The Transfer Learning-Based Approach for Electromagnetic Signal Classification Using Simulated HGCAL Data

This paper provides research results in the context of the Compact Muon Solenoid (CMS) detector design upgrade. Namely, a new mechanical structure called the High Granularity Calorimeter (HGCAL) will replace the current electromagnetic (EM) and hadronic endcap parts. The HGCAL will be an imaging calorimeter, offering the possibility to explore the usage of advanced machine learning methods, specifically neural networks (NN), for processing the detector event data. Detectors have already been used as imaging devices in high-energy physics applications, and we revisit this strategy for the future HGCAL. We use sensors measuring the values of the simulated event energy deposits to generate the NN input data, with the particular target of the HGCAL trigger selection and reduction. The detector layers are represented as images, and pixel intensities correspond to the deposited energy pattern of the signal (EM) and background pile-up (PU) data. The baseline study (BL) assumes electron-induced EM shower data to be a candidate for the EM versus PU classification. Since photon-induced showers are also EM signals, we reuse models pre-trained on electrons to classify the images of events containing photon showers. We show that the evaluated transfer learning (TL)-based approach is as successful as the BL, and the NN model is selected to provide the best performance without retraining the network.