Gamma Neutron Radioactive Source Identification in Water Cherenkov Detectors

This paper demonstrates the feasibility of gamma-neutron discrimination in Water Cherenkov Detectors (WCDs) using a combined methodology that integrates statistical analysis with machine learning techniques. The experimental setup employs different shielding configurations to isolate gamma and neutron contributions from a 241 AmBe source, while 60 Co and 137 Cs sources are used to establish a signal to energy calibration.

A statistical analysis based on a 3σ significance criterion is used to define energy thresholds, enabling a linear relationship between the measured charge spectrum and the deposited energy. Building on this calibration, pulse shape information is further exploited through machine learning methods to improve event classification. An ensemble model based on a soft-voting strategy combining a Bagging classifier, CatBoost, and a Multilayer Perceptron was trained on detector signals acquired under different shielding conditions, achieving an accuracy of 0.816 and an area under the Receiver Operating Characteristic (ROC) curve of 0.921.

The combined approach demonstrates that statistical thresholding provides a physically grounded discrimination baseline across the full energy range, while machine learning enhances classification performance at higher energies by leveraging pulse level information. This integrated strategy improves radiation identification capabilities in water Cherenkov detectors, with potential applications in nuclear security and radiation detection.

The detector comprises a cylindrical geometry with an active volume of 1.0 m3 (94 cm in diameter × 147 cm in height), using pure water as the active medium, a Photonics XP1802 9” photomultiplier tube (PMT), a Red Pitaya STEMLAB 125-14 FPGA board sampling at 125 MS/s, and Tyvek® lining for light collection. The primary interactions include gamma radiation through Compton effect, photoelectric effect, or pair production; charged particles directly generating Cherenkov photons; and neutron radiation moderated by water and captured by hydrogen nuclei, emitting 2.22 MeV gamma rays.

Three radioactive sources were used: 60 Co (emitting gamma rays at 1.17 MeV and 1.33 MeV), 137 Cs (emitting gamma rays at 0.66 MeV), and 241 AmBe (producing 4.44 MeV gamma rays and a continuous neutron spectrum). Shielding configurations included lead (10 cm) to attenuate gamma rays by 99%, borated paraffin (10 cm) with cadmium sheet (0.1 cm) to suppress more than 90% of neutron flux, and lead (10 cm) with pure paraffin (15 cm) to combine gamma attenuation with neutron moderation.

The cutoff point is defined as the first histogram bin at which the source spectrum becomes statistically indistinguishable from the background, using a 3σ compatibility criterion where σ = √(N src + N bkg). The cutoff values obtained were 1090 ± 80 ADU for 137 Cs, 2130 ± 180 ADU for 60 Co, 9000 ± 1300 ADU for 241 AmBe with Pb shielding, 8700 ± 1200 ADU for 241 AmBe with Pb/P-Paraffin, 14900 ± 3000 ADU for 241 AmBe with Cd/B-Paraffin, and 16500 ± 4500 ADU for unshielded 241 AmBe. The close agreement between Pb and Pb/P-Paraffin configurations supports the hypothesis that neutrons, once thermalized and absorbed in the detector, predominantly produce a secondary 2.22 MeV gamma. The similarity between Cd/B-Paraffin and Unshielded configurations suggests the cutoff is primarily determined by the highest energy event (4.44 MeV gamma emission).

The energy calibration established a linear relationship: E (MeV) = (4.00 ± 0.19) × 10−4 EqC (ADU) + 0.26, valid for this detector and configuration. The Cherenkov threshold for electrons in water (0.262 MeV) was included as an additional calibration point corresponding to 0 ADU. The residuals do not exhibit a clear systematic bias, but uncertainties increase significantly at higher energies due to poorer counting statistics, with the threshold selection procedure contributing approximately 80% of the total uncertainty.

The two stage gamma neutron classification method operates as a Traffic light system: Red (Below Threshold) indicates pure gamma radiation only; Green (At Threshold) confirms neutron presence with neutrons dominating the spectral endpoint; Yellow (Above Threshold) indicates exclusively high energy gamma rays, with neutrons possibly present at lower energies. The threshold is a necessary but insufficient condition for neutron detection, as sources above the threshold may contain neutrons (e.g., Unshielded configuration) or not (e.g., Cd/B-Paraffin configuration).

For pulse level classification, a specific charge range of 4500–8000 ADUs was defined to maximize neutron signal relative to background noise. Each detected pulse comprises 32 temporal bins, with most pulse structure development occurring between bins 4 and 17. The correlation matrix shows the 32 time bins are moderately correlated (ρ < 0.9), indicating sufficient feature independence for multivariate classification.

The data preprocessing pipeline uses stratified sampling to split data into training (80%) and testing (20%) subsets, applies Random Under Sampling to achieve a balanced 1:1 neutron-gamma ratio (82,967 events each), and standardizes features using StandardScaler. Over 25 classifiers were trained, with 12 classifiers achieving accuracy above 0.75. The final ensemble combined BaggingClassifier, CatBoostClassifier, and Multi-Layer Perceptron based on their individual performance metrics and prediction diversity, with low inter-correlation among base classifiers confirming their complementary nature.

