Gamma-ray track reconstruction using graph neural networks

Since the advent of the new generation of germanium detector arrays for low-energy nuclear physics experiments utilizing gamma-ray tracking, the challenges associated with track-reconstruction methods have been extensively studied. In the present work an approach based on recent developments in machine learning was used to address the problem. Here, a graph neural network was constructed and trained on data simulated in Geant4 in order to attempt track reconstruction of gamma rays below 1 MeV in a spherical shell geometry of pure germanium. Using a thick-shell geometry, and simulated data without energy- and position uncertainties the network achieved a reconstruction rate above 80% for complete tracks, and a combined peak-to-total value of 85% for energy spectra with four discrete peaks. For data with added noise, i.e. finite resolution in interaction-point position and energy, the corresponding peak-to-total ratio dropped to 74%. The track reconstruction was stable across multiplicities 1–10 but showed an increased error frequency in the energy range between 50 keV and 250 keV. To specifically study the complication of gamma tracks lost by out-scattering from the detector volume, a thin-shell (9 cm thickness) geometry was used together with a modified version of the GNN framework. By letting the GNN code identify and discriminate the out-scatter events, an improvement of the P/T value from 66% to 75% was found for the packed, noisy data. For the sake of comparison the new GNN model with existing gamma-ray tracking methods, a separate instance of the network was trained on slightly higher energies (up to 1.5 MeV) and multiplicities (up to 15) to evaluate 1.332 MeV photon cascade data in terms of P/T and photo-peak efficiency. The results for this GNN data set, with P/T values at 85% for single tracks and 74% for multiplicity 15, show clear promise when compared to the existing tracking methods.