Multi-facility virtual diagnostic for longitudinal phase space predictions

Abstract

Abstract A thorough understanding of the longitudinal phase space (LPS) of the electron beam is of great advantage to any modern linear accelerator (linac), and of critical importance for operating a free electron laser (FEL). While a transverse deflecting structure (TDS) allows full characterization of a beam’s LPS, measurements with a TDS system are often destructive and operationally complex. We present an application of machine learning in the form of a virtual diagnostic (VD) trained on destructive TDS measurements, which allows for online predictions of the beam’s LPS based on non-destructive measurements. We show the development and testing of such virtual diagnostics for three different accelerators: the MAX IV linac and the FELs FERMI and SwissFEL. We show how a single, general network architecture and training procedure can be used to reach reliable predictions of the LPS for all three facilities, achieving $$R^2$$ scores reaching 90% or higher across all test datasets. Further, we describe how a simplified architecture can be used for predicting key beam parameters of interest extracted from the full LPS, such as bunch length and slice energy chirp. Our results show how a generalizable VD framework can be rapidly deployed across multiple facilities to enable online monitoring of the beam LPS. For future work, we suggest how virtual diagnostics could be further developed to suit the specific needs of operations at each facility.