In late November, scientists from the Jülich Supercomputing Centre (JSC), the European Centre for Medium-Range Weather Forecasts (ECMWF), the German Weather Service (DWD), and the University of Bonn gathered for the annual RAINA project meeting. Hosted at the University of Bonn, the event combined a retrospective on 2025 with an outlook on the year ahead. Also joining for the first time was a new contact from the project’s responsible organisation at VDI/VDE.
As part of the broader WeatherGenerator initiative, a four-year EU project, RAINA contributes to improving extreme precipitation and storm forecasting. The team can look back on a productive year in which several key tasks were successfully advanced:
- Prototyping the WeatherGenerator to handle multi-modal inputs, including observation-based datasets such as IMERG, and experimenting with embedding networks and dataset-specific preprocessing strategies.
- Developing test-time adaptation methods.
- Collecting relevant datasets and advancing the downscaling–forecasting application.
- Adapting operational verification routines for RAINA and linking them with machine-learning model outputs.
The Jülich Supercomputing Centre (JSC) leads the core RAINA-specific development. This includes defining dedicated evaluation metrics for extreme-event performance and developing new loss functions that help the model better capture high-impact weather. Current work focuses on a new pre-training approach based on self-supervised learning (SSL), a methodology in which a model learns patterns from the data itself without relying on labelled samples. This strategy aims to improve spatial and temporal skill and strengthen the representation of extremes.
These efforts come in response to limitations observed in earlier masked-token pre-training, where weak precipitation was often overestimated, drizzle was exaggerated, and strong or extreme events were significantly underrepresented. By analysing the statistical behaviour of the datasets and exploring how different transformations affect distributions, such as standard transformer encodings versus Mixture-of-Experts (MoE) approaches, the team is addressing these shortcomings.
Researchers from the University of Bonn also reported substantial progress. Their contributions include:
- MSyncVP: Joint diffusion for synchronous multi-modal video prediction.
- Test-time adaptation via diffusion-noise optimisation for sequence-adaptive video prediction in continuous streams.
- RiverMamba: A state-space model for global river discharge and flood forecasting.
Weather services likewise contributed essential components: DWD prepared and integrated key datasets such as ICON-DREAM and the COSMO-REA2 reanalysis for model training, while ICON-D2-EPS served as comparison data. Next year, they will select test episodes and explore additional datasets for verification. The adaptation of operational verification routines is completed; methodological extensions will follow in 2026.
Looking ahead to 2026, the project will tackle:
- Fine-tuning the WeatherGenerator for targeted forecasting and downscaling applications.
- Comprehensive evaluation cycles to support further model development, with special emphasis on extreme events.
- Development of adapted data-sampling and loss strategies to better represent rare, high-impact phenomena.
- Short-range, high-resolution forecasting demonstrations using the WeatherGenerator.
The team is excited about the progress so far – and even more about what’s to come.
