15–18 Dec 2026
Duke University
America/New_York timezone
Artificial Intelligence for the Electron Ion Collider (2026)

Mixture-of-Detectors: A Foundation Model for Detector Simulation and Reconstruction at GlueX

Not scheduled
20m
LSRC B101 Love Auditorium (Duke University)

LSRC B101 Love Auditorium

Duke University

Durham, NC 27710
Trends in Data Science Trends in Data Science

Speaker

Cole Granger

Description

Foundation models are emerging as a promising paradigm for scientific computing, offering the potential to replace collections of task-specific models with shared representations that can be adapted across multiple applications. In nuclear and particle physics, this shift is motivating new approaches to detector modeling, ranging from self-supervised representation learning to generative models for fast simulation.

In this talk, we present recent developments in foundation models for the GlueX experiment as part of the Genesis Mission. We first introduce Mixture-of-Detectors (MoD), which adopts a language-inspired representation of detector responses, using a sparse Mixture-of-Experts transformer to learn correlations across heterogeneous detector subsystems conditioned on particle kinematics. Through a unified pre-trained architecture, MoD supports both generative detector simulation and reconstruction, including particle identification. We discuss its application to these tasks, alongside scaling studies exploring model capacity and expert specialization.

Building on the same underlying transformer backbone, we also present a second foundation-model approach that operates directly on continuous detector representations, employing joint-embedding and generative learning objectives. Although the two approaches differ in their data
representations and training strategies, their shared architecture demonstrates how a common modeling framework can accommodate fundamentally different approaches to learning detector
physics.

Beyond model development, we explore how foundation models can be integrated into practical scientific workflows. We present agentic frameworks that connect tool-calling large language models
with detector simulation, reconstruction, and analysis, alongside a lightweight, high-throughput inference backend designed for on-demand detector-response generation. Together, these developments illustrate a pathway toward more flexible, scalable, and AI-assisted scientific computing. We conclude by discussing opportunities for extending these approaches to future experiments, including detectors at the Electron-Ion Collider.

Author

Presentation materials

There are no materials yet.