Hybrid Machine-Learning Particle Identification for the ePIC Proximity-Focusing RICH

TL;DR

This work tackles particle identification in the pfRICH proximity-focusing RICH detector for the ePIC experiment at the Electron-Ion Collider by integrating CNN-based pattern recognition with an XGBoost classifier in a Geant4-based simulation framework. The hybrid model processes per-event Cherenkov-hit images alongside beam-kinematic features to deliver high per-class efficiencies (e.g., protons ~98%, kaons ~97%) and separation powers exceeding three-sigma up to ~12 GeV/$c$ for relevant PID pairs. The key contribution is a practical, high-performance PID pipeline that leverages both image-like detector information and traditional kinematics, with strong interpretability evidenced by feature-importance analyses showing dominant physical features (e.g., $\theta_c$) and meaningful CNN contributions. This approach demonstrates the viability of ML-enhanced Cherenkov detectors for next-generation EIC experiments and provides a adaptable framework for broader detector-PID integration within the ePIC software ecosystem.

Abstract

We present a machine-learning-based particle-identification study for the proximity-focusing Ring Imaging Cherenkov (pfRICH) detector of the ePIC experiment at the Electron-Ion Collider. Operating in the backward region ($-3.5 \lesssim η\lesssim -1.5$), the pfRICH is designed to achieve at least $3σ$ separation among pions, kaons, and protons up to $7,\mathrm{GeV}/c$ for Semi-Inclusive Deep Inelastic Scattering measurements. Using a standalone Geant4 simulation of the pfRICH, we develop a hybrid machine-learning approach that combines convolutional neural-network-based feature extraction with gradient-boosted decision-tree classifiers. This method significantly enhances Cherenkov-ring pattern recognition and improves particle-separation performance, demonstrating the effectiveness of hybrid machine-learning techniques for next-generation Cherenkov detectors at the EIC.

Figures (10)

• Figure 1: Schematic view of the proximity-focusing Ring Imaging Cherenkov (pfRICH) detector.
• Figure 2: Left: Reconstructed Cherenkov angle (mrad) as a function of hadron's momentum (GeV/c) for electrons, pions, kaons and protons using Eq. \ref{['eq:cherenkov']}. Right: Overlapping reconstructed rings of pion and kaon particles on the sensor plane.
• Figure 3: Comparison of photon hit patterns on the pfRICH sensor plane before and after preprocessing. Left: raw event, including background contributions. Right: corresponding sanitized event used for machine-learning training. The sanitization procedure removes low-energy hits ($<1$ eV) and applies fiducial cuts to isolate the primary Cherenkov signal.
• Figure 4: Quality-assurance distributions for sanitized photon hits. Top: $x$ and $y$ coordinate distributions showing the segmented photosensor structure. Bottom left: radial distribution with peaks corresponding to Cherenkov rings at various momenta; the red dashed line indicates the $R_{\mathrm{max}} = 650$ mm acceptance cut. Bottom right: two-dimensional hitmap on the pfRICH sensor plane showing the integrated ring pattern.
• Figure 5: Aggregated hit maps on the pfRICH sensor plane for each particle species. Top left: electrons. Top right: pions. Bottom left: kaons. Bottom right: protons. The characteristic Cherenkov ring radius decreases with increasing particle mass at fixed momentum, providing the physical basis for particle identification. The red dashed circle indicates the $R_{\mathrm{max}} = 650$ mm acceptance boundary.