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

Reinforcement Learning for Autonomous Control of a High-Power Proton Accelerator: Successes and Lessons Learned at SNS

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

LSRC B101 Love Auditorium

Duke University

Durham, NC 27710
AI/ML in Production and Distributed ML AI/ML in Production, Distributed ML

Speaker

Armen Kasparian (Jefferson Lab)

Description

Deploying reinforcement learning (RL) for autonomous control on a high power proton accelerators is gated less by the choice of RL algorithm than by a stack of infrastructure and data realities: the fidelity of any available simulator, the effective sample rate of the live machine, the latency imposed by safe machine-interface layers, and the readiness of historical operational data. We report a staged development effort at the Spallation Neutron Source (SNS) and its Beam Test Facility (BTF), in which each stage’s outcome directly determined the design of the next on the way to the collaboration’s actual operational objective: standing, continuous autonomous control of SNS beam loss. The first stage demonstrated pretrain-and-transfer orbit correction in the SNS medium-energy beam transport under a faithful virtual accelerator, establishing that the shared simulator-to-machine pipeline works end to end. Due to fact that the target problem was already known to have no usable simulator, the second stage used the SNS’s own extended outage to develop and validate simulator-free strategies at the BTF, where a large simulator–machine gap produced negative transfer and from-scratch on-machine training and actor plus-critic imitation learning proved viable instead. Those validated strategies were then carried back to the SNS for the effort’s actual objective, beam-loss optimization, conducted without any usable simulator and constrained by non-AI-ready archival data, proxy-imposed latency, and an action-rate/data-rate mismatch. These are the obstacles this paper documents alongside the concrete progress made toward resolving them. Beyond reporting results, this paper is written as a practical entry point for practitioners starting a similar effort: we describe the shared control-and-learning infrastructure built to support all three stages, report both successes and significant null results with equal weight, and distill the concrete benefits realized, the challenges encountered, and the transferable lessons learned into a decision framework mapping simulator fidelity, data readiness, effective refresh rate, and safety constraints to a recommended RL deployment strategy for production accelerators.

Author

Armen Kasparian (Jefferson Lab)

Co-authors

Presentation materials

There are no materials yet.