Enhancing Photon Identification with Neural Network Methods

TL;DR

This work addresses photon identification in high-luminosity collider environments where electromagnetic showers overlap at fine calorimeter granularity. It compares a BDT baseline on shower-shape variables, a DNN on the same inputs, and a ResNet CNN operating on calorimeter cell energies, using full COCOA-HEP detector simulations. The authors introduce soft scoring and an auxiliary $\\Delta R$ regression head to improve discrimination in hard-background regions, with the ResNet achieving the strongest performance. The results indicate that physics-informed residual CNNs can substantially improve photon identification for HL-LHC-like conditions, offering a robust path toward improved photon reconstruction in high pile-up environments.

Abstract

We investigate photon--pion discrimination in regimes where electromagnetic showers overlap at the scale of calorimeter granularity. Using full detector simulations with fine-grained calorimeter segmentation of approximately $0.025\times0.025$ in $(η,φ)$, we benchmark three approaches: boosted decision trees (BDTs) on shower-shape variables, dense neural networks (DNNs) on the same features, and a ResNet-based convolutional neural network operating directly on calorimeter cell energies. The ResNet significantly outperformed both baseline methods, achieving further gains when augmented with soft scoring and an auxiliary $ΔR$ regression head. Our results demonstrate that residual convolutional architectures, combined with physics-informed loss functions, can substantially improve photon identification in high-luminosity collider environments in which overlapping electromagnetic showers challenge traditional methods.

Figures (11)

• Figure 1: Illustration of $\pi^0$ decay to two photons for well-separated and highly collimated cases.
• Figure 2: The two main channels for direct diphoton production at the LHC.
• Figure 3: $\pi^0$ decay into two photons.
• Figure 4: Distribution of $\Delta R$ between photon pairs from $\pi^0$ decay (blue histogram, left axis) and the corresponding soft-scoring target function (red curve, right axis). Background events ware assigned continuous labels between 0 and 0.4 using a Fermi-Dirac-inspired function, reducing the penalty for near-threshold cases while maintaining hard labels for well-separated pairs.
• Figure 5: ROC curves for BDT, DNN, and ResNet, shown in the high-purity regime ($10^{-4} \leq \epsilon_{\mathrm{bkg}} \leq 10^{-2}$). The ResNet (top curve) demonstrates superior performance than BDT and DNN, achieving higher signal efficiency at any given background efficiency, with the performance gap widening at lower background efficiencies (higher purity).