Helmholtz invests 23 million in research on AI foundation models

Foundation models are AI applications that, built upon a very broad knowledge base, are capable of solving a range of complex problems. They are significantly more powerful and flexible than traditional AI models, making them suitable for scientific applications. Through targeted training with extensive datasets and the use of generative AI, they can understand complex relationships based on learned patterns, generate new connections, and make predictions. This enables, for example, the stronger integration of global climate data or significant improvements in medical diagnoses. “We are convinced that with foundation models, we can push the boundaries of science. Helmholtz not only brings outstanding talents and comprehensive datasets from various research areas but also brings together a unique computer infrastructure,” says Otmar Wiestler, President of the Helmholtz Association.

The goal of the three-year Helmholtz Foundation Model Initiative (HFMI) is to develop fully functional models. Four pilot projects have been selected for this purpose, involving scientists from twelve Helmholtz Centers. Over a period of three years, the projects will receive funding totaling 11 million euros. An additional 12 million euros will be invested in expanding necessary infrastructure. A Synergy Unit will also research interdisciplinary questions, promote knowledge exchange between projects, and undertake overarching activities. The funded projects aim to not only provide clear value for science but also make their final results available to society as open source—this includes the code, training data, and trained models.

HClimRep: Capturing interactions between the atmosphere, ocean, and sea ice in a novel climate model

What if we could make predictions about future climate even more accurately, quickly, and efficiently? Could we better combat the causes of climate change and mitigate its consequences as a result? Could we make the impacts of global warming impressively visible to everyone? The HClimRep project aims to answer exactly these questions. By building one of the first AI foundation models for climate research, which combines data from the atmosphere, ocean, and sea ice, researchers are developing one of the most precise weather and climate models in the world. This deep-learning model, with billions of parameters, will be capable of conducting complex "what-if" experiments and other modeling tasks related to the ocean and atmosphere, thanks to extensive training on Europe's first exascale computer.

Participating Helmholtz Centers: Forschungszentrum Jülich, Alfred Wegener Institute, Helmholtz Center for Polar and Marine Research, Karlsruher Institute for Technology, and Helmholtz-Zentrum Hereon.

“The Human Radiome Project”: 3D radiology data for a fundamentally better understanding of anatomy and pathology

The more accurately tumors can be localized and marked, the more successful radiation therapy tends to be. However, medicine has so far reached its limits in this respect because it has been very laborious to relate results from different imaging techniques to each other and to present them in a three-dimensional format. Precise tumor localization and marking are just one of the many procedures that “The Human Radiome Project” will improve in the field of medical imaging. It brings together the world's most extensive and diverse collection of 3D radiological images, such as MRI and CT scans, in a foundation model. This gives researchers deep insights into human anatomy and pathology as well as an overview of the entire spectrum of radiological information. “The Human Radiome Project” not only improves personalized medicine but also enhances diagnostic efficiency by reducing the need for manual labeling of complex medical images.

Participating Centers: Deutsches Krebsforschungszentrum, German Center for Neurodegenerative Diseases, and Max Delbrück Center.

SOL-AI: Development and optimization of photovoltaic materials

Photovoltaics is a key technology for the energy transition. In order to achieve the necessary increase in global use of low-cost solar power, innovative solar cell concepts must be implemented more quickly. Activities in research and development in this area are rapidly increasing, leading to a wealth of scientific publications. However, the sheer volume of data is creating limitations in implementing the latest findings. SOL-AI aims to create a foundation model that will fundamentally reform materials informatics in this field. It is capable of integrating the diversity of experimental data and results in the research of photovoltaic materials, advancing innovations in various areas: from accelerated component development and optimization to the discovery of new solar materials. SOL-AI is expected to develop solutions that will have practical relevance for both research and industry.

Participating Helmholtz Centers: Forschungszentrum Jülich, Karlsruher Institute for Technology, Helmholtz-Zentrum Berlin für Materialien und Energie, and Helmholtz-Zentrum Hereon. 

3D-ABC: Calculation and visualization of the global carbon budget of vegetation and soils

To mitigate the effects of global climate change, we need comprehensive knowledge about the global carbon budget, which comprises CO2 sources and sinks such as wetlands, forests, or permafrost soils. Until now, researchers have struggled to quantify how changes in land areas, vegetation, or soils affect the carbon cycle due to heterogeneous and scattered data. The foundation model 3D-ABC will target the integration and modeling of data from various sources such as satellites, drones, or local CO2 monitoring stations. This allows key parameters of the global carbon cycle of vegetation and soils to be captured, quantified, and characterized with high spatial resolution.

Participating Helmholtz Centers: Alfred Wegener Institute, Helmholtz Center for Polar and Marine Research, Forschungszentrum Jülich, Helmholtz-Zentrum Dresden-Rossendorf, Helmholtz Center Potsdam – GFZ German Research Center for Geosciences, Helmholtz Center for Environmental Research, and German Aerospace Center.   

Synergy Unit: Developing, deploying, and connecting foundation models

While individual projects focus on their specific issues, a Synergy Unit concentrates on overarching questions relevant to all participating projects. For example, it addresses concerns such as model scalability or training with datasets. However, it's not just about exchanging solutions; it's primarily about advancing research on foundation models across disciplines as rapidly as possible. Thus, the Synergy Unit ensures a long-term impact of the Helmholtz Foundation Model Initiative for the benefit of the general public.

Participating Helmholtz Centers: Deutsches Krebsforschungszentrum, Helmholtz Munich, Forschungszentrum Jülich, and Max Delbrück Center.