Machine learning based longitudinal virtual diagnostics at SwissFEL

Abstract

The bunch length in a linac driven Free Electron Laser (FEL) is a major parameter to be characterized to optimize the final accelerator performance. In linear machines, this observable is typically determined from the beam imaged on a screen located downstream of a Transverse Deflecting Structure (TDS) used to impinge a time dependent kick along the longitudinal coordinate of the beam. This measurement is typically performed during the machine setup and only sporadically to check the beam duration, but it cannot be continuously repeated because it is time consuming and invasive. A non-invasive method to determine the electron bunch length has already been presented in the past. This method is based on the analysis of the synchrotron radiation light spot emitted by the bunch passing through a magnetic chicane, provided that the energy chirp impinged on the bunch by the upstream radio frequency structures is known. In order to overcome a systematic discrepancy affecting the synchrotron radiation monitor based results compared to the absolute TDS based ones, we implemented and optimized a machine learning approach to predict the bunch length downstream of the two SwissFEL compression stages—from about 10 fs up to about 2 ps—as well as the beam longitudinal profile at the first one.