Content: 

Students will learn how to work with state-of-the-art AI models for climate science and weather forecasting. 

 For example, typical AI models will include recent releases of

 ·         Foundation models for climate science and weather forecasting.

·         Generative AI models for tasks such as ensemble generation of weather forecasts and of climate change simulations for uncertainty quantification.

·         Transformer and graph neural network models for weather forecasting.

·         Climate model emulators.

 Each student will be able to select from a variety of topics to explore in their practical experiments. These could include, but are not limited to:

 ·         The representation of physical concepts in data-driven AI models (e.g., does the model indirectly learn to “understand physics”?).

·         Detecting and understanding failure modes of AI models.

·         Forecast accuracy and uncertainty quantification for AI-generated ensembles of simulations.

·         Effective solutions to post-processing AI results and/or to modifying AI model architectures.

·         Assessing if certain AI architectures perform significantly better for specific tasks.

Workload: 

In-person introductory session, individual and group meetings, final presentation sessions: 30h

Practical tasks – getting started, implementation, experiments, analysis: 100h

Write up results in the style of a scientific paper and preparation of final presentation: 50h