深度学习在江门中微子实验顶点重建中的应用

Neutrinos are important fundamental particles. In history, three Nobel Prizes were awarded to neutrino physics, the most recent being in 2015. The mass of neutrinos is the only experimental evidence that is found beyond the standard model of particle physics. The further study of neutrinos makes it possible to discover new physics. Jiangmen Underground Neutrino Observatory (JUNO) hosted by China will determine the order of neutrino mass. The research on this topic is of great significance to the human understanding of the basic microstructure of material as well as the origin and evolution of the macro-universe. In addition, three neutrino mixing parameters will be accurately measured, and JUNO will also make an important contribution to the study of Supernova / Atmospheric / Sun / Eearth Neutrino...JUNO will accurately measure the reactor neutrino energy spectrum in order to achieve the scientific goal of determining the mass hierarchy of neutrinos. To achieve this scientific goal, the energy resolution of JUNO is required to reach 3%. Achieving such a high energy resolution requires a very accurate vertex reconstruction algorithm to accurately reconstruct the vertex of events. This project focuses on the vertex reconstruction algorithm applied to JUNO and the corresponding deep learning application, constructs the optical model of liquid scintillator to meet the physical goal of JUNO.

中微子是一种重要的基本粒子，历史上共有三次诺贝尔奖颁给了中微子物理，最近一次是2015年。中微子有质量是超出粒子物理标准模型所发现的有实验证据的唯一现象，对它的深入研究有发现新物理的可能。由我国主持的江门中微子实验，将决定中微子质量顺序。这课题的研究对于人类了解物质微观的基本结构以及宏观宇宙的起源与演化具有重要意义。此外，还将精确测量3个中微子混合参数，在研究超新星/大气/太阳/地球中微子等方面，江门中微子实验也将作出重要贡献。..江门中微子实验将精确测量反应堆中微子能谱，从而实现确定中微子质量顺序的科学目标。为了实现该科学目标，江门中微子实验要求达到3%的能量分辨率。而达到如此高的能量分辨率需要有很精准的顶点重建算法，从而精确重建出事例的顶点。本项目正是着重研究江门中微子实验的顶点重建算法及深度学习在其中的应用，构造出符合江门中微子实验的液闪光学模型，从而满足江门中微子实验的物理目标。

开发出一套适用于江门中微子实验的深度学习顶点重建方法，包括在TensorFlow框架中实现了对数据读取和模型训练的优化、完成卷积神经网络所需的探测器平面投影算法的设计和优化、平面投影VGG-J、ResNet-J卷积神经网络的结构优化和训练超参数调参、研究电子学产生渡越时间涨落与暗噪声对模型的影响程度、完成球面卷积神经网络所需的三维球面投影算法的设计和优化。最终结果显示，使用深度学习顶点重建，可以获得精确度高，系统偏差小，重建速度快等优势，满足实验的要求，证实了深度学习在高能实验重建中应用的可行性，未来可用于其他相关类型的实验。在本项目的支持下，我们共发表了期刊论文4篇，会议论文1篇，已培养毕业博士生3名，硕士生1名。