⚛️ high-energy experiments

Gamma Neutron Radioactive Source Identification in Water Cherenkov Detectors

This paper demonstrates that integrating statistical energy thresholding with an ensemble machine learning model significantly improves gamma-neutron discrimination in Water Cherenkov Detectors, achieving an accuracy of 0.816 and enhancing radiation identification capabilities for nuclear security applications.

Original authors: A. Núñez Selin, C. Sarmiento Cano, H. Asorey, I. Sidelnik

Published 2026-08-10

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: A. Núñez Selin, C. Sarmiento Cano, H. Asorey, I. Sidelnik

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are standing by a giant, crystal-clear swimming pool at night. If a fast-moving particle, like a tiny cosmic bullet, zooms through the water faster than light can travel in that water, it creates a shockwave of blue light, much like a sonic boom but with photons. This is called Cherenkov radiation, and scientists use massive pools of water as detectors to catch these flashes. Usually, these detectors are famous for spotting cosmic rays from space or studying mysterious particles called neutrinos. But recently, scientists have been asking a different question: Can these same water pools help us spot dangerous radioactive materials on Earth? The challenge is that radioactive sources often send out a mix of two things: gamma rays (which are like high-energy X-rays) and neutrons (which are neutral particles that can sneak through lead). To keep people safe, security teams need to tell these two apart instantly. If a detector sees a flash, is it just a harmless gamma ray, or is it a neutron from a hidden nuclear source?

This is exactly the puzzle a team of researchers from Argentina, Colombia, and Spain set out to solve. They wanted to see if a standard water Cherenkov detector could act like a smart security guard, distinguishing between gamma rays and neutrons without needing expensive or scarce equipment. They didn't just guess; they built a two-step "traffic light" system. First, they used old-school math to set a hard energy limit: if the light flash is too weak, it's definitely just gamma rays. If it's strong enough, it might be a neutron. But since high-energy gamma rays can sometimes look like neutrons, they added a second, smarter step. They trained a computer program (a type of artificial intelligence) to look at the tiny "shape" of the light flash itself. Just as a human can tell the difference between a drumbeat and a snare hit by listening to the sound's texture, the computer learned to tell the difference between a neutron signal and a gamma signal by analyzing the pulse's waveform.

The team tested this idea using a real detector filled with pure water, sitting in a lab in Bariloche, Argentina. They used three different radioactive "toys" to train their system: Cobalt-60 and Cesium-137 (which only shoot out gamma rays) and an Americium-Beryllium source (which shoots out both neutrons and gamma rays). To make the test fair, they wrapped the sources in different materials. Sometimes they used lead to block gamma rays, and other times they used special wax and metal sheets to soak up neutrons, letting them isolate exactly what the detector was seeing.

The results were promising. The researchers found that their first step, the "traffic light" math, worked well to filter out low-energy gamma rays. They established a specific energy threshold: if the signal was below a certain point (around 9,000 units of digital charge), they could safely say, "No neutrons here, just gamma rays." However, for the stronger signals where things got tricky, the math alone wasn't enough. That's where the machine learning stepped in. They taught an "ensemble" of three different AI models—a Bagging classifier, a CatBoost model, and a neural network—to work together like a panel of judges. By combining their votes, the system became very good at spotting the difference.

When they tested this combined system on a mixed signal (the Americium-Beryllium source with no shielding), the AI successfully separated the neutron flashes from the gamma flashes. The system achieved an accuracy of about 81.6%, meaning it correctly identified the particle type roughly 8 out of 10 times. Even more impressively, the "Area Under the Curve" (a score that measures how well the system separates the two types) was 0.921, which is a very strong score for this kind of difficult task. The study suggests that by combining simple statistical rules with modern AI, water detectors can become powerful tools for nuclear security. They can quickly screen for dangerous materials, using the water itself to slow down neutrons and the AI to read the story written in the light. While the system works best for higher-energy sources and needs to be recalibrated if the detector changes, the authors show that this approach is a solid, scalable way to make water tanks smarter guardians against hidden radiation.

Technical Summary: Gamma Neutron Radioactive Source Identification in Water Cherenkov Detectors

Problem Statement
Water Cherenkov Detectors (WCDs) are established technologies in astrophysics and high-energy physics but are increasingly sought for nuclear security applications requiring the discrimination of mixed neutron-gamma radiation fields. Conventional neutron detection often relies on $^3$ He proportional counters, which face global supply constraints. While WCDs offer scalability and chemical safety, they lack inherent, robust methods for distinguishing between gamma rays and neutrons in operational security scenarios. The primary challenges addressed in this work are: (1) establishing reliable energy deposition thresholds to perform coarse particle type discrimination, and (2) developing robust pulse-level classification methods to identify mixed radiation fields without relying on external shielding for every measurement.

Methodology
The authors propose a two-stage discrimination framework that integrates statistical spectral analysis with machine learning (ML) pulse classification.

