Speaker
Description
Hybrid detection of Extensive Air Showers (EAS) by using Large High Altitude Air Shower Observatory (LHAASO) and Electron - Neutron Detector Array (ENDA) can provide a full secondary particle measurement of EAS including electrons, muons, atmospheric Cherenkov light and hadrons, exhibiting the unique capacity for separating composition. This study presents a deep learning-based approach for cosmic ray composition separation using experimental data from LHAASO-KM2A and ENDA. Several measured or reconstructed parameters such as direction, core position, electrons, muons, thermal neutrons and their combinations are optimized as input of machine learning algorithm, and composition separation results such as quality factor, purity and efficiency are obtained. We developed a feedforward neural network framework with PyTorch, incorporating two fully connected layers optimized through cross-entropy loss function and AdamW optimizer. During model training, we conducted systematic optimization of hyperparameters (including learning rate, batch size, and network depth) with maximum iterations set at 20,000 to enhance predictive accuracy. The progressive reduction in training loss was observed throughout the optimization process, while ROC curve analysis and comprehensive quality metrics further validated the model's effectiveness. This work not only establishes a novel methodology for cosmic ray composition separation, but also provides foundational insights for subsequent research in similar experimental contexts, demonstrating the application potential of deep learning techniques in particle physics.
