New Tools for Old Problems: Generative AI for Discovery in Fundamental Physics

Abstract

Generative AI — a class of algorithms that learn complex probability distributions from data — is opening new avenues for discovery across fundamental physics. In this talk, I will highlight several recent applications. At the LHC, we are ushering in a new paradigm of model-agnostic searches for new physics, developing powerful anomaly detection methods using generative AI techniques such as normalizing flows. These same methods, applied to astronomical data, have enabled discovery of stellar streams and ultra-faint dwarf galaxies. Generative methods also accelerate simulation-based inference across many domains, including gravitational wave detection at pulsar timing arrays. When combined with physics-informed neural networks, generative AI enables an entirely new model-free measurement of the dark matter density in the local Galactic neighborhood. Finally, I will discuss how generative AI in the form of agentic LLMs is creating new modes of human-AI collaboration that promise to further accelerate research. Taken together, these advances signal that we have entered a new and exciting era — one where the oldest questions about our universe may very well be answered with the help of the newest AI tools.