Experimental Setup: Experiments were conducted using a 1.0 m $^3$ cylindrical WCD equipped with a 9" photomultiplier tube (PMT) and a Red Pitaya FPGA-based data acquisition system (125 MS/s sampling). The detector utilized pure water as both the Cherenkov radiator and neutron moderator.
Sources and Shielding: Three radioactive sources were used: $^{60}$ $^{60}$ Co and $^{137}$ $^{137}$ Cs (pure gamma emitters) and $^{241}$ $^{241}$ AmBe (mixed neutron-gamma emitter). To isolate specific radiation components, various shielding configurations were employed:
- Lead (10 cm): Attenuates gamma rays, isolating the neutron-dominated signal (where the detected spectrum is primarily driven by the 2.22 MeV gamma rays from neutron capture in water).
- Borated Paraffin + Cadmium: Moderates and absorbs neutrons, isolating the 4.44 MeV gamma component.
- Lead + Pure Paraffin: A mixed configuration to study combined effects.
Stage 1: Statistical Energy Thresholding:
- The authors defined a "Cutoff Point" (CP) in the charge spectrum (measured in Analog-to-Digital Units, ADU) where the source spectrum becomes statistically indistinguishable from the background.
- This was determined using a $3\sigma$ significance criterion based on Poisson statistics: $|N_{src} - N_{bkg}| < 3\sqrt{N_{src} + N_{bkg}}$ .
- A linear calibration was established between the measured charge (ADU) and deposited energy (MeV) using the known energies of the calibration sources and the Cherenkov threshold for electrons (0.262 MeV).
- This stage functions as a "traffic light" system: spectra with cutoffs below the neutron threshold indicate pure gamma sources; those at or above suggest potential neutron presence.
Stage 2: Machine Learning Pulse Classification:
- For sources exceeding the statistical threshold, a supervised ML approach was applied to the pulse shapes.
- Data Preprocessing: A specific charge range (4500–8000 ADU) was selected to maximize signal-to-noise ratio. Pulses were represented as 32 temporal bins. Random Under Sampling (RUS) was used to balance the training dataset (1:1 neutron-to-gamma ratio).
- Ensemble Model: A soft-voting ensemble classifier was constructed using three base models: a Bagging Classifier, CatBoost, and a Multilayer Perceptron (MLP).
- Optimization: Hyperparameters were tuned via RandomizedSearchCV. The final ensemble weights were derived from the individual F1-scores of the base models. A decision threshold of 0.52 was optimized to balance True Positive and True Negative rates.

Key Contributions

Systematic Calibration: A procedure linking deposited energy to incident radiation characteristics in pure water WCDs, utilizing a statistically rigorous $3\sigma$ cutoff definition.
Two-Stage Framework: A novel workflow combining physical/statistical energy thresholds for initial screening with ML-based pulse analysis for detailed classification.
Ensemble Assessment: A quantitative evaluation of ensemble model diversity, demonstrating that combining models with low inter-correlation (Bagging, CatBoost, MLP) enhances robustness in radiation signal processing.

Results

Calibration: The study established a linear relationship between charge and energy: $E (MeV) = (4.00 \pm 0.19) \times 10^{-4} E_{qC} (ADU) + 0.26$ . The cutoff points for various configurations were quantified (e.g., 9000 $\pm$ 1300 ADU for neutron-dominated signals).
Classification Performance: The ensemble classifier achieved an accuracy of 0.816 and an Area Under the Receiver Operating Characteristic (ROC) curve (AUC) of 0.921 (95% CI: 0.919–0.922).
Mixed Field Decomposition: The trained model was successfully applied to an independent dataset (unshielded $^{241}$ AmBe) not seen during training. It effectively decomposed the mixed radiation field into predicted neutron and gamma components, validating the model's ability to separate signals in real-world scenarios.
Operational Limits: The ML classifier demonstrated high efficacy in the mid-energy region (1.5–2.5 MeV) where background noise is low. The authors note that performance degrades at lower energies due to higher background contamination, a limitation acknowledged for future work.

Significance and Claims
The paper claims that this integrated strategy significantly improves radiation identification capabilities in WCDs, making them viable for nuclear security applications. The significance lies in the hybrid architecture:

Physical Interpretability: The statistical thresholding provides a physically grounded baseline that allows for rapid screening of low-energy gamma emitters without heavy computational load.
Enhanced Discrimination: Machine learning leverages pulse-level information to resolve ambiguities in high-energy mixed fields where simple energy thresholds are insufficient.
Scalability: The approach preserves the intrinsic advantages of water-based detectors (safety, scalability) while mitigating their traditional limitations in low-energy signal discrimination.

The authors emphasize that the method is specific to the detector configuration and requires recalibration if hardware parameters (e.g., PMT gain) change. They conclude that this framework positions WCDs as versatile instruments for both fundamental research and applied radiation detection, offering a robust alternative to supply-constrained technologies.

More like this