Registration now open for the WeatherGenerator Community Hackathon
The WeatherGenerator Community Hackathon
📍 Location: Max Planck Institute for Biogeochemistry, Jena, Germany
📅 Date: 26–28 January 2027
📝 Registration: 1 September – 1 November 2026, via REGISTRATION
Join our first Community Hackathon and become part of the WeatherGenerator community!
The WeatherGenerator project will host its first Community hackathon on 26–28 January 2027 in Jena, Germany. This in-person, hands-on event will bring together researchers and professionals working across Earth System sciences to explore the WeatherGenerator, exchange ideas and become part of its growing community. The goal is that each participating team can run and fine-tune the WeatherGenerator for its own application while learning from and contributing to the community around the open-source model. The WeatherGenerator is an open-source European foundation model of the Earth system aiming to provide new machine-learned Digital Twin for Destination Earth. It is developed by 16 partners in a Horizon Europe project coordinated by ECMWF. The WeatherGenerator learns an unified latent representation from a wide variety of Earth system datasets - reanalyses, model output, satellite products and in-situ observations - using a self-supervised, masked-token approach. Rather than training a separate model for their task, users can use this general WeatherGenerator foundation model and fine-tune it towards the outputs that matter for their application.
What the WeatherGenerator can do
The model is task-independent and covers a broad range of downstream use cases. It can produce forecasts for nowcasting, medium range to sub-seasonal and seasonal time scales with probabilistic ensemble outputs. It can be used for downscaling of forecasts, reanalyses and climate projections to kilometre scale and finer, including over complex terrain, and for data fusion, where incomplete or inconsistent observational sources are combined into consistent gridded products. Because decoder networks can be trained on any target quantity the user provides, the WeatherGenerator is not limited to standard meteorological variables. Applications within the project already address river flow, renewable energy production and consumption, sea ice, land surface temperature and land–atmosphere fluxes of carbon, water and energy. Very high resolution land and remote sensing applications are addressed through WeatherGenerator-Land, the projects’ land surface model trained on satellite observations.
The programme
The WeatherGenerator Community Hackathon will open with an introduction and a tutorial on the WeatherGenerator core model, its architecture, training approach and data formats. This is followed by short presentations in which each team introduces its application. The remainder of the event is dedicated to hands-on work on the applications, supported throughout by core model developers and by application scientists from across the consortium. The hackathon concludes with final presentations from all teams on the afternoon of the last day. A social programme will accompany the event.
Who can take part
The event is open to teams of two to five people from research institutes, national meteorological and hydrological services, agencies, SMEs and industry, working on weather, climate, water, energy, agriculture, ecosystems or health. Teams from regions and application areas that are currently underrepresented in machine-learning-based Earth system modelling are particularly encouraged to apply. Teams are expected to arrive with a well-defined target application — for example weather forecasting for a particular country, downscaling of a given variable over a defined region, or predicting an impact variable relevant to their sector. Data for the application should be prepared in advance if possible. Technical guidance on data formats will be shared with teams after acceptance. Familiarity with Python and machine learning is expected; prior experience with foundation models and the WeatherGenerator is not.
An overview of current WeatherGenerator applications can be found here.
What the project provides
The tutorial and training material, technical mentoring, access to pretrained checkpoints and the compute resources needed to fine-tune the WeatherGenerator during the event will all be provided by the project. Participation is free of charge but travel and accommodation is self-funded. A limited number of travel grants may be available on a needs-based basis to support participants who could not otherwise attend; interest can be indicated in the registration form. Invitation letters for visa applications can be issued on request.
Registration and selection
Registration opens on 1 September 2026 and closes on 1 November 2026 and is available via the REGISTRATION page. One registration is submitted per team, describing the intended application, the data available and in which format, the team’s experience with machine learning, any previous contact with the WeatherGenerator, the proposed team members, and, if applicable, the need for travel support. Members of accepted teams will be asked to complete individual registrations afterwards. Places are limited and registration is moderated by the organisers. Registrations will be reviewed on a rolling basis from 1 October 2026, so early submission is encouraged. Teams will be selected on the technical feasibility and readiness of the proposed application, its pathway to impact, its fit with the goals of the WeatherGenerator project, the potential for building lasting capacity, the diversity of teams, and the complimentary of the target applications across the event as a whole. Priority will be given to teams not already involved in the WeatherGenerator consortium or in closely related projects.
More information on this event including a preliminary agenda is available on the Event page.
Further information on the project is available at weathergenerator.eu.
