EarthGenerator will deliver a foundation model of the Earth system, extending the WeatherGenerator model by integrating atmosphere, ocean, and land in a single, physically consistent model. By adopting a generative AI approach, EarthGenerator will provide a general-purpose capability adaptable to multiple downstream tasks from seasonal forecasting to multi-annual climate projections with minimal additional training.

General Objectives

– Collect the leading datasets in Earth system science, encompassing selected observations, analyses, and reanalyses, as well as the output of conventional Earth system model simulations and ancillary datasets. This will facilitate the understanding of climate-environment-society-economy correlations.

– Build the EarthGenerator as an extension of the WeatherGenerator to longer lead times and more Earth system components, to create a Foundation Model of the Earth system. It will be based on representation

– learning to use the full potential of Europe’s pre-exascale and exascale machines of EuroHPC.

– Validate and apply the EarthGenerator in strategic application domains, at both global and regional scales, over time scales from seasons to a few years. Among the case studies, we will consider extreme events in the ocean and atmosphere, the land carbon cycle, as well as impacts of climate change on food security and human migrations.

– Engage with the broader scientific community to maximize impact and adoption of the EarthGenerator, and to facilitate its application by external groups across Europe for a wide range of scientific and operational uses.

CMCC role
CMCC is the coordinator of the project.

Expected Results

1. Develop the best Foundation Model of the Earth system to advance the understanding of the planet’s complex dynamics, enabling forecasting, reconstruction, downscaling, and simulation across seasonal to multi-annual timescales and different resolutions.
2. Create a Generative AI-based Foundation Model for the Earth system, building upon the existing WeatherGenerator model.
3. Create a foundation model of the Earth system, analogous to large language models in NLP but tailored to Earth sciences.
4. Improve forecasts for high-impact weather and long-term climate patterns; develop a prototype for early warning systems (EWSs) that predict food (in)security; integrate multimodal datasets to forecast human mobility.
5. Gather diverse datasets