Helmholtz-ELLIS Foundation Models in Science Workshop: A Wrap
© Svea Pietschmann, Max Delbrück Center
Artificial intelligence is transforming the way we conduct research, and foundation models are at the heart of this revolution. From language models like GPT to AI-driven scientific discovery, these models have the potential to accelerate breakthroughs across disciplines. Yet, key questions remain: How can they be tailored to scientific challenges, and what does responsible AI integration in research look like?
Artificial intelligence is transforming the way we conduct research, and foundation models are at the heart of this revolution. From language models like GPT to AI-driven scientific discovery, these models have the potential to accelerate breakthroughs across disciplines. Yet, key questions remain: How can they be tailored to scientific challenges, and what does responsible AI integration in research look like?
Last week, the Helmholtz-ELLIS Foundation Models in Science Workshop brought together leading minds to tackle these issues. Hosted in Berlin, the event provided a platform for experts in machine learning, physics, health, and materials science to exchange ideas on how foundation models can push the boundaries of discovery.
With speakers from Meta, Microsoft, ETH Zurich, Simons Foundation, ELLIS, and top research institutions, the workshop fostered discussions on AI’s role beyond traditional benchmarks, the challenges of building domain-specific models, and the importance of interdisciplinary collaboration.
The event was organized by the Helmholtz Foundation Model Initiative – Synergy Unit & the ELLIS - European Laboratory for Learning and Intelligent Systems (ELLIS) and was supported by Helmholtz AI, Helmholtz Imaging, HEIBRiDS - Helmholtz Einstein International Berlin Research School in Data Science & Zuse School ELIZA.
Throughout the two-day event, discussions emphasized that AI is no longer just a tool for scientific research - it is becoming an essential pillar of efficient research as its relevance and impact expand. Talks covered a wide range of topics, from self-supervised learning and scalable protein modeling to the environmental impact of large-scale AI models.
© Svea Pietschmann, Max Delbrück Center
A key takeaway was the importance of transparency, benchmarking, and open science in ensuring that foundation models remain accessible and interpretable for researchers across disciplines. Experts highlighted the need for efficient architectures that balance computational cost with performance and reliability.
Panel discussions brought forth thought-provoking questions on the future of open large language models for scientific applications. The workshop also served as a catalyst for new partnerships, with researchers exchanging ideas on how to tailor foundation models to the specific needs of scientific domains.
© Svea Pietschmann, Max Delbrück Center
As foundation models continue to evolve, the Helmholtz Association remains committed to advancing their use - not just as experimental tools, but as a driving force for scientific discovery. This commitment is reflected in initiatives like the Helmholtz Foundation Model Initiative (HFMI), which has earmarked €28 million over three years for foundation model development, and Helmholtz AI, which provides essential expertise and infrastructure to support AI-driven research.
We would like to extend our special thanks to the organizing team for making this workshop such a success - Stefan Bauer (Helmholtz AI PI, Helmholtz Munich), Dagmar Kainmüller (Helmholtz Imaging, Max Delbrück Center), Stefan Kesselheim (Helmholtz AI Consultancy Head, Forschungszentrum Jülich), Fabian Isensee (DKFZ German Cancer Research Center, Helmholtz Imaging), and Irene Kouskoumvekaki (Max Delbrück Center ).