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

Detector-Driven Streaming Inference for Beam-Steering at Jefferson Lab

Not scheduled
15m
LSRC B101 Love Auditorium (Duke University)

LSRC B101 Love Auditorium

Duke University

Durham, NC 27710
Calibration, Monitoring, and Experimental Control in Streaming Environments Calibration, Monitoring, and Experimental Control in Streaming Environments

Speaker

Cameron Clarke (Thomas Jefferson National Accelerator Facility)

Description

End-station detectors watch the beam continuously and are a natural candidate for beam optics monitoring, but they only partially capture the beam tune with nontrivial underlying correlations. A change in a detector rate, helicity correlate asymmetry, or spatial profile carries information about the upstream beam state, but inferring that state and how to adaptively steer the beam to restore to an ideal state is an ill-posed inverse problem: many upstream states may be consistent with one downstream observation. As streaming-readout detectors deliver their data continuously, including otherwise overlooked beam-state dependent background signals, it becomes practical to ask whether detector streams can serve as a virtual diagnostic and knob for controlling the beam.

A newly funded Jefferson Lab LDRD project applies a multi-modal shared latent space machine learning approach to address this question. We begin by taking the historical accelerator datasets and the integrating mode data from the CREX parity-violation experiment, whose truly triggerless acquisition approximates continuous counting mode data streams, along with its occasional destructive beam modulation-measured detector and beam-monitor responses that are used for calibrating away any beam optics dependent backgrounds from the integrated signals. The learned latent space describes the structure shared across these continuous streams and destructive measurements, without the need for deriving labels. Before deployment, the model will move onto live inference against streaming data through a JLab EPSCI group guided extension of Hydra to label-free models, with DAOS evaluated as a burst buffer for distributed training and inference. The model will be transferred to a Monte Carlo digital twin first and then the live Upgraded Injector Test Facility (UITF), where optics lattice and resulting beam position, angle, spot size, raster, and more can be deliberately driven out of distribution while streaming end-station detectors observe the response. Beyond validating the streaming inference platform, such beam tests will also prove the ML-driven beam-steering recommendation validity and provide valuable real-world training data.

Running such models on live streams sets concrete requirements for streaming ML tools generally: time alignment of fast detector data with slow EPICS channels and trustworthy uncertainty estimates from confidence intervals over learned distributions to inform operator decision making. We will show the CREX training-data assembly, initial model design, and Hydra extension plan, and invite discussion on coupling detector-side state estimates to accelerator optimizers and online models. Beyond the UITF demonstration, the same tools point toward detector-driven monitoring at ePIC, helicity-correlated beam control for future precision polarized-beam measurements, and, further afield, adaptive proton radiotherapy, where daily cone-beam CT, currently only used for patient positioning, is the intended partial downstream observation against which the ensemble of pencil-beam dose delivery treatment plan must be adapted.

Author

Cameron Clarke (Thomas Jefferson National Accelerator Facility)

Presentation materials

There are no materials yet.