Gamma/hadron separation in the TAIGA experiment with neural network methods

TL;DR

This work tackles the challenge of isolating rare very-high-energy gamma rays from a dominant hadronic background in TAIGA-IACT observations of the Crab Nebula. It proposes a convolutional neural network (CNN) framework that processes Imaging Atmospheric Cherenkov Telescope images to produce a gamma-score, leveraging tailored preprocessing, Wobble-mode transformations, and angle-based selection to boost separation. Training relies on a large, carefully constructed dataset of Monte Carlo gamma events and real hadron data, achieving high background suppression and substantial gamma retention, with a validation threshold set to $0.9965$ and an accuracy of $99.93\%$. On real Crab data, the CNN approach delivers a signal significance around $6.5σ$ over $21$ hours (68 gamma events), performing comparably to traditional Hillas-parameter cuts and offering a viable alternative for gamma/hadron separation and future energy-spectrum reconstruction.

Abstract

In this work, the ability of rare VHE gamma ray selection with neural network methods is investigated in the case when cosmic radiation flux strongly prevails (ratio up to {10^4} over the gamma radiation flux from a point source). This ratio is valid for the Crab Nebula in the TeV energy range, since the Crab is a well-studied source for calibration and test of various methods and installations in gamma astronomy. The part of TAIGA experiment which includes three Imaging Atmospheric Cherenkov Telescopes observes this gamma-source too. Cherenkov telescopes obtain images of Extensive Air Showers. Hillas parameters can be used to analyse images in standard processing method, or images can be processed with convolutional neural networks. In this work we would like to describe the main steps and results obtained in the gamma/hadron separation task from the Crab Nebula with neural network methods. The results obtained are compared with standard processing method applied in the TAIGA collaboration and using Hillas parameter cuts. It is demonstrated that a signal was received at the level of higher than 5.5σ in 21 hours of Crab Nebula observations after processing the experimental data with the neural network method.

Figures (5)

• Figure 1: Architecture of convolutional neural network considered in this work.
• Figure 2: Angle $\alpha$ distribution for gamma-events (left) and hadrons (right).
• Figure 3: Distribution of CNN classifier output for validation data, consisting of 9500 MC gamma events and 9600 experimental hadrons.
• Figure 4: Results obtained in the classification task to separate signal from the Crab Nebula using standard and deep learning methods. On the left picture diagram with average hadron background (calculated in 10 OFF points; filled columns) and excess in ON point at each observation night (empty columns) are shown. Right figure demonstrates the calculated signal significance at each night using LiMa calculationbib20.
• Figure 5: Distributions and interdependence of Hillas parameters for events selected at ON point by standard and CNN with strict selection conditions during observation of the Crab Nebula for 21 hours.