A Deep Learning Approach to Improving the Arrival-Direction Accuracy of Gamma-Ray Air-Shower Measurements with the Tibet-III Air-Shower Array

Abstract

The Tibet AS$γ$ experiment observes cosmic gamma rays from several teraelectron volts to the petaelectron volt range using an array of ground-based surface air-shower detectors (Tibet-III), which is composed of plastic scintillators. The arrival direction is conventionally reconstructed by fitting the shower front to detector hit-timing data, and the resulting angular resolution is approximately 0.5$^\circ$ to 0.2$^\circ$ in the energy range from 10 to 100 TeV. In this study, to further improve the directional reconstruction accuracy, we developed a new arrival-direction reconstruction method that integrates a convolutional neural network (CNN) with the conventional method. Evaluations using gamma-ray events generated by Monte Carlo simulations show that the angular resolution is improved by approximately 10\,\%--15\,\% compared with that achieved with the conventional method. In addition, this performance improvement shows little dependence on the zenith angle up to 40$^\circ$. The proposed method provides a performance gain equivalent to increasing the Tibet-III array, which consists of approximately 600 detectors, by roughly 150 to 200 detectors.