Helmholtz-ELLIS Workshop: Foundation Models and Generative AI Are Transforming Scientific Research Across Disciplines
Foundation Models in Science: JSC's ESDE Research Group Participates in the Helmholtz-ELLIS Workshop
On March 18-19, the Helmholtz-ELLIS Workshop Foundation Models in Science took place in Berlin, creating an excellent opportunity for members of JSC's Earth System Data Exploration (ESDE) research group and their partners in the Helmholtz Foundation Model Initiative to mingle with leading minds at the intersection of machine learning and scientific discovery across various disciplines. The initiative plays a central role in this emerging field, advancing its potential through seven interdisciplinary pilot projects.
The workshop made it clear: foundation models are no longer a niche topic. From large language models and multimodal systems to domain-specific applications, their influence is rapidly expanding. Day one delved into technical advances, while day two showcased applications in physics, biology, astronomy, and materials science. With contributions from Microsoft Research, the Simons Foundation, ETH Zurich, Meta AI, the Alan Turing Institute, and others, the event offered a broad yet deep look into this fast-evolving field.
Key Takeaways
A few themes particularly stood out to the ESDE team: the urgent need for better evaluation standards, the persistent challenge of bias in scientific datasets, and the question of how far these models can generalise beyond their training. Equally important was the strong emphasis on collaboration between machine learning experts and scientists who bring deep domain knowledge. These are all topics that the ESDE team is also grappling with in their strategy, and the workshop reinforced how central they are to progress in this space.
Scientific Sparks: Protein Folding, Materials Discovery, and Beyond
ESDE team members Ankit Patnala (JSC), Savvas Melidonis (JSC), and Sindhu Vasireddy (JSC) came back especially inspired by the breadth of applications. Talks on protein modelling and drug discovery demonstrated how generative models can simulate complex molecular dynamics, helping to accelerate therapeutic development. In materials science, foundation models are being used to identify new materials with properties like superconductivity or heat resistance – an exciting prospect for sustainable energy and electronics.
One recurring theme was the models’ ability to connect across scales – capturing fine details while keeping the big picture intact. In astrophysics, for example, AI is becoming a universal lens: powerful enough to uncover structure in massive datasets, and flexible enough to make sense of the unknown.
Panel Reflections: Fast, Powerful – But How Do We Measure Progress?
The ESDE's lead Martin Schultz (JSC) moderated a lively panel discussion on the future of foundation models in science. Panelists included Michael Gastegger and Tian Xe (Microsoft Research AI4Science), Anna Scaife (University of Manchester), and Sarath Chandar (MILA – Quebec AI Institute). Together, they explored where these models can take us – and what hurdles still lie ahead.
A clear takeaway: AI offers more than just speed. It opens new doors in fields where traditional simulations struggle – either because the data is overwhelming or the physics is just too complex to model explicitly. But even with these strengths, defining what “good” looks like remains a challenge. Benchmarks from classical machine learning often fall short, especially once models have seen the test data during training.
The conversation also touched on the big drivers of progress: is it data, compute, or conceptual innovation that matters most? While there’s no simple answer, most agreed that access to high-quality, diverse scientific data is likely the key ingredient for the next breakthroughs.
Finally, the panel tackled a provocative question: could one foundation model eventually address all scientific domains? For now, the consensus is cautious. Task-specific models still tend to perform better when data is abundant. But foundation models bring unique strengths in robustness and transferability – especially between related tasks. Whether this will translate into cross-domain breakthroughs remains to be seen.
Shaping the Future Together
As part of the Helmholtz Association, Forschungszentrum Jülich is proud to contribute to this emerging field. Through the HClimRep project of the Helmholtz Foundation Model Initiative (HFMI), the BMBF-funded RAINA project, and our contribution to the EU-funded WeatherGenerator, we are exploring how foundation models can support long-term forecasting and help societies prepare for extreme weather.
The ESDE team left Berlin with fresh momentum and new ideas – and a clear sense of being part of a larger movement to make generative AI a tool for both scientific progress and societal resilience.
Learn more more about the Helmholtz-ELLIS Workshop here.
"Virtually all research fields are exploring large, multi-purpose machine learning models – and the pace is staggering. With this powerful new tool comes a shared challenge. Every field now faces the same core questions:
How does science actually work? If a model predicts well, is the science done? What does it take to move from prediction to understanding? How do we design experiments that deepen understanding or open paths to more advanced techniques – going beyond what the model already tells us?
We need to find out – by building these tools, learning how to evaluate them, and figuring out how to embed them into the way science generates knowledge."