The ensemble system combines predictions through optimized soft voting: P ensemble = 0.3318 × P Bagging + 0.3342 × P CatBoost + 0.3340 × P MLP. The optimal decision threshold was determined to be 0.52, ensuring balanced classification of both neutron and gamma signals. The final test accuracy reached 0.816, with ROC-AUC = 0.921 (95% CI: 0.919-0.922) through 1000 bootstrap iterations.

The classifier was successfully applied to an independent dataset of unshielded 241 AmBe pulses, effectively decomposing the mixed radiation field into its constituent neutron and gamma components. The trained classifier is most effective in the mid energy region (1.5–2.5 MeV), where the natural background count rate is low. The chosen charge window (4500–8000 ADU) restricts the method to sources whose signal emerges above background in the high energy, low background region.

The total cutoff uncertainty accounts for statistical fluctuations (via bootstrap resampling with N = 1000 realizations), threshold sensitivity (repeating the procedure using 1σ and 5σ compatibility thresholds), and discretization effects (σ bin = w bin/√12). The threshold sensitivity term typically accounts for 70–80% of the total uncertainty, the statistical component contributes 20%, and the discretization term remains subdominant (≲ 10%).

Improvements to AI Systems:

1. Hybrid Physics-Informed ML Pipeline: Integrate the statistical 3σ thresholding method as a hard constraint or feature into ML classifiers, rather than using raw pulse data alone. This grounds the model in physical reality, reducing false positives at low energies and improving generalization across detector configurations.

2. Energy-Aware Adaptive Classification: Train the classifier to dynamically switch between statistical thresholding (for low-energy events) and ML-based pulse analysis (for high-energy events), mimicking the traffic light system. This improves accuracy across the full energy spectrum, not just in the mid-range where the current model excels.

3. Uncertainty-Quantified Decision Making: Incorporate the uncertainty decomposition (threshold sensitivity 70–80%, statistical 20%, discretization 10%) into the model's confidence scores. The AI can output calibrated confidence intervals, enabling safer deployment in nuclear security where false negatives are costly.

4. Transferable Calibration Module: Use the linear energy calibration (E = 4.00×10−4 EqC + 0.26) as a pretrained embedding layer, allowing the AI to adapt to new WCD geometries with minimal retraining by fine-tuning only the calibration constants.

5. Ensemble Diversity Optimization: Apply the soft-voting weight optimization (0.33/0.33/0.33) and decision threshold tuning (0.52) as a reusable meta-learning framework. The AI can automatically select and weight base classifiers based on inter-correlation metrics, improving robustness in other radiation detection tasks.

6. Shielding-Aware Data Augmentation: Use the shielding configurations (Pb, Cd/B-Paraffin, etc.) as labeled conditions to train a domain-adversarial network, enabling the AI to disentangle neutron vs. gamma signals from mixed radiation fields without needing physical shielding in deployment.

What the Improved AI System Can Do:

• Operate across the full energy range (0.26–4.44 MeV) with high accuracy, not just the 1.5–2.5 MeV sweet spot, by combining physics-based thresholds with ML pulse classification.

• Provide reliable neutron detection with quantified uncertainty, critical for border security or nuclear material monitoring, where a 3σ threshold alone is insufficient.

• Deploy to new detector geometries (different PMTs, volumes, water purity) with minimal recalibration, using the transferable calibration layer.

• Automatically decompose mixed radiation fields (e.g., unshielded 241 AmBe) into neutron and gamma components in real time, without manual shielding experiments.

• Flag anomalous events that fall between the green and yellow zones, triggering human review—reducing false alarms while maintaining sensitivity.

• Scale to large-area detector arrays (e.g., neutrino observatories) by using the ensemble's low inter-correlation to parallelize classification across nodes, with consistent decision thresholds.

Water Cherenkov Detectors (WCDs) are a robust technology widely used in astrophysics, high energy physics, and recently nuclear security applications. They detect high energy interactions through the Cherenkov light emitted by charged particles traveling faster than the speed of light in water. In this work, we demonstrate the feasibility of gamma-neutron discrimination in WCDs using a combined methodology that integrates statistical analysis with machine learning techniques. The experimental setup employs different shielding configurations to isolate gamma and neutron contributions from a AmBe source, while Co and Cs sources are used to establish a signal to energy calibration. A statistical analysis based on a 3 sigma significance criterion is used to define energy thresholds, enabling a linear relationship between the measured charge spectrum and the deposited energy. Building on this calibration, pulse shape information is further exploited through machine learning methods to improve event classification. An ensemble model based on a soft-voting strategy combining a Bagging classifier, CatBoost, and a Multilayer Perceptron was trained on detector signals acquired under different shielding conditions, achieving an accuracy of 0.816 and an area under the Receiver Operating Characteristic (ROC) curve. The combined approach demonstrates that statistical thresholding provides a physically grounded discrimination baseline across the full energy range, while machine learning enhances classification performance at higher energies by leveraging pulse level information. This integrated strategy improves radiation identification capabilities in water Cherenkov detectors, with potential applications in nuclear security and radiation detection.