Reconstruction of electron radiation spectra and beam parameters from photon spectrometer data in accelerator experiments using machine learning

Recovering key aspects of an incoming photon stream during projected FACET-II experiments, such as their energy distributions and the original electron beam's parameters, remains an unsolved computational problem. This paper utilized data from simulated plasma wakefield acceleration betatron radiation experiments and electron-positron pair production to determine which methods could most reliably reconstruct these key properties. The data from these two cases provided a large range of photon energies to help increase confidence in each of the tested methods. In both cases, we compared the performance of maximum likelihood estimation (MLE), a statistical technique used to determine unknown parameters from the distribution of observed data, neural networks, which detect patterns between datasets through repeated training, and a hybrid approach combining the two. Furthermore, in the electron-positron production case, the paper also compared QR decomposition, a matrix decomposition method. The betatron radiation case demonstrated the effectiveness of a hybrid ML-MLE approach, while the electron-positron pair production case illustrated the effectiveness of the ML model in the face of noise. As such, the ML-MLE hybrid approach proved to be the most generalizable of the methods.