WeatherGenerator Hackathon 2025: Connecting Developers and Application Teams
The WeatherGenerator is a four-year EU project with 16 international partners, working together to build a foundation model for weather forecasting. WeatherGenerator provides a flexible, pre-trained architecture that can be fine-tuned for a wide variety of applications – from medium- and long-range climate projections to high-resolution forecasts.
From 27-29 August, the European Centre for Medium-Range Weather Forecasts (ECMWF) hosted the first not-only-developer WeatherGenerator Hackathon in Bonn - a milestone event where model developers and application teams met face-to-face for the first time. Over three days, participants collaborated on hands-on challenges, explored the potential of WeatherGenerator across different use cases, and exchanged perspectives on how to make the model truly useful for the wider community.
Talks and Challenges
The hackathon began with introductions from 22 application leads, followed by two challenge tracks:
- Medium- and long-range forecasting
- Short-range, high-resolution forecasting
On Day 1, Christian Lessig (project lead) provided an overview of the WeatherGenerator architecture, setting the stage for practical sessions. Day 2 included presentations on general software development tooling and Machine Learning Operations (practices for deploying and managing machine learning models at scale), as well as a discussion on extending WeatherGenerator to downstream applications. On Day 3, the event concluded with group discussions, experience sharing, and short presentations from participants.
To accelerate progress, participants were given pre-trained model checkpoints for both challenges, later extended with fine-tuned versions to enable direct experimentation with inference and evaluation plots. Many tried their hands at both challenges, gaining valuable insights into how datasets such as CERRA, SYNOP, SEVIRI, and OPERA could be integrated into the training process. With HPC resources reserved on the supercomputer JURECA, teams could test their workflows on real infrastructure efficiently.
Perspectives from Participants
For the ESDE (Earth System Data Exploration) team, the hackathon was a chance to support application partners in setting up the WeatherGenerator environment and tutorials – but also to learn. Some team members’ quotes on the event:
“Talking to application teams gave us a better understanding of how they want to use the pre-trained WeatherGenerator architecture. We also gained background knowledge about reanalysis, post-processing, and operational workflows.” – Jifeng Wang
“Hearing the talks from ECMWF helped me see WeatherGenerator in a broader perspective – not just as another model, but as a foundation model that could enable downstream applications in ways that current state-of-the-art systems like Pangu or Aurora cannot.” – Jifeng Wang
“For me, the hackathon was more about connecting with application developers and understanding their needs. It was also a great way to stress-test our guides and onboarding workflows for WeatherGenerator.” – Simon Grasse
And of course, not everything was code and compute – the group also enjoyed pizza together on the first evening, fueling lively discussions.
Looking Ahead
The hackathon was more than a technical exercise – it marked the start of a closer collaboration between WeatherGenerator’s core developers and its diverse community of application partners. The gathered feedback will guide the next steps in model development, from exploring new embedding and decoding architectures to refining fine-tuning strategies.
The key takeaway: WeatherGenerator is not just about building a foundation model – it’s about building a community around a model.
More information about the hackathon: weathergenerator.eu/first-internal-hackathon