Weather, Oceans, and AI: HClimRep Annual Meeting 2025 in Cologne
On 1 and 2 September, the HClimRep team gathered in Cologne for the annual project meeting. Nearly 20 researchers from different Helmholtz centers participated in two days of lively discussions on progress, challenges, and the road ahead.
The aim of HClimRep is ambitious: to develop one of the first AI foundation models for climate research. By combining data from the atmosphere, ocean, and sea ice, we are working toward one of the most precise weather and climate models worldwide. The model will eventually run on Europe’s first exascale computer, enabling complex “what-if” experiments and faster, more detailed climate projections.
The project is a collaborative effort across Helmholtz institutions: the Jülich Supercomputing Centre, the Alfred Wegener Institute in Bremerhaven, Helmholtz-Zentrum Hereon in Geesthacht, and the Karlsruhe Institute of Technology. Together, we bring expertise in atmospheric and ocean sciences, stratospheric processes, machine learning, and high-performance computing.
A central topic at the meeting was the WeatherGenerator prototype, which integrates atmosphere, ocean, and stratosphere into a single framework. This prototype is becoming the backbone of our work, but it also highlights the main challenges we face. Achieving stable forecasts over longer timescales is still a bottleneck, reflecting the fundamental limits of current modeling approaches. We also see constraints on computing resources and data infrastructure, which require creative solutions and close coordination across the consortium.
Accepting these challenges, the meeting showed that we are making solid progress. New datasets are being created that allow us to better couple atmospheric and oceanic processes. First tests of stratospheric chemistry tracers have been carried out, and alternative modeling strategies - such as diffusion-based approaches - are being advanced in parallel. Stronger evaluation methods and more structured collaboration across teams are also helping us to move forward.
Looking ahead, we agreed to focus our efforts on refining the WeatherGenerator, with the goal of achieving a stable release by 2026. Next milestones include developing monthly-to-seasonal rollouts, integrating climate forcing, and designing downscaling applications that connect large-scale climate models to regional impacts. These steps will help us better understand how the climate system responds under different scenarios and provide insights that are both scientifically robust and societally relevant.