RAINA - High-resolution nowcasting of precipitation and wind extremes with a foundation model for the atmosphere

Data-driven weather prediction models based on deep learning have been on the rise for several years and have outperformed traditional physics-based numerical models in various benchmark forecasting scores. However, a significant challenge remains: accurately predicting extreme events on a local scale, such as thunderstorms and wind gusts. Previous models struggle in this area, as they were primarily developed for medium-range forecasting and operate at relatively coarse spatio-temporal resolutions. However, the capability of weather models to predict extreme events at a local level is essential for preventing severe consequences for communities, ecosystems, and the financial and material losses they entail. Recently, task-agnostic foundation models, trained on extensive and diverse datasets using self-supervised methods, have demonstrated remarkable skill and robustness, especially in their ability to generalize to rare extreme events. 

The RAINA project aims to develop a foundation model for the atmosphere, with an emphasis on delivering reliable, high-resolution forecasts of extreme wind and precipitation events. In partnership with the EU Horizon-funded WeatherGenerator project, which aims to create advanced digital twins for Destination Earth, RAINA will extend the pioneering AtmoRep model (Lessig et al., 2023) by employing a multi-modal learning approach.
The foundation model seeks to develop a comprehensive, statistically robust, and multi-scale understanding of atmospheric dynamics by incorporating a wide range of meteorological datasets from both models and observations. Innovative deep learning methods, including diffusion models and test-time adaptation, will be investigated to facilitate short-range forecasts of temperature, wind, and precipitation at kilometer-scale resolution over Germany.

In a first demonstrator, short-range forecasts are generated using the AtmoRep model and subsequently refined with the CorrDiff downscaling approach (Mardani et al., 2024) that combines a generative diffusion model with a residual approach. This two-step strategy delivers high-resolution forecasts with a maximum lead time of six hours while disentangling uncertainties inherent in the forecasting and downscaling processes, a separation that can enhance training quality when properly applied. By using ERA5 and COSMO REA2 reanalysis data, the approach enhances the precision of high-resolution forecasts over Germany. 
Initial results from the first demonstrator will be presented in a poster, along with the overall timeline and key milestones of the RAINA project.