The Helmholtz Foundation Model Initiative
Advancing the capabilities of AI through Foundation Models
What is the Helmholtz Foundation Model Initiative (HFMI)?
The Helmholtz Foundation Model Initiative (HFMI), launched in February 2024 by the Helmholtz Association, aims to harness the transformative power of advanced AI solutions, specifically "foundation models" to address significant societal challenges and push forward AI development both within Helmholtz and beyond, bringing innovation and novel solutions into a variety of scientific fields.
Through this strategic initiative, high-impact pilot projects have been selected for funding, with a total investment of 11 million euros over three years. Additionally, 12 million euros have been allocated to enhance essential infrastructure.
The HFMI has paved the way for 7 pilot projects across various Helmholtz domains and established a Synergy Unit that addresses overarching challenges, facilitates exchange between projects, and drives common initiatives to accelerate the development of the foundation models.
Building on the strong communities established through the Helmholtz Information & Data Science Framework, HFMI connects experts across fields, fostering collaboration and innovation. These communities, such as Helmholtz AI and Helmholtz Imaging, aid the initiative, provide the foundation for advancing AI-driven research and addressing complex societal challenges.
Still not sure what foundation models are?
Foundation models are machine learning models pre-trained on large datasets, allowing them to understand complex patterns and adapt to various tasks with minimal additional data. Unlike traditional AI models, tailored for specific tasks, foundation models have broad applicability - beyond just natural language processing (e.g. ChatGPT) - making them valuable across different research domains.
About the projects:
Computing resources: HAICORE
The HFMI underwent two selection rounds with an international panel, securing a total investment of 16 million euros from central funds and matching contributions from participating centers. Over three years, the selected projects will receive 11 million euros in funding, while an additional 12 million euros will be dedicated to expanding the necessary infrastructure. Given that foundation models require immense computational power, Helmholtz provides access to cutting-edge supercomputing resources, including Europe’s first exascale computer at Forschungszentrum Jülich. Complementing this, the Helmholtz AI computing resources (HAICORE) strategically enhance supercomputing capabilities, ensuring efficient resource allocation and fostering a world-class research environment.
What leading scientists say about the projects
Through advanced cellular modeling, VirtualCell is poised to drive breakthroughs in disease pathology, drug design, and patient stratification, ultimately driving advancements in biotechnology applications and our understanding of complex diseases.
Prof. Fabian Theis on VirtualCell
Director of Helmholtz AI | Director of Helmholtz Munich’s Computational Health Center
...To this end, we very much look forward to collaborating very closely with our fantastic pilot projects and further Helmholtz-based foundation model endeavors in a broad comparative approach, where we derive general principles and best practices for efficient and effective foundation model development, deployment, and adaptation.
Prof. Dr. Dagmar Kainmüller on the Synergy Unit
Spokesperson for Helmholtz Imaging & Head of Research Unit at MDC
Each individual image can contain a wealth of information. This will create a fundamental understanding in times of global change and additionally help to assess improved carbon budgets and evaluate emerging ocean-based technologies for CO₂ storage.
Dr. Klas Ove Möller on AqQua
Biological Carbon Pump - Head of Department at Helmholtz-Zentrum Hereon
The 3D-ABC foundation model development is a great opportunity to compile, process and integrate the multiple carbon measurements that have been done by different HGF centers. These includes long-term ground measurements, those from drones and planes all the way to satellite-based data incl. laser data and radar.
Prof. Dr. Martin Herold on 3D-ABC
Head of Remote Sensing and Geoinformatics at GFZ Helmholtz Centre for Geosciences
By modeling proteins in motion, PROFOUND opens the door to designing dynamic enzymes and programmable molecular machines.
Prof. Stefan Bauer on PROFOUND
Senior PI at Helmholtz AI
Contacts
Eirini Kouskoumvekaki
Manager Data Science & AI - HFMI Synergy Unit, Max Delbrück Center
eirini.kouskoumvekaki@mdc-berlin.de