenerator to enhance high-intensity precipitation nowcasts across sub-Saharan Africa. By combining WeatherGenerator forecasts at convection-permitting scales with existing satellite-based SEVIRI observations, the application aims to improve the localization and estimation of extreme rainfall. The WeatherGenerator will be fine-tuned with (i) the MSG-CPP precipitation product based on daytime SEVIRI cloud property observations and, additionally, investigate the added value of improving retrieval by finetuning on Euradclim data over the Europe, which fuses radar and rain gauge data in a single gridded product. The resulting nowcasting methods will be compared against TAHMO rain gauge data over the target region. This task will happen in collaboration with the Dutch water management and consultancy firm HKV. Finally, the potential of diffusion-based ensemble techniques to further enhance forecast reliability could be explored.

    Key features:

    • Main developer: eScience
    • Time horizon: Up to 4 hours
    • Coverage: Senegal/Burkina Faso/Ghana
    • Spatial resolution: 3 km

  • 40-year analyses for the Nordics

    AP11

    40-year analyses for the Nordics

    AP11
    About
    Consistent historical weather time series are essential for assessing regional climate signals, but existing archives are often limited in length or costly to produce. In this application, MetNor will use the WeatherGenerator’s data fusion capabilities to create a 40-year, high-resolution reanalysis of standard surface parameters for the Nordic region. By leveraging the WeatherGenerator’s ability to model complex relationships between observation sources and global reanalyses, the application will generate a time-consistent dataset without additional training. The dataset will be evaluated against out-of-sample measurements from weather stations.

    Key features:

    • Main developer: MetNor
    • Time horizon: past 40 years
    • Coverage: The Nordic region

  • High-resolution reanalysis and climatology for the Alps

    AP12

    High-resolution reanalysis and climatology for the Alps

    AP12
    About
    Accurate weather information in mountainous regions is limited by coarse model resolution and complex topography. In this application, the WeatherGenerator will be used to downscale conventional model output over the Alpine region, with a particular focus on topographic effects such as valley winds. This will be used to generate a high resolution climatological regional dataset from the existing ERA5 reanalysis. Such a dataset will be very valuable for applications for renewable energy and will be compared to conventional downscaling approaches.

    Key features:

    • Main developer: Meteo Swiss
    • Time horizon: Current conditions
    • Coverage: Central Europe, centred on Switzerland
    • Spatial resolution: 1 km

  • Climate Prediction

  • Spatio-temporal downscaling of climate extremes

    AP13

    Spatio-temporal downscaling of climate extremes

    AP13
    About
    Accurately representing climate extremes at regional scales remains a major challenge for coarse-resolution climate models. In this application, CMCC will use the WeatherGenerator as a spatio-temporal downscaling tool to produce ensembles of high-resolution decadal forecasts with more realistic extremes. Historical simulations and initialised daily decadal predictions will serve as input, and the added value of downscaling will be assessed by evaluating temperature and precipitation extremes against observations at the weekly timescale. The use of ensembles will allow for more robust statistics of extreme events, while the increased resolution and the WeatherGenerator implicit physics are expected to improve the realism of the events.

    Key features:

    • Main developer: CMCC
    • Task: Downscaling
    • Time horizon: Up to 7 days
    • Coverage: Europe

  • Forecasting Arctic sea-ice

    AP14

    Forecasting Arctic sea-ice

    AP14
    About
    Understanding and predicting Arctic sea ice is essential for climate monitoring and polar operations, but is challenged by complex interactions between atmospheric, oceanic and cryospheric processes. In this application, the WeatherGenerator will use satellite and model-based data to forecast Arctic sea ice. The approach will integrate key drivers of ice dynamics, such as surface temperature, snow depth, salinity and large-scale climate indicators. The application aims to enhance ice prediction accuracy, while also strengthening process understanding through the identification of previously unrecognized relationships governing Arctic sea ice variability.

    Key features:

    • Main developer: CMCC
    • Time horizon: Up to 7 days
    • Coverage: Arctic Ocean sector
    • Spatial resolution: ~10 km

Our partners

A Europe-wide alliance of weather agencies, HPC centres, universities and industry pioneers driving an open next-generation climate model.

Europe Map

Météo-France

France
About
Météo-France, based in Toulouse, is the French National Weather Service, providing weather forecasts, alerts, and guidance while advancing research on weather and climate through its CNRM joint research unit.
Role in WeatherGenerator
  • Applying the WeatherGenerator model to produce high-accuracy predictions for Western Europe
  • Contributing to the development of next-generation, AI-driven forecasting models
  • Integrating research into operational weather prediction for extreme events and public services
  • Collaborating with top-level partners to advance data-driven weather forecasting

MPI for Biogeochemistry

Max Planck Institute for Biogeochemistry

Germany
About
The Max Planck Institute for Biogeochemistry in Jena, Germany, is a leading research institute studying Earth system and biogeochemical processes using advanced measurements and modeling.
Role in WeatherGenerator
  • Building a foundation model of land ecosystems
  • Integrating measurements from experiments, in-situ stations, and Earth observations
  • Applying advanced modeling, including machine learning, to understand ecosystem-climate interactions
  • Supporting early warning systems for extreme climate events like droughts and heatwaves

