Bringing together 13 research institutions across 9 countries, the CMCC-led Horizon Europe project EarthGenerator will develop an AI-based foundation model of the Earth system, advancing climate science, food security forecasting, and early warning for climate-induced displacement.

A new line of research applying artificial intelligence to climate science officially begins today. Coordinated by CMCC (Euro-Mediterranean Center on Climate Change), EarthGenerator, a next-generation Foundation Model of the Earth system, is a four-year Horizon Europe project launched on 1 June 2026. It is designed to integrate AI-driven modelling of the atmosphere, ocean, and land into a single, physically consistent framework for climate science and decision-making.

The project brings together 13 institutions from 9 countries, Italy, Germany, the United Kingdom, the Netherlands, France, Spain, Denmark, Finland, and Canada, including research centres, meteorological agencies, universities, and SMEs, combining expertise in climate science, high-performance computing (HPC), and generative artificial intelligence.

EarthGenerator builds on the WeatherGenerator model, an existing AI-based weather prediction system, and extends it into a general-purpose Earth system foundation model. Its central innovation is to represent the atmosphere, ocean, and land within a single architecture, enabling the model to learn interactions across these domains directly, rather than coupling them after the fact as in conventional Earth system models. By adopting a generative AI approach, the project will provide a single, adaptable capability for a range of downstream applications, from seasonal forecasting to multi-annual climate projections, requiring only minimal additional training for each new task.

Foundation models, large AI systems pre-trained on extensive data and then fine-tuned for specific tasks, have reshaped fields such as natural language processing and computer vision. EarthGenerator aims to bring this approach to climate science, offering the research community a shared, open, and physically grounded tool spanning the full complexity of the Earth system.

The project is structured around four interconnected themes:

• Data curation and optimisation, developing AI-ready, FAIR-compliant datasets that integrate satellite observations, reanalysis data, socio-economic indicators, and climate model outputs;
• Model development and training, advancing transformer-based architectures with modules that enforce physical constraints such as conservation laws, integrate ocean and land components, and incorporate causal machine learning to support interpretable analysis, alongside scalable training on European supercomputers (EuroHPC);
• Use-case applications, delivering fine-tuned models for the prediction of climate extremes (heatwaves, droughts, marine heatwaves, urban hazards), quantification of the land carbon cycle at high spatial resolution (~1 km), early warning systems for food insecurity, and analysis of climate-induced migration and displacement;
• Engagement and dissemination, supporting open science, community adoption, and a sustainable exploitation framework beyond the project’s lifetime.

The societal relevance of these use cases is significant. Extreme weather events are increasing in frequency and intensity, with consequences for agriculture, ecosystems, and human mobility. EarthGenerator will deliver risk products across multiple timescales and spatial resolutions, supporting decision-makers in civil protection, food security agencies, and climate adaptation policy.

“EarthGenerator represents a meaningful step forward in how artificial intelligence can serve climate science,” says Prof. Italo Epicoco, Scientific Coordinator of the EarthGenerator project, “by integrating the atmosphere, ocean, and land into a single generative model, we are building a tool that can help societies anticipate and respond to a warming world.”

The project will use Europe’s leading HPC infrastructures, including Jupiter, MareNostrum5, Leonardo, LUMI, and Alps, to train models at billion-parameter scale, and will align its data infrastructure with the European Open Science Cloud (EOSC) and Destination Earth (DestinE) initiatives. In line with its open science commitment, the model’s training data will be openly accessible. EarthGenerator will also provide the scientific community with an LLM-based interpretation interface, making Earth system model outputs more accessible to experts, decision-makers, and non-specialist users across sectors from transport to agriculture to humanitarian response.

By contributing to Europe’s capacity in weather and climate modelling and combining AI methods with climate science expertise across an international consortium, EarthGenerator aims to accelerate research in climate change science and deliver societally relevant solutions to global challenges.

The EarthGenerator project is funded by the European Union under Horizon Europe (Grant Agreement No. 101294938 – European Health and Digital Executive Agency, HADEA).

More information about the project click here