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

Beamline Model Calibration from Beam Images Using Conditional Diffusion

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

LSRC B101 Love Auditorium

Duke University

Durham, NC 27710
AI/ML for Accelerators AI/ML for Accelerators

Speaker

Pengyu Li (old dominion university)

Description

We address the inverse problem of estimating solenoid field errors and misalignments from downstream viewer images to calibrate a digital twin of the Upgraded Injector Test Facility (UITF) at Jefferson Lab. This study uses the simulation code Cheetah to model a two-solenoid section of the beamline, with the field-strength error and transverse offsets of each solenoid device as unknown parameters. We train a conditional diffusion model on simulated cases that pair viewer images, acquired at several solenoid settings, with the corresponding device-error vectors. Conditioned on measured images, the trained model samples an approximate posterior over the device errors, yielding parameter estimates with quantified uncertainty. Updating the digital twin with the inferred parameters and compensating with solenoid and corrector settings reduces the beam-centroid error at an unseen setting by an order of magnitude. The study will be extended to test with real measurement data from the UITF.

Authors

Pengyu Li (old dominion university) Karen Makino (Old Dominion University) Mr Jacob Bird (old dominion university) Prof. Yaohang Li (old dominion university) Prof. Monika Yadav (old dominion university)

Presentation materials

There are no materials yet.