A Method for Neutron-Gamma Pulse Shape Discrimination of CLYC Detector Based on a Gated Residual-Linear Attention Network

This paper proposes an enhanced recursive gated cyclic residual-sparse linear attention network for CLYC detectors that achieves high-accuracy neutron-gamma pulse shape discrimination (98.7% accuracy, 2.2 quality factor) with robust noise resistance and ultra-low latency (0.05 ms) suitable for real-time embedded deployment.

The Gist

The Big Picture: Sorting the "Noise" from the "Signal"

Imagine you are at a crowded party where two types of people are shouting: Neutrons and Gamma rays. Both are shouting, but they have slightly different voices.

• Neutrons shout with a slow, heavy voice that takes a while to fade out.
• Gamma rays shout with a sharp, quick voice that stops abruptly.

In the real world, there is also background noise (like people coughing or music playing). The goal of this research is to build a "super-listener" that can instantly tell the difference between the Neutron and the Gamma ray, even when the party is very loud and chaotic.

The researchers built a special computer program (a neural network) to do this listening job using a specific type of sensor called a CLYC detector.

The Problem with Old Methods

Before this new method, scientists tried to sort these voices using two main ways:

1. The "Analogue" Way: Like using a simple mechanical ear. It works okay in a quiet room but gets confused easily if there is too much background noise.
2. The "Digital" Way: Like recording the sound and analyzing the frequency. This is very accurate but requires expensive, high-speed equipment (like a camera that takes a billion photos per second) and is slow to process.

Both old methods struggled when the signal was weak or the noise was high.

The New Solution: The "Smart Detective" (RGLR-SLA)

The authors created a new AI model called RGLR-SLA. Think of this model as a super-smart detective who looks at the shape of the shout (the pulse) from three different angles at the same time.

Here is how the detective works, broken down into three tricks:

1. The Three-Lens Camera (Multi-Scale Feature Detection)

Imagine looking at a wave in the ocean.

• Lens 1 (Zoomed In): Looks at the tiny ripples on the very top of the wave (the rising edge).
• Lens 2 (Medium Zoom): Looks at the main body of the wave (the middle part).
• Lens 3 (Wide Angle): Looks at the whole wave from start to finish (the long tail).

Old methods usually only looked through one lens. If the wave was small, the wide lens missed the details. If the wave was huge, the zoomed-in lens got lost. This new detective uses all three lenses at once, ensuring it catches every detail, whether the signal is tiny or huge.

2. The "Local vs. Global" Team (Gated Residual Fusion)

The detective has two assistants:

• Assistant A (Local): Focuses on the tiny, immediate details of the sound wave.
• Assistant B (Global): Remembers the long history of the sound to see the big picture.

Sometimes the room is quiet, and Assistant A is perfect. Sometimes the room is noisy, and Assistant A gets confused, but Assistant B can still hear the pattern. The detective uses a "Gating Mechanism" (like a smart traffic light) to decide how much to listen to Assistant A and how much to listen to Assistant B. If it's noisy, it listens more to the Global assistant. If it's clear, it listens more to the Local assistant. This teamwork makes the system very tough against noise.

3. The "Speed Reader" (Sparse Linear Attention)

Usually, AI models that look at long sequences of data (like a long speech) get slow because they try to compare every single word to every other word. This is like trying to read a book by checking every letter against every other letter in the book—it takes forever.

This new model uses a "Sparse Linear Attention" trick. Instead of reading the whole book, it learns to skip the boring parts and only focus on the most important words. This makes the detective 50 times faster than the standard "slow reader" AI, allowing it to process signals in real-time without needing a supercomputer.

The Results: How Good is the Detective?

The researchers tested this new detective on a dataset of nearly 20,000 pulses (some from neutrons, some from gamma rays). Here is how it performed:

• Accuracy: It got the answer right 98.7% of the time.
• Noise Resistance: Even when they added heavy static noise (simulating a very loud party), the detective still got it right 95.1% of the time. Old methods dropped below 80% accuracy in these conditions.
• Speed: It can process a single signal in 0.05 milliseconds on a standard graphics card. That is fast enough to be used in real-time monitoring systems, like those used for nuclear safety.

The Bottom Line

The paper claims that by combining a "three-lens" view, a smart "local/global" team, and a "speed-reading" attention mechanism, they have built a system that is:

1. More accurate than traditional methods.
2. Much better at ignoring noise.
3. Fast enough to be used in real-world, real-time safety equipment.

They successfully proved this using a specific detector (CLYC) and a custom-built radiation source, showing that this new "AI detective" is ready to help keep nuclear environments safe and monitored efficiently.

