Applying AtmoRep for Diverse Weather Applications

Machine learning has recently seen a rapid wide-spread adoption across various fields of science including atmospheric and weather research. The emergence of foundation models has marked a transformation in the science of machine learning. These foundation models are general-purpose models trained on huge amounts of data using self-supervised methods, eliminating the need for labeled data. Once trained, the parameters of these models can be utilized as a starting point for a range of domain-specific tasks. This approach is advantageous in terms of both cost and performance, as it minimizes the reliance on annotated data compared to models trained from scratch. Motivated by this, our study explores the foundational capabilities of AtmoRep, a stochastic atmospheric foundation model, for two distinct weather-related applications, data compression and statistical downscaling. The training of the 3.5 billion parameter AtmoRep model consumed about a few weeks of compute time on 32 JUWELS Booster nodes.