Motivation:
Reliable weather and climate forecasts and the associated adaptation strategies are more important than ever for society and politics. However, accurate forecasts across all possible natural phenomena are essential for this. However, short-term forecasts and extreme events are particularly challenging. The resolution of the models and data also plays a decisive role here. This is precisely where the targeted use of artificial intelligence (AI) methods has a promising effect and is being investigated and advanced in the “RAINA” research project.
Goals and approach:
The aim is to develop the world’s first flexible, resilient and efficient basic AI model of the atmosphere based on deep learning. To this end, the spatial resolution of the model is to be refined from the current approx. 25 km to around 1 km. The prototype will be set up for weather events caused by wind and precipitation. This approach is particularly demanding on both a scientific and technical level. The specific further development and implementation of the AI processes and the large amount of different data from the entire Earth system have to be managed and merged.
Innovations and perspectives:
The main result of RAINA will be a so-called “Foundation Model”, which offers a novel probabilistic description of the dynamic atmosphere. This should make it possible to quantify both the intrinsic uncertainty of a forecast and that of the simulation model itself. The prototype application will be demonstrated using examples of high-resolution weather forecasts for up to 24 hours and with unprecedented quality. The strong consortium ensures broad and high-profile utilization.