With an energy range from a few hundred keV up to GeV’s combined Compton-scattering and pair-creation telescopes are a central technology for future space-based gamma-ray telescopes such as the envisioned AMEGO, AMEGO-X, or e-ASTROGAM satellites. These missions can enable unprecedented measurements of extreme particle accelerators (e.g. GRBs, AGN), study the life cycle of matter generated by stars and supernovae, search for dark matter signatures, and more.
The standard analysis toolset for medium-energy gamma-ray astrophysics is MEGAlib. It was originally developed for the MEGA Compton and pair telescope. While the event reconstruction for untracked Compton events has been significantly advanced in the context of the COSI balloon missions, the analysis of Compton events with electron tracks and low-energy pair events has not seen similar advances yet. However, especially the data analysis of low-energy Compton tracks and the transition region between Compton events and pair events is most challenging, due to short and non-straight electron tracks, the need to distinguish turning Compton electron tracks from pair events, Bremsstrahlung hits masking as Compton interactions, energy lost in passive material deteriorating the angular resolution, Compton events escaping from the detector, etc.
The goal of this project is to look at all steps in the event reconstruction pipeline of Compton and pair events within MEGAlib and develop alternate machine-learning-based reconstruction approaches wherever possible.
Here we present the status of applying different deep learning approaches to the identification and tracking of pair and Compton events. The overall event reconstruction problem is split in several smaller tasks, for example, identify the event type, find the particle tracks, etc. The layout of the neural network, for example, for the event identification is a 3-dimensional convolutional neural network modeled after VoxNet; the electron tracking approach is based on a graph neural network (GNN); the Compton sequencing approach uses a shallow, fully connected neural network; the approach to identify incompletely absorbed events is based on a Random Forrest.
Comparing the new machine-learning approaches with the existing approaches shows, for example, that the GNN approach is on par with the existing approaches, the event type identification method significantly outperforms existing approaches with an identification rate of up to 99%, and the neural network Compton sequencing approach is better than all previously existing approaches.