Detector reconstruction of gamma-rays

Abstract

The study of nuclear reactions through measuring emitted gamma-rays becomes convoluted due to complex interactions with the detector crystals, leading to cross-talk between neighbouring elements. To distinguish gamma-rays of higher multiplicity, the detector data needs to be reconstructed. As a continuation on the works of earlier students, this thesis investigates the application of neural networks as a reconstruction method and compares it to the conventional addback algorithm. Three types of neural networks are proposed; one Fully Connected Network, a Convolutional Neural Network (CNN) and a Graph Neural Network (GNN). Each model is optimized in terms of structure and hyperparameters, and trained on simulated data containing isotropic gamma-rays, before finally being evaluated on realistic detector data.

Compared to previous projects, all presented networks showed a more consistent reconstruction across the studied energy range, which is credited to the newly introduced momentum-based loss function. Among the three networks, the fully connected performed the best in terms of smallest average absolute difference between the correct and reconstructed energies, while having the fewest number of trainable parameters. By the same metric, none of the presented networks performed better than addback. They did, however, show a higher signal-to-background ratio in the energy range of 3–6 MeV. Suggestions for further studies are also given.