Deep Learning Applications for Background Rejection in Very-High-Energy Gamma-Ray Observations with LHAASO-KM2A

The identification of gamma-rays and suppression of the cosmic ray hadron background are crucial for very-high-energy gamma-ray observations and related scientific research with LHAASO-KM2A. Traditional machine learning methods, such as support vector machines, decision trees and deep neural networks, have demonstrated promising performance in gamma-ray/hadron separation for ground-based experiments including H.E.S.S. and LHAASO-KM2A. Recent advances in deep learning offer new opportunities for improved particle identification. In this work, we present a deep learning approach for gamma-ray/hadron separation in the 10 TeV to 10 PeV energy range of the LHAASO-KM2A experiment. We model simulated air showers as graph structures and perform particle identification using a dynamic graph convolutional neural network (DGCNN) framework. Our results demonstrate improvements in background suppression, particularly at the energies around 100 TeV, outperforming the traditional method employed by KM2A.