Scientists are increasingly turning to artificial intelligence to observe our Earth system. This allows them to predict the weather more accurately or warn of natural events such as floods. However, for most AI models, our Earth and the processes taking place on it are too complex – especially when they are changing as rapidly as they are now due to climate change.

Reliability under constantly changing conditions

A research team from Friedrich Schiller University Jena, Jena University Hospital, the Max Planck Institute for Biogeochemistry in Jena, and the Senckenberg Institute for Plant Diversity in Jena therefore wants to develop AI models that function reliably under constantly changing conditions. The Carl Zeiss Foundation has now announced that it will support the new project “AI Generalizability in Non-stationary Environmental Regimes: The Case of Hydro-climatic Extremes (GENAI-X)” with a total of around six million euros over five years as part of its “CZS Breakthroughs” funding program. The project originates from the ELLIS Unit Jena and is therefore part of the Europe-wide AI excellence network “European Laboratory for Learning and Intelligent Systems.”

In the context of hydro-climatic extreme events

In their project, the Jena researchers are combining two strands of work: On the one hand, they want to fundamentally advance AI-based methods so that they can be applied in complex and changing environmental systems. On the other hand, the scientists want to test these new models in practical environmental research – especially in the context of hydro-climatic extreme events such as floods, droughts, or landslides.

Generalizability of AI systems

“Unlike many other areas of application, an AI model used for such applications should not be based exclusively on data from past events, but should also take into account changing conditions and mechanisms of action in space and time,” explains Prof. Dr. Alexander Brenning, who coordinates the project at the University of Jena. “Such generalizability of AI systems helps, for example, to develop models that work independently of specific locations or without time restrictions.” With their help, it should be possible, for example, to predict landslides in a specific area based on similar events in other regions with better data availability. Or it should be able to predict the risk of a landslide in a specific region, not only based on local conditions during past events of this type, but also taking into account, for example, changes in land use or environmental conditions, such as the frequency or intensity of heavy rainfall events.

Combining environmental processes and artificial intelligence

For its work, the team, which consists of both geoscientists and computer scientists, combines various AI methods. “We are pursuing a hybrid modeling approach in which we combine AI with physical, hydrological, or ecological laws, for example,” explains the Jena-based geoinformatics expert. "Current AI can recognize patterns, but it does not understand why something happens and is therefore unreliable when data is scarce or conditions change. Physical models are based on vast amounts of parameters, most of which are not fully known, which limits their predictive power. Combining both approaches provides AI with physical guidelines for its calculations and allows it to fill in the gaps left by missing parameters in the physical models."

Deriving cause-and-effect relationships

Equally promising are causal AI models that not only read statistical correlations from data, but also derive cause-and-effect relationships. To do this, researchers correlate variables – such as how ecosystems change during drought – describe these relationships using mathematical rules, and thus obtain more reliable statements about possible consequences. The “discovery” of mathematical relationships through symbolic regression also follows this approach. It attempts to cast hidden relationships into clear formulas that are easier to interpret and verify.

Identifying key factors

The project also incorporates dimension reduction, in which AI models use complex data sets – such as satellite images, on-site observations of plant growth, and sensor data – to identify the key factors that are decisive for the respective questions.

Better early warning systems

The Jena team is closely networked with partners from science and practice, such as the German Centre for Integrative Biodiversity Research (iDiv), the Centre for Medium-Range Weather Forecasts (ECMWF), and GeoSphere Austria – Federal Institute for Geology, Geophysics, Climatology, and Meteorology. This network helps to test the new models in practical applications and to generate momentum beyond the project itself. In this way, the researchers are helping to develop better early warning systems for specific disasters. In addition, new AI systems can identify sources of danger at an early stage, enabling timely countermeasures to be taken. For example, certain areas could be specifically reforested to prevent the danger of landslides.

Video