AtmoRep: Large Scale Representation Learning of Atmospheric Dynamics

• Name: AtmoRep: Large Scale Representation Learning of Atmospheric Dynamics , AtmoRep
• EuroHPC machine used: MareNostrum 5
• Topic: Natural sciences

AtmoRep is the first prototype of a probabilistic foundation model for weather and climate predictions. The model is trained on different data sources and used for a variety of applications, from weather forecasting, to downscaling (or super-resolution), spatio-temporal interpolations and scenario generation. More details can be found in the preprint:  https://arxiv.org/abs/2308.13280. The developments are continued as part of the WeatherGenerator project.

The support action of the EPICURE project helped improving AtmoRep’s performance in terms of scalability, inference speed, and data loading, with the ultimate goal of efficiently handling trainings and a more efficient fine-tuning for the different applications.

The primary benefit of the EPICURE support provided to AtmoRep lies in the significant savings of computing time and energy throughout the process.