Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Two-Stage Gamma-Neutron Source Classification in Water Cherenkov Detectors: Energy Threshold Screening and Machine Learning Pulse Analysis

Alejandro Núñez-Selin, Iván Sidelnik, Christian Sarmiento-Cano, Hernán Asorey, Luis A. Núñez

TL;DR

The paper tackles gamma–neutron discrimination in water Cherenkov detectors by introducing a two-stage approach: first, a physics-driven energy threshold screen distinguishes obvious gamma sources from neutron-emitting fields; then, for ambiguous events above threshold, a pulse-level machine learning ensemble refines classifications. A linear ADU–MeV calibration (R^2 = 0.966) establishes a neutron-detection threshold at $2.62 \,\pm\ 0.77$ MeV, derived from a $3\sigma$ significance analysis. The pulse-level stage employs a soft voting ensemble (Bagging, CatBoost, MLP) that achieves an accuracy of $0.816$ and AUC of $0.921$, demonstrating strong discriminative power while preserving interpretability. Overall, the hybrid framework offers a scalable, actionable solution for nuclear security and nonproliferation monitoring using water-based detectors, with potential extensions to deep learning for raw waveform analysis.

Abstract

Water Cherenkov detectors offer a robust and economical solution for real-time radiation monitoring by detecting Cherenkov light from charged particles moving faster than light in water. This work presents a novel two-stage classification framework for gamma-neutron discrimination: an initial physics-based energy threshold filters unambiguous low-energy gamma sources, followed by a machine learning ensemble that resolves ambiguities at higher energies. The detector response was characterized using $^{60}$Co (1.17/1.33~MeV), $^{137}$Cs (0.66~MeV), and a shielded $^{241}$AmBe source, with lead, paraffin, and cadmium shielding employed to isolate neutron and gamma interactions. Energy calibration established a linear ADU to MeV conversion ($R^2 = 0.966$), enabling identification of a neutron detection threshold at $2.62 \pm 0.77$~MeV via a $3σ$ significance analysis. Stage one categorizes sources as pure gamma (below threshold) or neutron-emitting (at threshold). For ambiguous cases above threshold, a machine learning pipeline utilizing pulse shape analysis was developed. A soft voting ensemble (Bagging, CatBoost, and MLP) achieved an accuracy of 0.816 and an AUC of 0.921. This hybrid scheme combines physics-based filtering with ML refinement, offering an interpretable and scalable solution for nuclear security, nonproliferation monitoring, and fundamental radiation research.

Two-Stage Gamma-Neutron Source Classification in Water Cherenkov Detectors: Energy Threshold Screening and Machine Learning Pulse Analysis

TL;DR

The paper tackles gamma–neutron discrimination in water Cherenkov detectors by introducing a two-stage approach: first, a physics-driven energy threshold screen distinguishes obvious gamma sources from neutron-emitting fields; then, for ambiguous events above threshold, a pulse-level machine learning ensemble refines classifications. A linear ADU–MeV calibration (R^2 = 0.966) establishes a neutron-detection threshold at MeV, derived from a significance analysis. The pulse-level stage employs a soft voting ensemble (Bagging, CatBoost, MLP) that achieves an accuracy of and AUC of , demonstrating strong discriminative power while preserving interpretability. Overall, the hybrid framework offers a scalable, actionable solution for nuclear security and nonproliferation monitoring using water-based detectors, with potential extensions to deep learning for raw waveform analysis.

Abstract

Water Cherenkov detectors offer a robust and economical solution for real-time radiation monitoring by detecting Cherenkov light from charged particles moving faster than light in water. This work presents a novel two-stage classification framework for gamma-neutron discrimination: an initial physics-based energy threshold filters unambiguous low-energy gamma sources, followed by a machine learning ensemble that resolves ambiguities at higher energies. The detector response was characterized using Co (1.17/1.33~MeV), Cs (0.66~MeV), and a shielded AmBe source, with lead, paraffin, and cadmium shielding employed to isolate neutron and gamma interactions. Energy calibration established a linear ADU to MeV conversion (), enabling identification of a neutron detection threshold at ~MeV via a significance analysis. Stage one categorizes sources as pure gamma (below threshold) or neutron-emitting (at threshold). For ambiguous cases above threshold, a machine learning pipeline utilizing pulse shape analysis was developed. A soft voting ensemble (Bagging, CatBoost, and MLP) achieved an accuracy of 0.816 and an AUC of 0.921. This hybrid scheme combines physics-based filtering with ML refinement, offering an interpretable and scalable solution for nuclear security, nonproliferation monitoring, and fundamental radiation research.
Paper Structure (12 sections, 4 equations, 2 figures, 1 table)

This paper contains 12 sections, 4 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Experimental Setup
  3. Detector System
  4. Radiation Interactions and Detection Principles
  5. Shielding Configurations
  6. Energy-Level Classification
  7. Cutoff Point Determination
  8. Energy Calibration
  9. Pulse-Level Classification
  10. Data Preprocessing
  11. Machine Learning Pipeline
  12. Conclusions

Figures (2)

  • Figure 1: (a) ADU to Energy conversion. The background-equivalent charge (ADUs) is converted into radiation energy (MeV) using Eq. \ref{['eq:fit']}. (b) The green band represents the neutron detection threshold, marking the transition from red (no neutrons) to yellow (possible neutron presence).
  • Figure 2: Classification workflow: Decision nodes (blue) route sources to pure gamma (red), confirmed neutron (green), or ambiguous cases (yellow). Machine learning refines high-energy ambiguities to final neutron/gamma states.