Agentic AI for Breakthroughs in High Energy Physics
Overview
Artificial intelligence is poised to transform experimental physics. Recent advances in mathematics show that the reasoning capabilities are already here.
But what would it take to extend these capabilities to research in which AI agents formulate hypotheses, design investigations and confront their predictions with sensed, physical reality? What would play the role of Lean in experimental physics?
This session, organized in conjoction with CERN openlab, will bring together CERN and industry researchers working on frontier AI, agents, reinforcement learning environments and HEP to explore what is needed to make this jump.
CERN offers a unique environment for this challenge, combining high-fidelity simulations of theories, digital twins of experimental systems and vast quantities of collected data. However, for AI agents to conduct long-horizon exploration, they must establish not only what they find, but also the conditions under which that evidence is valid and how it relates to other evidence. How should this scientific context be represented, queried and verified? Finally AI agents must interface with experimental platforms while safeguarding sensitive datasets and intellectual property.
We will focus on two directions:
1. Open-ended scientific exploration
How can AI agents identify promising research directions by combining theoretical predictions, mathematical consistency and observations across experiments? How can they explore large spaces of hypotheses, allocate limited computational resources and pursue discoveries that were not specified in advance?
2. Closed-loop verification
What scientific infrastructure would enable agents to conduct long-horizon investigations? Do we need a machine-readable, structured and verifiable representation of the experiment (its data, calibrations, software, configurations) and underlying physics assumptions?
Guiding question
What would HEP 2.0 look like? What combination of frontier models, scientific representations, verification infrastructure and agent architectures would enable AI systems to make discoveries in experimental physics?
The session will be informal and discussion-based, with minimal slides and a strong focus on whiteboard exchange. Coffee and snacks will be provided before the session.
Registration
Please register so that we can estimate the number of participants and plan logistics accordingly.