Study of residual artificial neural network for particle identification in the CEPC high-granularity calorimeter prototype

Particle Identification (PID) plays a central role in associating the energy depositions in calorimeter cells with the type of primary particle in a particle flow oriented detector system. In this paper, we propose a novel PID method based on the Residual Neural Network (ResNet) architecture which enables the training of very deep networks, bypasses the need to reconstruct feature variables, and ensures the generalization ability among various geometries of detectors, to classify electromagnetic showers and hadronic showers. Using Geant4 simulation samples with energy ranging from 5 GeV to 120 GeV, the efficacy of Residual Connections is validated and the performance of our model is compared with Boosted Decision Trees (BDT) and graph-based approaches, DGCNN, and GravNet. In shower classification, we observe an improvement in background rejection no matter whether data is image-based or graph-based, over a wide range of high signal efficiency (> 95%). These findings highlight the prospects of Artificial Neural Networks with Residual Blocks for imaging detectors in the PID task of particle physics experiments.