Speaker
Description
Machine-learning surrogate models for accelerators require compact beam representations that preserve physically meaningful structure while remaining useful for predictive and tuning tasks. In this work, multi-projection beam images from simulated LCLS injector data are studied as a representation-learning problem. The goal is to evaluate how ML model architecture choices impact the degree to which beam distributions, represented via multiple 2D projections, can be compressed into a low-dimensional latent-space vector while retaining important structure, such that the projections can be reconstructed from the latent vector with acceptable accuracy. We examine how model choice affects reconstruction quality and latent-space behavior. Results show that the beam projections are highly compressible, with low reconstruction error across the models studied, while architectural differences lead to clear tradeoffs between reconstruction fidelity and latent-space regularity. Initial latent-space analysis and ongoing neighborhood exploration aim to test whether local perturbations in latent space correspond to smooth variations in decoded beam structure. This study helps identify promising representation-learning approaches for later work in accelerator model calibration, adaptation to measured data, and model-assisted beam tuning.
Funding Agency
This work was supported by the U.S. Department of Energy, under DOE Contract No. DE-AC02-76SF00515 and the Office of Science, Office of Basic Energy Sciences
| I have read and accept the Privacy Policy Statement | Yes |
|---|