EarthGenerator
EarthGenerator a foundation model for Earth system modelling
The project
EarthGenerator is a Research and Innovation Action funded by the Horizon Europe programme, developing a foundation model of the Earth system. Extending the WeatherGenerator model, it integrates atmosphere, ocean and land into a single, physically consistent model that adapts to multiple downstream tasks, from seasonal forecasting to multi-annual climate projections, with minimal additional training.
The work is organised around four themes: curation of AI-ready datasets; model development and training, including a module enforcing physical constraints and scalable HPC training; use-case applications covering climate extremes, the land carbon cycle, early warning of food insecurity and climate-induced migration; and engagement and dissemination for open science and community adoption. The consortium brings together thirteen partners from across Europe and Canada, coordinated by CMCC, over 48 months.
What MEEO does
MEEO contributes its expertise in multi-source Earth Observation data exploitation and harmonisation, integrating satellite observations with in-situ, reanalysis and socio-economic datasets to build climate stress benchmarks for applied scenarios, with 26 person-months in total. Its work is concentrated in the development of task-specific models (WP3) and their application (WP8), and it leads the prototype deliverable on food insecurity models and the final cross-task impact assessment report.
In the food security use case, MEEO explores how EarthGenerator predictions can further evolve desert locust monitoring, an early warning service that forecasts swarm formation and migration across sub-Saharan countries. In the migration use case, it analyses the dynamics of informal settlements and refugee camps from Earth Observation and climate data, assesses their climate-related drivers, and provides benchmarks that support and validate the development of the foundation model itself.