Gamma Ray Discrimination Algorithm In LHAASO-KM2A Experiment With Graph Neural Network

The Large High Altitude Air Shower Observatory(LHAASO)is a major scientific research facility for observing cosmic rays.The Square Kilometer Array(KM2A)is the most important part of LHAASO,with the largest and most numerous muon detectors(MD)to date,and numerous electromagnetic particle detectors(ED).KM2 A can accurately identify muons produced by the extensive air shower in the high energy band,and thus identify gamma in cosmic rays.Since gamma photons are direction invariant in a magnetic field,they become the best messenger for exploring the origin of cosmic rays.However,the previous gamma identification relied on traditional physical methods with several variables combined with physical equations for gamma identification,which could not utilize the structural information between particles and the accuracy depended on manual experience.As deep learning develops,more and more traditional neighborhoods are beginning to use deep learning to gain breakthroughs.Since KM2 A particle information has graphical structure characteristics,this thesis constructs the first end-to-end graph convolution network model from node feature extraction,construction of adjacency matrix,and design of graph convolution.The main work of this thesis is as follows.(1)Currently,the physical method is to use the ratio of the number of electrons detected by ED,and the number of muons from MD detectors as the discriminant formula.To extract more features,this thesis visualizes and analyzes KM2 A data,and combines feature maps,Pearson correlation algorithm to calculate the correlation of features,and draws the distribution of gamma rays on the features to extract relevant features.The processing of the input is completed by one-hot coding and maximum-minimum normalization.Based on the nodes this thesis designs the graph neural network,and based on the KM2 A detector features,this thesis designs the physical composition method.In the experimental part,the Q-factor of gamma-ray discrimination is higher than the physical method by 9～31.(2)For graph neural networks,high-energy physics-related papers generally constitute all nodes as a fully connected undirected graph.To construct a better graph structure,this thesis improves the point cloud composition method: Radius,KNN,and proposes the Radius-pos,Radius-x,KNN-pos,KNN-x methods.To solve the Point Net++ node pooling problem,this chapter proposes a U-shaped network Point Net++(U)based on Point Net++ point cloud network,which is mainly divided into: encoder module and decoder module.In the experiment,the Q-factor of Point Net++(U)algorithm is improved by 76.(3)In view of the problem that the usual graph convolutional network is only for a single graph,it is not conducive to the effective use of the features of two different detectors(ED,MD).Based on this,in this thesis,the features are separated to obtain the features of different detectors.The different features are then cross constructed at four different distances r1,r2,r3,and r4.A modified graph convolution operation is used to fuse the different detector features.And a variety of different aggregation algorithms are explored: mean aggregation,maximum aggregation,and summation aggregation to verify the performance of different aggregation methods.It is verified in the experiments that the graph neural network algorithm proposed in this thesis outperforms other algorithms and improves Q-factor by 84 over the physical method.