We’re happy to announce that the BMFTR is funding a new Germany-wide effort “SciFM” to develop foundation models in astro- and particle physics. Our group will be part of the Munich hub together with the group of Daniel Grün at the LMU.
SciFM: Big Models for Big Science#
A major unsolved problem in Fundamental Physics is that even the best theories in physics - the Standard Models of Cosmology and Particle Physics - only explain a fraction of the Universe: 95% of the energy in the Universe appears to be either “dark matter” or “dark energy”. What are they exactly? Currently physicists can only speculate what the laws of nature look like that explain these phenomena.
To answer these questions experimentally, researchers in astro- and particle physics face the challenge of analyzing exabytes of data that modern particle colliders, such as the Large Hadron Collider at CERN or telescopes like the Square Kilometer Array, produce - a perfect proving ground for AI.
The SciFM project, funded by the German Federal Ministry of Research, Technology and Space (BMFTR), aims to create a new generation of advanced AI tools, so-called foundation models, to enable physicists to extract the best possible information out of the gigantic scientific datasets available to them. Instead of training small “AI specialists” to solve individual tasks, foundation models are “AI generalists” that are trained on data from a wide range of sources and capable of solving many tasks. Because they can see data from multiple experiments at once, they can outperform what a model trained on data from any single experiment could achieve.
The project brings together a group of leading researchers in AI for fundamental physics across four the country: Munich, Heidelberg, Hamburg and Aachen, and has been funded with over 3 million euros. The project is now hiring postdoctoral researchers and PhD students.