Speaker
Description
Manually calibrating the HPSim simulator to the LANSCE accelerator is time-intensive and demands substantial domain expertise. In this work, we investigate the use of machine learning (ML) to automate much of the calibration process and substantially reduce tuning time. Specifically, our focus is the calibration of the front-end of the accelerator, which involves obtaining the amplitudes and phases of the pre-buncher, main buncher and tank 1 of the drift tube linac. To get empirical data of the accelerator, we use current-phase curves obtained from absorber/collectors, both with the pre-buncher on and off. We derived features from the curves, such as standard deviation of each or average distance between them, which are then trained on ML models. By combining classical ML methods—gradient-boosted decision trees and random forests—with a state-of-the-art transformer model, we achieve a significant speed-up of the calibration process, from about a month of human expert labor to 2-3 days of mostly computational processing.
| In which format do you inted to submit your paper? | LaTeX |
|---|