Methodological innovations, new applications, and community-building all play a role in developing more powerful AI systems. ELISE’s Earth and climate sciences programme demonstrates how technical progress, interdisciplinary collaborations, and pilot projects are contributing to advancing AI R&D in an area of vital strategic importance.
The consequences of climate change can now be seen across Europe. Recent years have seen extreme weather events, including heat waves, droughts, floods, and wildfires, increase in frequency and intensity, affecting people, infrastructure, the environment, and the economy. As the need for accurate information about how the Earth’s climate is changing, and how these changes will influence local areas, intensifies, there is growing focus on using AI to help tackle climate change.
The Earth’s climate is a complex system, and a core challenge is to model this system in a way that allows researchers to identify changes, understand their causes, and predict their impact. Physics-based approaches construct models from first principles; known laws of physics that describe mechanistic relationships between different parts of the system. Despite much progress in these models, there continue to be uncertainties that limit how they can be used. Data-driven models offer a route to addressing these limitations. Large volumes of observational data available from different sources today – satellite imagery, for example – open the door to machine learning-led approaches to climate modelling. These approaches are the focus of ELISE’s Machine Learning for Earth and Climate Sciences Programme, led by Gustau Camps-Valls (Universitat de València) and Markus Reichstein (MPI for Biogeochemistry).
There are already examples of the power of these data-driven techniques in climate science. For example, data-driven quantification of the global carbon cycle has allowed researchers to track how carbon dioxide is taken up and released by different ecosystems, combining in-situ observations of carbon dioxide levels with satellite remote sensing and weather forecast data, producing estimates of carbon dioxide movement across the globe; effectively visualising how the Earth breathes. These estimates can be used to benchmark climate models, helping improve the climate projections that form the basis of policymaking, such as the IPCC reports.
This work is one of a collection of projects supported by the ELISE Machine Learning for Earth and Climate Sciences Programme, which also includes:
Across these projects, collaborative workshops have been influential in advancing scientific research, developing AI methods, and delivering new research tools. One example of such a collaboration is the EarthNet challenge, which brings together machine learning researchers and climate scientists to improve forecasts of the local impacts of climate change. The ability to forecast how climate change will influence seasonal weather locally is important in enabling climate adaption, allowing policymakers to plan interventions to prevent food insecurity, increase infrastructure resilience, and prepare communities for extreme weather events. The challenge is how to apply AI to improve these forecasts. Recognising that the local effect of extreme weather or seasonal change can be seen in satellite imagery – changes in vegetation cover, for example – the EarthNet project seeks to leverage advances in computer vision to deliver improved seasonal weather forecasts. Providing a rallying point for both machine learning and climate researchers, this initiative has already catalysed new research and applications of AI. A pilot study, for example, has explored how AI can predict changes to vegetation in Africa using satellite data, producing a model that can forecast seasonal changes and the impact of weather anomalies.
These collaborations are also driving advances in AI methods. Causality, for example, has been a long-standing challenge in AI and in climate science. Identifying cause-effect relationships is important for researchers to understand the drivers of change in the climate system and for policymakers to identify appropriate interventions. In complex systems, such as the Earth’s climate, which are characterised by dynamic interactions across spatial and temporal scales and emergent phenomena, it is challenging to discern which relationships are causal and which are co-occurring. Differentiating causal connections from a variety of correlations that might exist within a dataset is also a challenge for AI developers. A variety of approaches to causal AI exist: hybrid modelling combines mechanistic models about how a system works with data-driven models, allowing researchers to connect data-derived insights to known laws or principles; explainable AI methods could also help researchers scrutinise how a system works; and the field of causal discovery from observational data and hypothesis is also providing new methods. One challenge that follows is how to benchmark these new methods, validating them against a causal ground truth. In response, the CauseMe platform offers ground truth benchmark datasets that can be used to compare the performance of different causal discovery methods by providing real-world data sets where the causal structure is known and synthetic datasets mimicking real-world challenges.
Such efforts provide a rallying point for a growing community of researchers and practitioners in AI for Earth and climate, supporting a wider wave of action to deploy AI to tackle critical climate and environmental concerns.