Machine learning-based analysis of experimental electron beams and gamma energy distributions

TL;DR

This paper compares multiple data analysis methods, including neural networks, maximum likelihood estimation, and hybrid approaches, to reconstruct electron beam and photon energy distributions from experimental data in high-energy physics.

Contribution

It introduces a hybrid ML-MLE method for more reliable reconstruction of beam and photon properties in high-field experiments.

Findings

Hybrid ML-MLE approach is most effective and generalizable.

Machine learning models perform well even with noisy data.

QR decomposition is useful for high-energy photon analysis.

Abstract

The photon flux resulting from high-energy electron beam interactions with high field systems, such as in the upcoming FACET-II experiments at SLAC National Accelerator Laboratory, may give deep insight into the electron beam's underlying dynamics at the interaction point. Extraction of this information is an intricate process, however. To demonstrate how to approach this challenge with modern methods, this paper utilizes data from simulated plasma wakefield acceleration-derived betatron radiation experiments and high-field laser-electron-based radiation production to determine reliable methods of reconstructing key beam and interaction properties. For these measurements, recovering the emitted 200 keV to 10 GeV photon energy spectra from two advanced spectrometers now being commissioned requires testing multiple methods to finalize a pipeline from their responses to incident electron beam information. In each case, we compare the performance of: neural networks, which detect patterns between data sets through repeated training; maximum likelihood estimation (MLE), a statistical technique used to determine unknown parameters from the distribution of observed data; and a hybrid approach combining the two. Further, in the case of photons with energies above 30 MeV, we also examine the efficacy of QR decomposition, a matrix decomposition method. The betatron radiation and the high-energy photon cases demonstrate the effectiveness of a hybrid ML-MLE approach, while the high-field electrodynamics interaction and the low-energy photon cases showcased the machine learning (ML) model's efficiency in the presence of noise. As such, while there is utility in all the methods, the ML-MLE hybrid approach proves to be the most generalizable.