发表机构
Università degli Studi di Padova; INAF - Osservatorio di Astrofisica e Scienza dello spazio di Bologna; NASA Goddard Space Flight Center; Space Sciences Laboratory, UC Berkeley（帕多瓦大学; 意大利国家天体物理研究所博洛尼亚天文与空间科学观测站; 美国国家航空航天局戈达德太空飞行中心; 加州大学伯克利分校空间科学实验室）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本研究测试图神经网络（GraphSAGE和Interaction Networks）在AMEGO-X伽马射线望远镜中重建双光子产生事件，旨在提高低能段（<100 MeV）的角分辨率和探测效率，结果显示其性能优于标准方法。
AI 中文摘要
在下一代用于伽马射线空间观测的跟踪探测器的开发中，将先进的深度学习技术集成到事件重建算法中是提高任务性能、减少系统不确定性并最大化科学产出的有前景的方法。在本研究中，我们探讨了图神经网络（GNNs）等深度学习技术在AMEGO-X提议概念跟踪器内识别和重建由双光子产生事件产生的粒子轨迹的应用。目标是提高望远镜的角分辨率和探测效率，特别是在软能量范围（低于100 MeV）内，其中多次库仑散射显著降低轨迹重建质量，从而限制了所谓的MeV间隙中的灵敏度。我们使用基于Geant4的MEGAlib框架和基于Python的专用读出和数据处理处理器来格式化输入数据集，使其尽可能接近真实数据。然后使用模拟数据集训练和评估两种图神经网络架构，即GraphSAGE和Interaction Networks，并比较它们在双光子产生事件重建中的性能，以点扩散函数68%包含半径和有效面积来衡量，也与使用标准重建技术获得的结果进行比较。本研究的结果表明，与标准重建方法相比，基于图神经网络的重建是低于100 MeV双光子产生事件重建的有前景方法，并具有进一步优化的显著潜力。
英文摘要
In the development of next-generation tracking detectors for gamma-ray space observations, the integration of advanced deep learning techniques into event reconstruction algorithms is a promising approach for improving the performance of the mission, reducing systematic uncertainties and maximizing the scientific output. In this study, we investigate the application of deep learning techniques such as Graph Neural Networks (GNNs) for the identification and reconstruction of particle tracks resulting from pair production events within the tracker of the AMEGO-X proposed concept. The goal is improving the angular resolution and the detection efficiency of the telescope, especially in the soft energy range (below 100 MeV) where multiple Coulomb scattering significantly degrades track reconstruction and consequently limits the sensitivity in the so-called MeV gap. We use the Geant4-based MEGAlib framework and a Python-based dedicated read-out and data handler processor to format the input datasets as close as possible to the real data. The simulated datasets are then used to train and evaluate two graph neural network architectures, GraphSAGE and Interaction Networks, and to compare their performance in pair-production event reconstruction in terms of Point Spread Function 68% containment radius and effective area, also with that obtained using standard reconstruction techniques. The results of this study indicate that graph neural network-based reconstruction is a promising approach for pair-production event reconstruction below 100 MeV when compared with standard reconstruction methods, with significant potential for further optimization.
Journal refProc. SPIE 14146, Space Telescopes and Instrumentation 2026: Ultraviolet to Gamma Ray, 141465G (2026)
