The rapid proliferation of data in the new information era has increased the complexity of data-driven problems across various fields of science and engineering. This development has led to a paradigm shift in Machine Learning, moving towards unsupervised and self-supervised representation learning, as well as multimodal learning. Significant advancements have emerged not only in mainstream Natural Language Processing and Computer Vision but also in Earth Observation applications. These advancements exploit the synergies between self-supervised learning and the expanded availability of supercomputing systems, resulting in the emergence of Foundation Models (FMs).

To mitigate the effects of global climate change, we need comprehensive knowledge about the global carbon budget, which comprises CO2 sources and sinks such as wetlands, forests, and permafrost soils. Until now, researchers have struggled to quantify how changes in land areas, vegetation, or soils affect the carbon cycle due to heterogeneous and scattered data. The project Towards Global 3D Above and Below Ground Carbon Stocks (3D-ABC) aims to build a new generation of geospatial FMs. The goal is to integrate and model data from various sources such as satellites, drones, and local CO2 monitoring stations. This approach will allows key parameters of the global carbon cycle of vegetation and soils to be captured, quantified, and characterized with high spatial resolution. The final objective of 3D-ABC is to achieve a seamless quantitative understanding of the current state of carbon stocks in forests and soils.

FZJ is responsible for the deployment of the FMs on HPC systems, particularly on the upcoming first European Exascale supercomputer, JUPITER. This will facilitate exceptionally high performance for AI training and deployment at the scale needed for 3D-ABC. FZJ will exploit integration with existing geospatial FM models, carry out pre-processing of core datasets, and develop large-scale AI with innovative distributed training strategies.