LT

Latest Thinking

Germany
About
Latest Thinking is a science communication company that helps institutions and funded projects make research accessible and engaging. They provide tailored communication strategies and produce high-quality content across multiple formats - including videos, animations, infographics, podcasts, reports, and interactive media - bridging the gap between science and society.
Role in WeatherGenerator
  • Developing the project’s visual identity, including logo, color scheme, and website
  • Producing video content such as project intros, work package leader statements, science explainers, and researcher spotlights
  • Managing ongoing website updates and enhancements
  • Supporting communication, planning, coordination, reporting, and social media management

FZJ

Forschungszentrum Jülich

Germany
About
Forschungszentrum Jülich is a leading German research center leveraging high-performance computing to tackle global scientific challenges.
Role in WeatherGenerator
  • Developing the core models of WeatherGenerator through the ESDE group
  • Designing training protocols and maintaining the software framework
  • Optimizing computations for high-performance supercomputers
  • Fine-tuning models to handle diverse and complex datasets

CMCC

Centro Euro-Mediterrane sui Cambiamenti Climatici

Italy
About
The CMCC Foundation (Euro-Mediterranean Center on Climate Change) is an international, independent, multidisciplinary research center that includes climate modelling and the interaction between climate change and society.
Role in WeatherGenerator
  • Contributing to WeatherGenerator model architecture and development of applications for extreme events and sea-ice prediction
  • Improving Arctic sea ice predictions by correcting systematic errors using data assimilation
  • Providing cutting-edge machine learning solutions for model development
  • Leveraging high-performance computing to enhance WeatherGenerator models
  • Supporting the creation of a world-leading generative foundation model for the Earth system

Netherlands eScience Center

Netherlands
About
The Netherlands eScience Center, based in Amsterdam, is the Dutch national center of expertise in research software, supporting academic research through innovative and sustainable software development.
Role in WeatherGenerator
  • Connecting WeatherGenerator to the broader European geoscientific community
  • Embedding WeatherGenerator into existing scientific and operational workflows
  • Extending and improving data-driven solutions using the WeatherGenerator atmospheric foundation model

KNMI

The Royal Netherlands Meteorological Institute

Netherlands
About
The Royal Netherlands Meteorological Institute (KNMI) in De Bilt provides research, forecasts, and warnings in meteorology, climate, and seismology to ensure a safe and livable Netherlands.
Role in WeatherGenerator
  • Contributing to the development of the WeatherGenerator
  • Developing tail networks for improved forecasting
  • Enhancing high-resolution probabilistic extreme weather forecasts
  • Improving probabilistic sub-seasonal to seasonal forecasts

Statkraft

Norway
About
Statkraft is Europe’s largest producer of renewable energy, specializing in the development, operation, and ownership of hydropower, wind, solar, gas, and biomass assets, while also supplying district heating and trading energy across more than 20 countries.
Role in WeatherGenerator
  • Applying and validating WeatherGenerator results in real-world operations
  • Using weather forecasts for multi-year price forecasting, intraday production planning, and trading
  • Leveraging large models and reference data to integrate WeatherGenerator outputs
  • Supporting the project with practical energy sector applications

MetNor

Norwegian Meteorological Institute

Norway
About
MetNor is Norway’s national meteorological service, providing weather information to support public safety, economic activity, and environmental protection, with offices in Oslo, Bergen, and Tromsø.
Role in WeatherGenerator
  • Leading Theme 3 on applications
  • Developing two applications (AP7 and AP11) for the project
  • Integrating research advances into operational weather products
  • Enhancing public weather forecasts, including for the Yr app with over 14 million users
  • Collaborating with partners to share expertise and develop innovative applications

KAJO

Slovakia
About
KAJO is a Slovak-based consultancy that develops AI- and data-driven solutions for disaster risk management, environmental monitoring, and climate resilience.
Role in WeatherGenerator
  • Leading the global flood forecasting use case
  • Integrating AI-driven meteorological data into models for improved flood accuracy, depth estimation, and delineation
  • Combining predictive analytics with real-time satellite data
  • Feeding outputs into the GLOFAS system under Copernicus EMS for long-term adoption
  • Advancing rapid risk assessment tools to strengthen global resilience, including support for the UN’s Early Warnings for All initiative

SMHI

Swedish Meteorological and Hydrological Institute

Sweden
About
SMHI, the Swedish Meteorological and Hydrological Institute, provides science-based weather, water, and climate forecasts while conducting research in meteorology, hydrology, oceanography, and climate, with offices in Norrköping, Gothenburg, and Uppsala.
Role in WeatherGenerator
  • Contributing to WP5 and WP6 with expertise in nowcasting from satellite and radar data
  • Applying machine learning and generative methods for precipitation forecasting
  • Bringing experience in generative numerical weather prediction techniques
  • Exploring AI/ML applications to enhance operational decision support for society

