Abstract
Electron beam accelerators are essential in many scientific and technologicalfields. Their operation relies heavily on the stability and precision of theelectron beam. Traditional diagnostic techniques encounter difficulties inaddressing the complex and dynamic nature of electron beams. Particularly inthe context of free-electron lasers (FELs), it is fundamentally impossible tomeasure the lasing-on and lasingoff electron power profiles for a singleelectron bunch. This is a crucial hurdle in the exact reconstruction of thephoton pulse profile. To overcome this hurdle, we developed a machine learningmodel that predicts the temporal power profile of the electron bunch in thelasing-off regime using machine parameters that can be obtained when lasing ison. The model was statistically validated and showed superior predictionscompared to the state-of-the-art batch calibrations. The work we present hereis a critical element for a virtual pulse reconstruction diagnostic (VPRD) tooldesigned to reconstruct the power profile of individual photon pulses withoutrequiring repeated measurements in the lasing-off regime. This promises tosignificantly enhance the diagnostic capabilities in FELs at large.