AI4Science Seminar #1 – Advancing Science with AI
Focus session: AI for Particle Physics
The first edition of the AI4Science Seminar series, hosted by SCAI, brought together researchers and experts at the intersection of artificial intelligence and scientific discovery.
This inaugural session highlighted how AI is not only accelerating research workflows but also reshaping the way scientific problems are approached—by embedding domain knowledge directly into machine learning models.
“Symmetry-aware networks for Particle Physics: How equivariant transformers enable a deeper understanding of particle jets” by Antoine Petitjean (AIPHY – Universität Heidelberg)
Given the scale of data available from experiments, particle physics is particularly well-suited to machine learning applications. Off-the-shelf ML models already outperform classical methods, but incorporating physics knowledge in our algorithms leads to better results. To study jets, sets of particles produced in high-energy collisions, we can leverage their known symmetry group with equivariant transformers for several tasks that parallel standard ML problems, notably classification and inverse problems. We introduce a Lorentz-equivariant architecture and how it can be tuned for those two tasks. For trigger-level classification, real-time jet tagging in the experiments, we propose different ways of improving the efficiency of our networks. To reconstruct particles from measurements, we introduce a sequential approach to solve this inverse problem with generative models.