Event reconstruction of gamma-rays using neural networks

Abstract

This study aims to develop and investigate artificial neural networks that reconstruct the energy and emission angles of relativistic gamma-ray events in the CALIFA (CALorimeter for In-Flight detection of gamma-rays and high energy charged pArticles) particle detector. In the present, the addback algorithm is used to reconstruct such events, but only to a limited extent because of difficulties arising from Compton scattering and pair production of the gamma-rays. Both fully connected and convolutional neural networks were investigated and evaluated with detector data which was simulated with the software toolkits Geant4, ggland and ROOT. When finally comparing the neural networks with addback they showed potential to have equal or better accuracy than the addback for gamma-ray energies of 3.5 to 10 MeV. However, the addback had a better signal-to-background ratio between 1 and 10 MeV whereas the neural networks also showed significant issues reconstructing lower energies. The best performing networks were fully connected neural networks trained on simulated detector data that took relativistic effects into account with a mean square-based cost function. Furthermore, this study concludes some hyperparameters of the neural networks which are suitable for the event reconstruction as well as suggestions for further investigation.