2025 workshop program
Deep Learning for Compton Image Reconstruction in Gamma-Ray Astrophysics
The work
Abstract
Compton telescopes are essential instruments for observing MeV gamma rays in astrophysics. However, reconstructing source images from Compton scattering events presents a challenging inverse problem, often limited by ambiguities and noise amplification in traditional iterative algorithms like Richardson-Lucy. This work introduces a novel deep-learning framework to enhance Compton image reconstruction. Trained on simulated Compton event data from MEGALib, our method utilizes deep neural networks, including architectures such as Deep Sets designed for permutation-invariant event data, to learn a direct mapping from processed event parameters to high-fidelity gamma-ray source images. The model robustly infers scattering orders by leveraging both spatial and spectral features of the scattering events and suppresses reconstruction artifacts. This leads to improved angular resolution and source localization accuracy compared to conventional techniques. Extensive evaluations on simulated Gamma-Ray Burst (GRB) datasets demonstrate that our approach accurately recovers transient point-like emissions, substantially reduces computational overhead, and offers enhanced robustness against statistical fluctuations. This advancement paves the way for potential real-time data analysis in next-generation gamma-ray missions such as NASA’s COSI. It provides a powerful tool for studying nucleosynthesis, positron annihilation, and other high-energy phenomena, including transient events like GRBs, in our Galaxy.