A Comparison of Deep Learning Models for Proton Background Rejection with the AMS Electromagnetic Calorimeter

TL;DR

This work investigates deep learning architectures for proton background rejection in the AMS ECAL by treating ECAL shower deposits as image-like data. Across MC and ISS datasets, Convolutional Vision Transformers (CvT) achieve the strongest proton rejection at fixed electron efficiency, outperforming MLP, CNN, and ResNet variants, with physics-based feature engineering (Phys+CvT) offering gains on limited ISS data. The results show CvT models generalize well from sub-TeV to TeV energies in MC, while ISS data benefit from physics-informed inputs, though larger datasets further unlock CvT potential. Overall, the study demonstrates the viability of CvT-based calorimeter showers for improved positron identification in AMS, with practical implications for high-energy cosmic-ray physics and dark-m matter studies.

Abstract

The Alpha Magnetic Spectrometer (AMS) is a high-precision particle detector onboard the International Space Station containing six different subdetectors. The Transition Radiation Detector and Electromagnetic Calorimeter (ECAL) are used to separate electrons/positrons from the abundant cosmic-ray proton background. The positron flux measured in space by AMS falls with a power law which unexpectedly softens above 25 GeV and then hardens above 280 GeV. Several theoretical models try to explain these phenomena, and a purer measurement of positrons at higher energies is needed to help test them. The currently used methods to reject the proton background at high energies involve extrapolating shower features from the ECAL to use as inputs for boosted decision tree and likelihood classifiers. We present a new approach for particle identification with the AMS ECAL using deep learning (DL). By taking the energy deposition within all the ECAL cells as an input and treating them as pixels in an image-like format, we train an MLP, a CNN, and multiple ResNets and Convolutional vision Transformers (CvTs) as shower classifiers. Proton rejection performance is evaluated using Monte Carlo (MC) events and ISS data separately. For MC, using events with a reconstructed energy between 0.2 - 2 TeV, at 90% electron accuracy, the proton rejection power of our CvT model is more than 5 times that of the other DL models. Similarly, for ISS data with a reconstructed energy between 50 - 70 GeV, the proton rejection power of our CvT model is more than 2.5 times that of the other DL models.

Figures (11)

• Figure 1: The cosmic positron flux measured by AMS Aguilar2021TheYears compared to (a) the theoretical results of the cosmic positron source being cosmic ray collisions Trotta2011ConstraintsAnalysis and (b) a dark matter model Kopp2013ConstraintsResults.
• Figure 2: The process of how misidentified protons are calculated. The dataset is first split into its constituent proton and electron datasets. The electrons are passed through the model being tested and the model's score for each electron event is recorded. Electron efficiency is the percent of electrons that are correctly classified given a specific model score cutoff, $E_{\%}$. Then, the protons are passed through the model and each proton's score is recorded. For a given $E_{\%}$, proton events that score above this value are classified incorrectly as electrons and are considered misidentified protons.
• Figure 3: Reconstructed energy histograms for the train, val, and two test sets, containing MC events with a reconstructed energy between 0.2–2 TeV.
• Figure 4: Steps taken to separate ISS electrons from protons for the reconstructed energy range of 50–70 GeV. Due to the abundance of matter compared to antimatter, positively and negatively charged particles were labeled as ISS protons and electrons, respectively. The charge sign was determined using the Tracker. (a) Histogram of the TRD Likelihood. (b) Histogram of $|\frac{\text{Energy}}{\text{Rigidity}}|$ (c) 2D histogram of TRD Likelihood vs. $\log_{10} |\frac{\text{Energy}}{\text{Rigidity}}|$. (d) Histogram of the TRD Likelihood after learned cuts (filtering) on TRD Likelihood, energy, and rigidity is made.
• Figure 5: Reconstructed energy histograms for the train/val/test sets created from the dataset consisting of ISS data with a reconstructed energy between 50–70 GeV. The class imbalance (percentage of protons) and energy distribution is the similar between the three sets.