Speaker
Description
Reliable injector operation depends on diagnostics that are both informative and non-destructive, yet tuning of the ALS injector still relies on intercepting fluorescence screens in the linac and the transfer line to booster to assess some of the beam parameters used for tuning. We investigate a neural-network-based virtual diagnostic that reproduces quantities derived from a dispersive screen in the linac-to-booster transfer line using only always-on measurements. The model takes as input a set of non-intercepting signals and is trained to predict secondary parameters extracted from TV images, such as a scalar bunching metric. Paired data are collected during standard injector tuning by interleaving TV insertions with normal operation over a range of RF and optics settings; simulation-derived samples are optionally used to seed coverage of rare operating points. We compare architectures, regularization strategies, and uncertainty estimates, and study robustness to drifts via time-weighted training and periodic re-labeling. Initial offline and limited online tests indicate that the virtual screen can reproduce TV-derived figures of merit within operational tolerances while substantially reducing the required number of destructive screen insertions. We summarize the training/validation workflow, constraints from booster acceptance and injector timing, and outline integration of the virtual TV into future automated tuning frameworks for ALS.
Funding Agency
Funded by the US Department of Energy (BES Accelerator and Detector Research Program), and supported by the US Department of Energy, Director of the Office of Science under Contract DEAC02-05CH11231.
| In which format do you inted to submit your paper? | LaTeX |
|---|