REPRESENTATION LEARNING FOR ATMOSPHERIC DYNAMICS

The AtmoRep project is a joint effort between climate scientists and CERN physicists. Its goal is to improve our understanding of atmospheric dynamics and how it changes over time, starting from the large amount of observational data accumulated in the last 70 years. This approach is made possible by the newly released ERA5 reanalysis data set, procured by the European Center for Medium Weather Forecats (ECMWF), that integrates a diverse range of historical measurements into a seamless data set of unprecedented resolution and quality. In this project I have contributed to use state-of-the-art machine learning, specifically transformer architectures, to obtain a data driven description of atmospheric dynamics.