Advancing High Energy and Heavy Ion Physics with Artificial Intelligence

by Yeonju Go (Brookhaven National Laboratory (US))

Thursday 21 May 2026, 10:00 → 12:00 Europe/Zurich

4/2-037 - TH meeting room (CERN)

Artificial intelligence (AI) is rapidly becoming a powerful tool for high-energy nuclear and particle physics. In this talk, I will present AI approaches that are opening new capabilities for fast simulation, reconstruction, and data analysis in collider experiments.

First, I will discuss a deterministic AI surrogate for the 2+1D and 3+1D space-time evolution of viscous hydrodynamic solutions from IP-Glasma+MUSIC. TC-QGP, a time-conditioned neural operator, provides orders-of-magnitude speedup with percent-level accuracy, enabling large-scale event-by-event inference, multi-observable global analyses, and hard-probe studies in evolving QGP media that were previously computationally prohibitive.

Second, I will introduce an unsupervised AI approach to jet background subtraction in heavy-ion collisions. UVCGAN-S [1] learns directly from unpaired data without truth labels and generalizes robustly across domains, opening new opportunities for precision studies of jet modification in high-background environments.

Finally, I will discuss how foundation models can move beyond task-specific AI. FM4NPP [2], a foundation model for nuclear and particle physics, uses self-supervised pretraining on low-level detector signals and exhibits neural scaling up to 188 million parameters, similar to trends seen in large language models. Together, these results point toward a new AI-native paradigm for physics discovery across experiments.

[1] arXiv:2510.23717 "Robust and Generalizable Background Subtraction on Images of Calorimeter Jets using Unsupervised Generative Learning”
[2] arXiv:2508.14087 "FM4NPP: A Scaling Foundation Model for Nuclear and Particle Physics"