Extending the high-energy cosmic ray electron spectrum with deep learning

宇宙射线电子（CRe）光谱是一个独特的工具，可以提高我们对地球附近高能粒子（宇宙射线，CRs）来源的理解。在这个项目中，我们将研究CR电子的局部源的性质，并通过将能谱扩展到20 TeV来限制这些粒子在我们附近的传播性质。CR质子（CRp）探测距离很远，而由于冷却，能量高于1tev的CRe应该来自距离地球约1kpc的源。目前天基和地面仪器对CRe光谱的测量扩展到约5 TeV。由于这些仪器的有效面积小，用天基望远镜测量高于5 TeV的能量是有问题的。因此，测量更大能量的光谱只能用地面望远镜，最突出的是成像大气切伦科夫望远镜（IACTs），如H.E.S.S.和未来的CTA。我们建议使用18年的可用H.E.S.S.数据，目标是将CRe光谱的测量扩展到~ 20 TeV（确切的值取决于对超过5 TeV的光谱形状的假设）。先前发表的（未发表的）利用H.E.S.S.对CRe光谱的测量包括3年（6年）的数据收集。我们还将使用深度学习方法确定未来CTA对CRe测量的敏感性。为了充分发挥IACTs的潜力，我们将使用不同类型的望远镜（H.E.S.S.的中型和大型望远镜或CTA的小型，中型和大型望远镜）-这种混合CRe光谱分析（与使用相同类型望远镜的立体分析相比）迄今尚未进行过。考虑到1 TeV以上强子CRs的通量比CRe通量大1000倍以上，强子CRs需要1000数量级以上的高抑制因子，而大部分CRe应保留。我们建议使用现代深度学习方法（特别是图神经网络）来使用全相机淋浴图像分离强子和轻子淋浴。我们将测试系统不确定性的主要来源，包括：(1)强子相互作用模型（影响基于MC模拟的强子阵雨的抑制性能），(2)18年观测期间大气波动和望远镜性能变化，以及(3)重建方法的选择（通过与基于增强决策树和空气阵雨半解析建模的标准背景抑制方法进行比较）。测量的CRe光谱将允许我们推断地球附近的超新星速率，约束CRe加速的效率，对从超新星遗迹中逸出的CRe的最高能量设置下限，并推导出它们传播到地球的性质。

The cosmic ray electron (CRe) spectrum is a unique tool to improve our understanding of the sources of high-energy particles (cosmic rays, CRs) in the vicinity of Earth. In this project, we will study properties of local sources of CR electrons and constrain the properties of the propagation of these particles in our vicinity by extending the energy spectrum to 20 TeV. CR protons (CRp) probe large distances, while, due to cooling, CRe with energies above 1 TeV should come from sources within about 1 kpc from Earth. Current measurements of the CRe spectrum by space-based and ground-based instruments extend to about 5 TeV. Measurements at energies higher than 5 TeV by space-based telescopes are problematic due to small effective area of these instruments. As a result, the measurement of the spectrum to larger energies is only feasible with ground-based telescopes, most prominently with imaging atmospheric Cherenkov telescopes (IACTs), such as H.E.S.S. and the future CTA. We propose to use 18 years of available H.E.S.S. data with the goal of extending the measurement of the CRe spectrum up to ∼ 20 TeV (the exact value depends on assumptions about the shape of the spectrum beyond 5 TeV). The previously published (unpublished) measurements of CRe spectrum with H.E.S.S. include 3 years (6 years) of data collection. We will also determine the sensitivity of the future CTA to measurements of CRe using deep learning methods. In order to realize the full potential of IACTs, we will use telescopes of different types (medium and large telescopes in H.E.S.S. or small, medium and large telescopes in CTA) – such hybrid analysis of the CRe spectrum (compared to stereo analysis where telescopes of the same type are used) have not been performed up to date. Given that the flux of hadronic CRs above 1 TeV is more than a factor of 1000 larger than the CRe flux, a high rejection factor on the order of 1000 or better for hadronic CRs is needed, while the majority of CRe should be kept. We propose to use modern deep learning methods (specifically, graph neural networks) for the separation of hadronic and leptonic showers using full camera shower images. We will test the main sources of systematic uncertainties related to (1) the hadronic interaction models (which impact the performance of the rejection of hadronic showers based on MC simulations), (2) the atmospheric fluctuations and changes in the telescope performance during the 18 years of observations, and (3) the choice of reconstruction method (by comparing with the standard background rejection methods based on boosted decision trees and semi-analytic modelling of air showers). The measured CRe spectrum will allow us to infer the supernova rate in the vicinity of Earth, constrain the efficiency of CRe acceleration, put a lower limit on the highest energies of CRe escaping from supernova remnants and derive properties of their propagation to Earth.