Pulse Shape Analysis for Particle Discrimination in Neutron Detection

Introduction

Pulse Shape Analysis (PSA)–based signal discrimination is a powerful technique for enhancing measurement sensitivity and suppressing unwanted background, especially in scintillator and silicon detectors. One prominent challenge in scintillator-based detection is neutron–gamma discrimination for precise cross-section measurements. Traditional PSA approaches often involve comparing the integrated charge over short- and long-gated portions of the detector signal, exploiting the distinct decay times characteristic of different types of radiation.

In addition to neutron–gamma discrimination, different charged particles (e.g., protons and alpha particles) deposit energy differently depending on their atomic number, particularly in scintillator and silicon detectors. In these cases, the crucial information for discrimination often lies in the signal rise time.

We explored an alternative PSA method using Artificial Neural Networks (ANNs) with both supervised and unsupervised learning. Several datasets were used to evaluate the performance of different network architectures and compare them with conventional PSA methods.

Pulse shape discrimination: supervised learning approach

Clyc scintillators

CLYC detectors rely on the ³⁵Cl(n,p) and ³⁵Cl(n,α) reactions to capture neutrons. The resulting light output correlates with the incident neutron energy, enabling neutron spectroscopy without relying on time-of-flight (TOF) measurements. TOF-based methods typically require large detectors placed at significant distances from the source or target, whereas using CLYC detectors to directly exploit these reactions could eliminate the need for such large-scale setups.

A key advantage of these CLYC-based reactions is their positive Q-value, which helps avoid issues with minimum detectable neutron energies often limited by electronic thresholds. However, the main challenge lies in discriminating between the (n, p) and (n,α) channels, as their different Q-values can produce overlapping spectra, complicating neutron energy identification.

A dataset was collected using a ²⁵²Cf source alongside both CLYC and BaF₂ scintillators. One detector was placed approximately 1 m away from the source to facilitate neutron–gamma discrimination via TOF. This provided traces from the CLYC detector with a precise determination of the incident neutron energy. By correlating the TOF-derived energy with the measured light output, it became possible to distinguish between the (n,p) and (n,α) reactions.

To further investigate the discrimination power for charged particles, we performed short- vs. long-gate charge integrations and a Fast Fourier Transform (FFT) analysis. Results indicate that reliable discrimination between protons and alpha particles is feasible only above approximately 3.5-4 keVee.

We next used the TOF information to label training examples for the ANN-based approach. Several network topologies were tested:

Fully Connected Network
Convolutional Neural Network (CNN)
Long Short-Term Memory (LSTM)

The CNN and LSTM architectures showed the highest overall accuracy (about 70% across the entire energy range) on a dataset trained with 12,000 samples. We also tested wavelet transforms (specifically, the Ricker wavelet) for feature extraction but observed similar performance compared to using the raw signal traces as input.

A notable benefit of the ANN-based method is the ability to set a threshold on the classification confidence. By doing so, one can selectively identify the (n,p) reaction with high certainty and thereby suppress background contributions from the (n,α) channel.

Figure 1: Correlation of the light yield with the energy extracted by the neutron time of flight. Events labelled in red represent the (n,p) reaction while events labelled in green represent the (n,α).

*Figure 2: Pulse shape discrimination with the short and long integration gate, correlated with the light yield. Neutrons and gammas are clearly discriminated.*

*Figure 3: Distribution of neural network predictions. Uncertain events are contained in values lower than 1.*

Pulse shape discrimination: unsupervised learning approach

Autoencoder topology

In many scenarios, the labels for different particle types are unavailable or difficult to obtain, making supervised network training infeasible. However, it is possible to exploit the feature-extraction capabilities of autoencoder architectures. In an autoencoder, the encoder portion is given as input the raw signals and progressively compresses their information into a low-dimensional latent layer (in this case, as few as two floating-point values). The decoder then mirrors the encoder by expanding these latent representations back into signal space.

By minimizing the mean squared error between the original and reconstructed signals, the autoencoder learns to retain the most essential features, while discarding uninformative aspects, such as electronic noise. Consequently, different particle types tend to map onto distinct regions of the latent space, enabling unsupervised discrimination even in the absence of labeled data.

This approach was tested on silicon detector signals, where the latent-space position clearly correlates with the rise time, thus demonstrating the feasibility of autoencoder-based particle separation.

Figure 4: Top panels: comparison of real signal with reconstructed signals. Electronic noise is removed by the network. Bottom-left panel: histogram of the latent space that shows different regions of accumulation of features. Bottom right: Generated signals based on different regions of the latent space.

Pulse shape discrimination: a mixed approach

Autoencoder for gamma/neutron discrimination with Liquid Scintillators

In this study, a Gaussian-Mixture Variational Autoencoder (GMVAE) neural network was developed and applied to enhance neutron/γ-ray discrimination in EJ-309 liquid scintillator detectors. This approach significantly improves performance over traditional pulse-shape discrimination (PSD) methods, particularly at low neutron energies—a critical regime for the SHADES project targeting astrophysical cross-section measurements.

The GMVAE architecture combines three core elements:

A Variational Autoencoder (VAE), which compresses input scintillator waveforms into a latent space and attempts reconstruction, enabling unsupervised feature extraction.
A Gaussian Mixture Model (GMM) integrated into the latent layer to allow clustering across multiple distributions, supporting distinct separation between event types.
A classifier, trained in a semi-supervised fashion, to assign tags (neutron or γ-ray) to the events.

Training involved a curated dataset of 140,000 waveforms from beam runs, with 10,000 samples per run. A fraction of events was pre-tagged using traditional PSD methods to assist in supervised learning via a composite loss function that combined reconstruction error, KL divergence, binary cross-entropy, and a triplet loss to enforce intra-class compactness and inter-class separation in the latent space. Optimal training was achieved using 100 epochs, a learning rate of 10⁻⁴, and a batch size of 512, with dropout layers preventing overfitting.

Post-training, the GMVAE model successfully distinguished neutrons from γ-rays down to a quenched energy of 60 keVee (corresponding to neutron energies below 475 keV), well beyond the limits of traditional PSD, which becomes ineffective below ~2200 keV proton beam energy. The model demonstrated a clean separation in the PSD-energy plane, and significantly improved tagging performance for low-energy events.

This neural network implementation confirms the potential of machine learning to enhance particle discrimination in low-background, low-energy environments—providing a critical tool for next-generation s-process experiments in nuclear astrophysics.