Technical Summary

Technical Summary: A Method for Neutron-Gamma Pulse Shape Discrimination of CLYC Detector Based on a Gated Residual-Linear Attention Network

Problem Statement
The discrimination of neutron and gamma ray pulses is a critical technology for nuclear safety monitoring and radiation assessment. While neutrons possess high penetrating power and pose significant hazards, their detection is frequently complicated by the simultaneous presence of gamma rays, which introduce interference into data analysis. Traditional Pulse Shape Discrimination (PSD) methods, including analogue circuit-based approaches and digital time-domain techniques (e.g., charge comparison, rise-time), often suffer from sensitivity to noise and limited applicability in low-activity mixed radiation fields. Conversely, frequency-domain methods offer strong noise immunity but impose stringent hardware requirements, typically necessitating sampling rates exceeding 1 GS/s. Although recent intelligent algorithms based on machine learning and deep learning have shown promise, challenges remain regarding weak noise resistance, limited feature extraction capabilities, and insufficient real-time performance for embedded deployment.

Methodology
To address these limitations, the authors developed an enhanced Recursive Gated Loop Residual-Sparse Linear Attention (RGLR-SLA) network, implemented on a CLYC (Cs2LiYCl6:Ce) detector experimental platform. The methodology involves three primary architectural components:

1. Multi-Scale Block Embedding: The network utilizes parallel convolutional channels to extract features across different temporal scales. Specifically, it captures nanosecond-level rising edge details, microsecond-level decay rates, and 10-microsecond-level overall contour features. This approach overcomes the single-scale limitations of traditional algorithms, allowing the model to adaptively weight features based on pulse energy (e.g., focusing on rising edges for low-energy pulses and decay trends for high-energy pulses).
2. Enhanced RGLR Module: This core module employs a synergistic fusion of local and global features. It utilizes deep separable convolutions to extract local waveform details and bidirectional Gated Recurrent Units (GRUs) to capture long-range temporal dependencies. A learnable gating mechanism adaptively weights and fuses these features, while residual connections mitigate vanishing gradients. This design aims to suppress interference in high-noise environments while maintaining resolution in low-noise conditions.
3. Enhanced Sparse Linear Attention (SLA): To meet real-time processing demands, the network replaces standard self-attention (O(n²) complexity) with a sparse linear attention mechanism. By combining sparsification strategies with linear projection, the computational complexity is reduced to O(n), significantly improving efficiency without sacrificing the ability to model global features.

Experimental Setup and Data
The study utilized an experimental platform comprising an NG-9 D-D neutron generator (producing 2.5 MeV neutrons) and standard 137Cs and 60Co gamma sources. A CLYC detector (25 mm × 25 mm crystal) coupled with a photomultiplier tube (PMT) and high-speed data acquisition system was used to capture pulse signals. A dataset of 19,971 valid samples (8,974 neutrons and 10,977 gammas) was pre-processed through noise suppression, amplitude normalization, and quality verification (ensuring SNR ≥ 30 dB). The data was stratified into training (75%), validation (15%), and test (10%) sets.

Key Results
The proposed RGLR-SLA network demonstrated superior performance across multiple metrics compared to traditional methods (charge comparison, rise-time, frequency gradient) and benchmark deep learning models (CNN+LSTM+Transformer):

• Classification Performance: The model achieved a Figure of Merit (FoM) of 2.2, a classification accuracy of 98.7%, a precision of 96.4%, a recall of 99.4%, and an F1 score of 97.9%.
• Noise Robustness: Under low signal-to-noise ratio conditions (20 dB), the model maintained an accuracy of 95.1%, significantly outperforming traditional methods which fell below 80% under similar conditions.
• Computational Efficiency: The model contains approximately 2.8 million parameters. On an NVIDIA RTX 4070 GPU, the processing time per pulse is 0.05 ms; on an Intel Core i7-12800HX CPU, it is 0.3 ms. This throughput satisfies the requirements for real-time monitoring and embedded deployment.

Significance and Claims
The paper claims that the RGLR-SLA network provides a feasible technical solution for high-precision, robust, and real-time pulse shape discrimination in neutron-gamma mixed radiation fields. The authors assert that their approach successfully balances accuracy and efficiency by:

1. Capturing multi-scale physical characteristics of pulses that traditional single-scale methods miss.
2. Achieving adaptive fusion of local and global features to enhance noise immunity.
3. Reducing computational complexity through sparse linear attention, enabling deployment on resource-constrained hardware without sacrificing performance.

The study concludes that the proposed method significantly outperforms existing algorithms in both discrimination quality and real-time capability, offering a viable path for advanced nuclear safety monitoring systems.