Can Foundation Models for the atmosphere such as the WeatherGenerator overcome the limits of today’s Machine Learning and physics-based systems in predicting extreme weather?
How can the WeatherGenerator unlock very high-resolution forecasts of temperature, wind, and precipitation?
What training strategies make Machine Learning-based predictions of extreme weather statistically more robust?
Can we trust Foundation Models to anticipate rare and dangerous weather events before they strike?
Extreme weather events—like the devastating Ahr Valley flood of 2021—pose growing threats to lives, infrastructure, and ecosystems. Accurate forecasting of such events is a critical scientific and societal challenge.
While AI and machine learning are advancing the field, current models often struggle with key issues, such as underestimating peaks in wind and precipitation. The RAINA project sets out to change that.
Building on the WeatherGenerator model, one of the first Foundation Models for atmospheric dynamics, RAINA aims to develop a next-generation weather model capable of forecasting at an unprecedented 1 km resolution.
By integrating diverse data sources, including reanalysis and remote sensing data as well as SYNOP observations, and the combined expertise of the European Centre for Medium-Range Forecasting (ECMWF), German Weather Service (DWD), Forschungszentrum Jülich, and the University of Bonn, RAINA will explore cutting-edge strategies such as diffusion-based modeling. The ultimate goal: a prototype system delivering precise, short-term predictions of temperature, wind, and precipitation that outperform today’s operational models.
Project Partners
