Scholarlink

Abstract <p>The SPHERE-3 telescope, currently under development, is designed to study cosmic rays in the energy range of 1–1000 PeV. In this work, we propose a method for filtering extensive air shower (EAS) events whose axes lie outside the telescope’s field of view by employing two complementary machine learning approaches. The first approach uses a convolutional neural network (CNN) to regress the distance between the telescope axis and the EAS axis. The second relies on an autoencoder with an embedded spatial transformer network, trained to reconstruct only ‘‘true’’ EAS images and to discriminate events based on reconstruction error. We demonstrate that the combined approach achieves high accuracy: most ‘‘true’’ events are retained while ‘‘false’’ ones are effectively suppressed. Since the method does not depend on the primary particle’s energy, mass, or arrival direction, it provides robust and universal filtering. The method can be integrated into a trigger system or event selection pipeline for subsequent physical analysis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Filtering of False EAS Maxima Using Neural Network Methods in the SPHERE-3 Experiment

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • E. L. Entina,
  • D. A. Podgrudkov,
  • T. A. Kolodkin,
  • E. A. Bonvech,
  • O. V. Cherkesova,
  • D. V. Chernov,
  • V. I. Galkin,
  • V. A. Ivanov,
  • N. O. Ovcharenko,
  • T. M. Roganova,
  • M. D. Ziva

摘要

Abstract

The SPHERE-3 telescope, currently under development, is designed to study cosmic rays in the energy range of 1–1000 PeV. In this work, we propose a method for filtering extensive air shower (EAS) events whose axes lie outside the telescope’s field of view by employing two complementary machine learning approaches. The first approach uses a convolutional neural network (CNN) to regress the distance between the telescope axis and the EAS axis. The second relies on an autoencoder with an embedded spatial transformer network, trained to reconstruct only ‘‘true’’ EAS images and to discriminate events based on reconstruction error. We demonstrate that the combined approach achieves high accuracy: most ‘‘true’’ events are retained while ‘‘false’’ ones are effectively suppressed. Since the method does not depend on the primary particle’s energy, mass, or arrival direction, it provides robust and universal filtering. The method can be integrated into a trigger system or event selection pipeline for subsequent physical analysis.