MeteoSwiss

Switzerland
About
MeteoSwiss is Switzerland’s federal meteorology and climatology office, providing weather and climate data, forecasts, warnings, and research-driven services.
Role in WeatherGenerator
  • Developing the core WeatherGenerator model
  • Working on two applications: high-resolution weather forecasting over the Alps and downscaling re-analysis to finer resolution
  • Generating a new km-scale dataset with the ICON model for training and improving applications
  • Integrating advanced physical modeling to enhance weather and climate predictions

ETH Zurich

Eidgenössische Technische Hochschule Zürich

Switzerland
About
ETH Zurich is a world-leading research university in Switzerland, advancing climate science and scalable computing to tackle complex Earth system challenges.
Role in WeatherGenerator
  • CSCS: Developing a data transfer service to efficiently move large datasets between supercomputing centers
  • SPCL: Creating compression schemes for weather and climate data
  • Supporting scalable computing solutions and high-performance data handling
  • Leveraging expertise in parallel computing, storage, and data access for the project

Buluttan

Turkey
About
Buluttan is Turkey’s first private meteorology-technology company delivering AI-powered, ultra-local weather intelligence to enhance decision-making across sectors like renewable energy, aviation, logistics, and utilities.
Role in WeatherGenerator
  • Developing robust AI-driven weather models to support Europe’s sustainability strategies
  • Delivering reliable energy production forecasts to reduce operational costs and support renewable energy investment
  • Demonstrating the financial impact of models in renewable energy and increasing grid stability
  • Contributing to Europe’s transition to sustainable energy solutions

Met Office

United Kingdom
About
The Met Office, based in Exeter, UK, is the national meteorological service providing weather and climate services to support public safety, business, and policy decisions, with 160 years of expertise in global weather science.
Role in WeatherGenerator
  • Evaluating WeatherGenerator outputs using scientific expertise in forecast model assessment
  • Applying eXplainable AI (XAI) to understand and build trust in model outputs
  • Contributing to data compression and dataset optimization for efficient model training
  • Enhancing the handling of very large and diverse datasets for WeatherGenerator

ECMWF

European Centre for Medium Range Weather Forecasting

United Kingdom
About
ECMWF is an international organization providing world-leading weather forecasts and climate data, while advancing innovation in Earth system science.
Role in WeatherGenerator
  • Coordinating the entire project across all partners
  • Exploring the potential of foundation models for weather and climate applications
  • Bringing expertise from operational forecasting and climate science
  • Leveraging resources and collaborations for maximum project impact

Community

The WeatherGenerator project is bringing together partners from various fields related to weather and climate modeling and research. Through a series of internal and external hackathons, core and application developers, as well as external innovators collaborate and build community. These collaborative coding events focus on embedding the WeatherGenerator into applications, testing data APIs, sharing development experiences, and enabling external users to integrate the system with their own weather-related projects.

View in Events

Synergies

WeatherGenerator is developed in close alignment with major European research and innovation initiatives. By fostering synergies with projects such as TerraDT, UrbanAIR, and Destination Earth, the WeatherGenerator contributes to a shared vision of advancing high-resolution climate modeling, digital twins, and data-driven environmental intelligence. These collaborations enhance interoperability, promote knowledge exchange, and maximize the impact of European investments in climate resilience and sustainability.
  • Destination Earth
    Destination Earth
    Destination Earth
    https://destination-earth.eu

    Funding: EU DEP funding, Horizon Europe

    Runtime: Launched in 2021

    Synergies:
    While the WeatherGenerator will train on the km-scale data of the Extremes Digital Twin and the Climate Digital Twin of Destination Earth, WeatherGenerator also provides a machine-learned foundation model for Earth sciences to the Destination Earth programme.  The WeatherGenerator will be applicable for many different application areas in weather and climate science, for both global and local use cases, and will run natively from the snapshots of the digital twins of Destination Earth. WeatherGenerator will therefore serve as additional machine-learned Digital Twin for the Destination Earth user community. WeatherGenerator will also link to the machine-learned ocean model that is developed in Destination Earth as part of a machine-learned Earth system model.
  • TerraDT
    TerraDT
    TerraDT
    https://terradt.eu

    Funding: Horizon Europe

    Runtime: 2025-2028

    Synergies:
    TerraDT will enhance the Climate Digital Twin of Destination Earth. This will bring direct benefits to the WeatherGenerator that is partly trained on the datasets developed for the Climate Digital Twin. On the other hand, the WeatherGenerator could likely be used as cheap emulator for the more expensive simulations of the Climate Digital Twin, and the enhancements of the atmospheric chemistry from TerraDT within the Climate Twin will potentially help to add atmospheric chemistry fields as outputs to the WeatherGenerator.
  • UrbanAIR
    UrbanAIR
    UrbanAIR
    https://www.urbanair-project.eu

    Funding: Horizon Europe

    Runtime: 2025-2028

    Synergies:
    UrbanAir is an ambitious programme that will close the gap between global and street-level simulations for both physical and machine-learned models, therefore covering many orders of magnitude in resolution. WeatherGenerator has the ambition to be applicable for both global simulations at coarse and limited area simulations at high resolution. We therefore will explore whether the tool can be used as alternative modelling infrastructure for the use-cases of UrbanAir.