Machine Learning Optimizes BEGe Detector Event Selection, Achieving Efficiency for 10 keV Radiation Detection

The search for physics beyond our current understanding requires increasingly sensitive detectors, and a team led by Simone Manti, Jason Yip, and Massimiliano Bazzi are pushing the boundaries of low-energy detection with a novel approach to data analysis. Researchers at the VIP collaboration operate a Broad Energy Germanium detector, seeking evidence of new phenomena in the realm of quantum mechanics, and this work presents a significant upgrade to its capabilities.

The team developed a machine learning workflow, employing autoencoders and convolutional neural networks, to dramatically improve the detection of low-energy events, extending sensitivity down to 10 keV. This advancement, validated on a large dataset of detector waveforms, not only lowers the minimum detectable energy but also enhances spectral quality, promising a measurable increase in the detector’s ability to identify rare signals and refine fundamental tests of physics.

Discriminating Dark Matter Signals Via Pulse Shape Analysis This research investigates new ways to improve the sensitivity of dark matter searches by better distinguishing between genuine dark matter interactions and background noise. Current experiments struggle with events that mimic dark matter signals, primarily from environmental radioactivity and neutrons. This work develops and validates a method to identify nuclear recoils, expected from weakly interacting massive particles (WIMPs), by analysing the shape of the scintillation light pulses they produce. The approach utilises cryogenic detectors with high energy resolution and timing capabilities to characterise the emitted light in detail, employing organic scintillators coupled with silicon photomultipliers. Detailed simulations, incorporating realistic detector responses and background conditions, demonstrate the potential to improve the sensitivity of WIMP searches by a factor of 10. The method relies on observing that nuclear recoils produce slower decaying scintillation signals compared to electron recoils, allowing for effective separation through pulse shape discrimination. A key achievement is the development of a novel algorithm for pulse shape discrimination based on machine learning, specifically utilising boosted decision trees, which achieves greater than 99% efficiency in identifying nuclear recoils while minimising misidentification of electron recoils. This detector benefits from substantial shielding, creating a low-background environment crucial for detecting rare events.

The team engineered a new event selection strategy focused on improving the detection of low-energy signals down to 10 keV, a critical threshold for observing subtle quantum phenomena. This innovative approach employs a denoising autoencoder, a type of artificial neural network, to suppress both electronic noise and microphonic vibrations, effectively reconstructing the original pulse shapes. Following noise reduction, a convolutional neural network classifies the waveforms, distinguishing between normal single-site events and those exhibiting anomalies. Validated on a dataset of over 20,000 waveforms, this method achieves high accuracy, with a receiver operating characteristic curve area of 0. 99 and 95 percent accuracy in event classification. Applying this procedure lowers the minimum detectable energy to approximately 10 keV, substantially enhancing the detector’s sensitivity, and yields a measurable 14 percent improvement in the signal-to-background ratio, alongside a reduction in energy resolution for characteristic gamma lines. These advancements enhance the detector’s capability to detect rare low-energy signals and establish a scalable framework for future precision tests of quantum foundations. VIP-2 Searches for Pauli Exclusion Violation This research details the ongoing efforts of the VIP experiment, specifically the VIP-2 phase, to search for violations of the Pauli Exclusion Principle. The core goal is to test this fundamental principle, which states that no two identical fermions can occupy the same quantum state simultaneously; a violation would have profound implications for our understanding of quantum mechanics. The detector is designed to identify anomalous events where an electron appears to occupy an already occupied state. The methodology involves advanced signal processing techniques and the application of deep learning, specifically Denoising Autoencoders (DAEs), to improve the signal-to-noise ratio and enhance the identification of potential violation events. These autoencoders clean up the waveforms, making subtle signals more apparent, and pulse shape discrimination techniques, aided by machine learning, distinguish between different types of events and reduce background noise. The experiment continues to set stringent limits on the probability of Pauli Exclusion Principle violation, contributing to the broader field of quantum foundations and the search for new physics beyond the Standard Model. Researchers developed a workflow that utilises a denoising autoencoder to suppress electronic and microphonic noise, effectively reconstructing pulse shapes, and then employs a convolutional neural network to classify waveforms as either standard events or those containing anomalies. These improvements demonstrably lower the minimum detectable energy to approximately 10 keV, and yield a measurable 14 percent improvement in the signal-to-background ratio, alongside a reduction in energy resolution for characteristic gamma lines. This enhanced performance directly benefits searches for rare events, specifically violations of the Pauli Exclusion Principle and signatures of spontaneous wave-function collapse, both of which require extremely sensitive detection of low-energy signals. 👉 More information 🗞 Machine Learning Optimization of BEGe Detector Event Selection in the VIP Experiment 🧠 ArXiv: https://arxiv.org/abs/2512.09777 Tags: