Scholarlink

The Cherenkov Telescope Array Observatory (CTAO) will provide incredible opportunities for the future of ground-based very-high-energy gamma-ray astronomy. Its Science Alert Generation (SAG) system, part of the Array Control and Acquisition (ACADA) system, will handle real-time reconstruction, data quality monitoring, and scientific analysis to issue candidate science alerts. Real-time technical and observational variability, along with performance demands, can affect automated pipeline sensitivity. We developed two Convolutional Neural Network (CNN) prototypes to perform analyzes when requirements for standard techniques are not satisfied. The first is an autoencoder that removes background noise from counts maps without needing target positions, background templates, or Instrument Response Functions (IRFs). The second is a 2D regressor that localizes candidate sources in the field of view, also independent of background templates and IRFs. Verification against the gammapy analysis as implemented in ACADA/SAG v1.0.0 shows comparable results, with the added advantage of reduced dependency on prior knowledge. The CNN regressor achieved a 68% containment radius of approximately 0.07 \(^\circ \) . Meanwhile, the autoencoder significantly improved localization accuracy by removing background noise and demonstrated promising potential for data quality applications.

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

Machine Learning Enhancements for Cherenkov Telescope Data Analysis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ambra Di Piano,
  • Nicolò Parmiggiani,
  • Andrea Bulgarelli,
  • Domenico Beneventano,
  • Gabriele Panebianco,
  • Valentina Fioretti,
  • Luca Castaldini,
  • Riccardo Falco,
  • Daniele Gregori,
  • Elisabetta Boella,
  • Mattia Paladino

摘要

The Cherenkov Telescope Array Observatory (CTAO) will provide incredible opportunities for the future of ground-based very-high-energy gamma-ray astronomy. Its Science Alert Generation (SAG) system, part of the Array Control and Acquisition (ACADA) system, will handle real-time reconstruction, data quality monitoring, and scientific analysis to issue candidate science alerts. Real-time technical and observational variability, along with performance demands, can affect automated pipeline sensitivity. We developed two Convolutional Neural Network (CNN) prototypes to perform analyzes when requirements for standard techniques are not satisfied. The first is an autoencoder that removes background noise from counts maps without needing target positions, background templates, or Instrument Response Functions (IRFs). The second is a 2D regressor that localizes candidate sources in the field of view, also independent of background templates and IRFs. Verification against the gammapy analysis as implemented in ACADA/SAG v1.0.0 shows comparable results, with the added advantage of reduced dependency on prior knowledge. The CNN regressor achieved a 68% containment radius of approximately 0.07 \(^\circ \) . Meanwhile, the autoencoder significantly improved localization accuracy by removing background noise and demonstrated promising potential for data quality applications.