Appears in collection : School/Workshop: Energy, mathematics, and theoretical challenges

Artificial intelligence, and particularly deep learning, revolutionised numerical weather prediction (NWP) in 2023. Several teams from giant tech companies have proposed surrogate models for high-resolution global atmospheric dynamics. These models achieve the performance levels of the deterministic IFS of the European Centre for Medium-Range Weather Forecasts, as well as its ensemble prediction variant. In this presentation, I will discuss the techniques used to construct these models, their scope and limitations, and illustrate the concepts with our own models and results, in NWP and sea-ice models for climate. I will also discuss the integration of such surrogate models with data assimilation for the improvement of NWP, as well as some more fundamental issues related to the end-to-end approaches to data assimilation.