Machine Learning in Gamma Astronomy

TL;DR

The paper surveys deep learning methods applied to Imaging Atmospheric Cherenkov Telescopes (IACTs) data to address gamma-ray event classification and reconstruction of extensive air shower (EAS) parameters amid large charged-particle backgrounds. It covers CNN-based image analyses on irregular hexagonal camera grids, recurrent architectures for time-ordered telescope sequences, self-supervised and anomaly-detection approaches, and generative models for fast Monte Carlo simulation. A key finding is that multi-task CNNs with attention can improve angular and energy reconstruction, achieving angular resolution near $0.3^{\circ}$ at low energies in some setups, and that hex-grid-aware architectures (e.g., IndexedConv) facilitate effective image processing. The review also emphasizes practical software tools such as CTLearn and GammaLearn that enable scalable DL deployment for IACT data, supporting real-time analysis and guiding future observatories with stereoscopic multi-telescope capabilities.

Abstract

The purpose of this paper is to review the most popular deep learning methods used to analyze astroparticle data obtained with Imaging Atmospheric Cherenkov Telescopes and provide references to the original papers.

Figures (5)

• Figure 1: Examples of simulated EAS images in an IACT camera for the TAIGA experiment (on the left: for an EAS initiated by a gamma ray, on the right: for an EAS from a charged particle (proton)).
• Figure 2: The key components of the deep learning based IACT data processing.
• Figure 3: The simplified $\gamma$-PhysNet architecture.
• Figure 4: The general GAN architecture.
• Figure 5: The simplified general architecture of GammaLearn.