Fuhao Ji
National Accelerator Laboratory
ABSTRACT
SLAC MeV-UED, part of the LCLS user facility, is a powerful “electron camera” for the study of ultrafast molecular structural dynamics and the coupling of electronic and atomic motions in a variety of material and chemical systems. At MeV-UED, the growing demand of scientific applications calls for highly automated and rapid switch between different beamline configurations for delivering electron beams meeting specific user run requirements, i.e., pulse length, beam energy, beam spot size, q-resolution, pulse charge, energy spread, etc. However, accelerators are complex systems which involve a large number of parameters, non-linear behavior and many interactive systems. Currently, beam optimizations at MeV-UED rely on hand tuning by experienced operators, requiring ~1h for setting up the desired beam. Numerical simulations can be utilized to understand the beamline and provide guidance. However, traditional simulation tools are too slow to be directly used during operation. Here we utilize machine learning based techniques for beam optimization at MeV-UED. In particular, a multi-objective Bayesian optimizer(MOBO) was utilized for fast searching the parameter space and mapping out the Pareto Front which gives the trade-offs between key beam parameters, i.e., spot size vs q-resolution, pulse length vs charge, etc. Algorithm, model deployment and first test results will be presented. The measurement data collected during this beamtime will be beneficial to future upgrades at UED and photoinjector R&D projects at SLAC.
Poster Session Link:
https://gather.town/invite?token=0pEoq7VP
If you have any questions for the presenter, please contact them by either one of the following ways:
Email: fuhaoji@slac.stanford.edu