AI-based particle track identification in scintillating fibres read out with imaging sensors

TL;DR

The paper tackles rapid identification of particle tracks in scintillating-fibre detectors read by SPAD imaging, where data volumes are dominated by noise. It introduces a variational autoencoder (VAE) trained only on background frames, using loss terms $L_{BCE}$ and $L_{KLD}$ so that $L = L_{BCE} + \beta L_{KLD}$ yields an effective anomaly score; anomalies are flagged when BCE or KLD exceed BG-derived thresholds (e.g., the 98th percentile). Validation on experimental data shows the VAE can distinguish signal frames from background, with substantially faster processing than traditional track-selection methods, especially at low photon counts. The work demonstrates the viability of combining advanced sensor technology with ML-based anomaly detection for real-time particle tracking and event selection, highlighting a path toward hardware-embedded fast inference.

Abstract

This paper presents the development and application of an AI-based method for particle track identification using scintillating fibres read out with imaging sensors. We propose a variational autoencoder (VAE) to efficiently filter and identify frames containing signal from the substantial data generated by SPAD array sensors. Our VAE model, trained on purely background frames, demonstrated a high capability to distinguish frames containing particle tracks from background noise. The performance of the VAE-based anomaly detection was validated with experimental data, demonstrating the method's ability to efficiently identify relevant events with rapid processing time, suggesting a solid prospect for deployment as a fast inference tool on hardware for real-time anomaly detection. This work highlights the potential of combining advanced sensor technology with machine learning techniques to enhance particle detection and tracking.

Figures (13)

• Figure 1: Variational Autoencoder simplified structure.
• Figure 2: (a) SciFi setup used for the acquisition of the data samples. Scintillating fibres are independently coupled to SiPM on their left side end, while on the opposite side they are bundled together and coupled to a single SwissSPAD2 sensor. (b) Close-up photograph of the face of $4\times4$ scintillating fibre bundle. (c) Coupling of the scintillating fibre bundle with SwissSPAD2. Black glob top was used to protect the wire bonding around the active area. Taken from franks2023demonstration.
• Figure 3: Distribution of number of counts per 1 $\mu s$ frame corresponding to the fibre bundle region on SwissSPAD2. The hollow red markers show the distribution obtained from data collected without $^{90}$Sr using SwissSPAD2, known as the background ($\mathrm{BG}$) data. The grey histogram is a Poisson distribution with $\lambda$ equal to mean counts of the $\mathrm{BG}$ data. The solid blue dots show the distribution of the number of counts per frame taken with the $^{90}$Sr source positioned above the fibre bundle. Adjusted from franks2023demonstration.
• Figure 4: The standard track selection process, following the approach described in Ref. franks2023demonstration, is illustrated for three example frames. The red lines indicate the fibre contours, while the filled squares highlight fibres with at least one photon count. The left frame is identified as a signal event, as it contains detected photons on at least three vertically down-going fibres. Similarly, the central frame is classified as signal due to the presence of a diagonal down-going track. In contrast, the right frame is categorised as noise, as it lacks a discernible down-going pattern.
• Figure 5: The plot illustrates the training and validation losses of a Variational Autoencoder (VAE) across iterations, with Binary Cross-Entropy (BCE) loss shown in red and Kullback-Leibler Divergence (KLD) loss in green. Solid lines represent validation losses, while dashed lines denote training losses. The learning rate (blue dotted line) and the $\beta$ weight of the KLD term (purple dash-dot line) are also plotted in arbitrary units.