Machine Learning Optimization of BEGe Detector Event Selection in the VIP Experiment

TL;DR

The paper addresses the challenge of detecting rare low-energy signals (down to ~10 keV) with a BEGe detector in the VIP experiment to test collapse models and Pauli exclusion principle violations. It introduces a denoising autoencoder to clean waveforms and reconstruct pulse shapes, followed by a CNN to distinguish normal single-site events from anomalous ones, trained on synthetic pulse classes and validated on 2021 data. Key results show a ROC AUC of 0.99 and 95% accuracy, with the effective energy threshold lowered to ~10 keV and a ~14% improvement in signal-to-background, along with improved spectral resolution at characteristic lines. The work demonstrates a scalable, online-friendly framework for low-background experiments, complemented by open-source tools and data for reproducibility and adaptation to future precision tests in quantum foundations.

Abstract

The VIP collaboration operates a Broad Energy Germanium detector at the Gran Sasso National Laboratory to measure radiation in the few keV to 100 keV range, aiming to search for spontaneous collapse induced radiation and atomic transitions that violate the Pauli Exclusion Principle. Here we present a machine learning based upgrade for the BEGe detector using an event selection strategy aimed at improving the efficiency in detecting low energy events down to 10 keV. The method employs a denoising autoencoder to suppress electronic and microphonic noises and to reconstruct pulse shapes, followed by a convolutional neural network that classifies waveforms as normal single site or events with anomalies. The workflow was validated on a dataset comprising more than 20000 waveforms recorded in 2021. The classifier achieves a receiver operating characteristic curve with an area under the curve of 0.99 and an accuracy of 95 percent. Applying this procedure lowers the minimum detectable energy of the final spectrum to approximately 10 keV. It also yields a measurable enhancement in spectral quality, including an improvement of about 14 percent in the signal to background ratio and a reduction of the energy resolution for the characteristic Pb and Bi gamma lines. These developments enhance the sensitivity of the BEGe detector to rare low energy signals and provide a scalable framework for future precision tests of quantum foundations in low background environments.

Figures (6)

• Figure 1: Examples of the four pulse types in the synthetic dataset used to train the DAE: a normal single-site pulse (green) in panel a, and anomalous pulses (red): multi-site (panel b), slow-rise (panel c), and saturated (panel d).
• Figure 2: The rise time (left) is defined as the interval between 10% and 90% of the waveform amplitude (red). The FWHM time (right) is defined from the derivative of the waveform, with the FWHM interval highlighted in green.
• Figure 3: Rise time (left) and FWHM time (center) distributions for the events before and after applying the DAE, with mean and variance intervals. On the right is the mean squared error (MSE) of the DAE on the real and synthetic waveform data.
• Figure 4: Effect of the DAE on a real waveform (left) and on its derivative (right) for a low-energy pulse which corresponds to an energy of 15 keV.
• Figure 5: Performance of the CNN classifier applied to DAE-processed waveforms. Left: ROC curve with an AUC of 0.99, with the chosen operating threshold marked in red. Right: accuracy and F1-score as a function of the classification threshold, with the selected operating point indicated by the dashed red line.