Enhancing Neutrinoless Double-Beta Decay Sensitivity of Liquid-Xenon Time Projection Chamber with Augmented Convolutional Neural Network

Abstract

Dual-phase time projection chamber (TPC) that employs a multi-ton-scale liquid xenon (LXe) target mass is a pioneering detector technology to search for dark matter. Beyond its advantage in dark matter direct detection efforts, the natural xenon target allows it to search for the neutrinoless double-beta decay ($0νββ$) process, which would violate lepton number conservation and indicate that neutrinos are Majorana particles. However, such $0νββ$ searches have been limited by gamma-ray backgrounds originating from the detector materials. In this work, we designed an augmented convolutional neural network (A-CNN) model to extract additional event-topology information from detector data. Using simulation and calibration data from XENONnT, a leading LXe TPC experiment, our model achieved over 60% background rejection while maintaining 90% signal acceptance. This rejection power improves XENONnT's projected sensitivity of the $^{136}$Xe $0νββ$ search by about 40%. The implementation of A-CNN in the data analysis of future liquid xenon observatories, such as XLZD, will further enhance their sensitivities for $0νββ$ with $^{136}$Xe.

Figures (8)

• Figure 1: Left: TPC schematic with illustration of a gamma event. Right: top and bottom PMT hit patterns and waveforms.
• Figure 2: Energy spectrum of all backgrounds relevant to the $0\nu\beta\beta$ search in XENONnT. Dominant contributions around Q$_{\beta\beta}$ arise from material background (solid orange) and, in particular from $^{214}$Bi in the TPC (solid green), $^{214}$Bi in the LXe shell (dash green), and $^{137}$Xe (solid purple). $2\nu\beta\beta$ (solid blue) of $^{136}$Xe and $^{8}$B solar neutrinos (solid pink) are subdominant. The shaded light blue area denotes the 2 $\Delta E$ ROI.
• Figure 3: Training input construction example. First (from top to bottom): the main S2$_{\text{top}}$ and its resampled waveform; Second: the alternative S2$_{\text{top}}$ and its resampled waveform; Third: the aligned resampled main S2$_{\text{top}}$ and alternative S2$_{\text{top}}$ comparison; Fourth: the complete A-CNN input waveform.
• Figure 4: Schematic diagram of the A-CNN model. Left: the feature extractor; Right: the linear classifier.
• Figure 5: The trained A-CNN performance on the test set. Left: A-CNN score distribution on 0-4 MeV events. The inset shows events with energy lower than 300 keV, which the population at A-CNN score around 0.5 mostly comes from low-energy events. Right: the A-CNN ROC curve without traditional cuts (solid blue, 62% background rejection at 90% signal acceptance), the ROC curve with traditional cuts in $Q_{\beta\beta}$ (dashed blue, 50% background rejection at 90% signal acceptance), and the 90% signal acceptance (TPR) line (dashed red). Background rejection = 1-FPR.