TL;DR
This paper explores using machine learning (specifically a fully connected neural network and LightGBM) to distinguish between background single-electron events and the rare neutrinoless double-beta decay signal (two electrons) in a multi-electrode high-purity germanium (HPGe) detector. The study demonstrates the feasibility of this approach, with LightGBM offering a significantly faster training time and competitive discrimination performance.
What Problem Does This Paper Address?
The search for neutrinoless double-beta (0νββ) decay is a crucial experiment in particle physics, aiming to determine if neutrinos are their own antiparticles (Majorana particles). Detecting this extremely rare decay requires highly sensitive detectors and sophisticated background suppression techniques. A significant challenge in these experiments is distinguishing the 0νββ signal, which produces two electrons, from background events, particularly single-electron events caused by various sources like radioactive decays within the detector or gamma-ray interactions. Traditional methods struggle with this discrimination due to the similarity in signal characteristics. This paper tackles this challenge by investigating advanced machine learning techniques applied to the detailed pulse shapes recorded by a novel multi-electrode HPGe detector.
Methods
The researchers developed a multi-electrode HPGe detector system and a corresponding waveform simulation platform. The core of their approach lies in applying machine learning to analyze the electrical pulse shapes generated by particle interactions within the detector.
- Detector System: They utilized a multi-electrode HPGe detector with seven segmented readout electrodes. This design allows for better spatial information compared to traditional single-electrode detectors.
- Waveform Simulation: A detailed simulation pipeline was established, involving:
- Calculating the electric field within the germanium crystal.
- Simulating charge carrier drift and generating induced waveforms using the Shockley-Ramo theory.
- Monte Carlo simulations (GEANT4) to determine energy depositions of single and double electrons.
- Digital Pulse Processing (DPP): A system was implemented to clean and process the raw waveforms, including data quality checks, baseline correction, and energy reconstruction using a trapezoidal filter.
- Machine Learning Models:
- Fully Connected (FC) Neural Network: A standard deep learning architecture was employed for binary classification of single- vs. double-electron events.
- Light Gradient Boosting Machine (LightGBM): A gradient boosting decision tree algorithm known for its efficiency and speed was also used for the same classification task.
- Feature Extraction: For both models, the input was derived from the rising edge of the pulse shape across all seven electrodes, specifically a ±25 point window centered around the pulse's 50% amplitude point.
Datasets
The study primarily used simulated data for training and evaluation.
- Training/Validation/Test Sets: A dataset of 180,000 waveform data sets was generated, comprising 90,000 simulated <sup>76</sup>Ge 0νββ signal events and 90,000 simulated 2039 keV single-electron background events. This dataset was split into training (70%), validation (20%), and test (10%) sets.
- Experimental Data: While the primary analysis relied on simulated data, experimental waveforms from a <sup>57</sup>Co source were used to validate the pulse-shape simulation's accuracy by comparing rise time distributions.
Key Insights and Ideas
- Pulse Shape as a Discriminator: The paper strongly emphasizes that the detailed characteristics of the electrical pulse shape, captured by multi-electrode detectors, contain sufficient information to distinguish between single- and double-electron events, even when traditional track-based methods are insufficient.
- LightGBM's Efficiency: The study highlights the remarkable speed advantage of LightGBM, achieving comparable or better results than the FC neural network with a training time reduced from hundreds of minutes to just a few minutes. This is crucial for large-scale experiments where rapid data processing is essential.
- Multi-electrode HPGe Detectors' Potential: The research demonstrates that the position sensitivity of multi-electrode HPGe detectors, even with resolutions in the sub-millimeter range, can be leveraged indirectly through pulse-shape analysis for background discrimination.
- Feasibility of ML for 0νββ: The work provides strong evidence that machine learning, particularly gradient boosting methods, is a viable and powerful tool for enhancing the sensitivity of future <sup>76</sup>Ge 0νββ experiments by improving background rejection.
Key Results
- FC Neural Network:
- Discrimination efficiency (recall): 77.4%
- Precision: 57.7%
- Training time: 430 minutes
- AUC: 0.61
- LightGBM Model:
- Discrimination efficiency (recall): 73.1%
- Precision: 64.0%
- Training time: 1.5 minutes
- AUC: 0.69
- Track Feature Method: Found to be highly dependent on detector position resolution and not directly applicable with the current detector structure for this specific discrimination task.
- Simulation Accuracy: The simulated waveforms closely matched experimental measurements in terms of rise time distribution, validating the simulation platform.
Limitations and Future Work
- Simulation-Based: The primary analysis relies on simulated data. While the simulation was validated against experimental data, real-world noise and detector imperfections might affect performance.
- FC Network Feature Extraction: The authors note that the FC neural network method took a long time and suggest that further improvements in feature extraction for FC networks could be beneficial in future work.
- Track Feature Limitations: The current detector structure's position resolution was insufficient for effective discrimination using track-based methods. Future detector designs or advanced track reconstruction algorithms might overcome this.
- Noise Interference: The reported discrimination efficiencies were achieved in the absence of significant noise interference. Real experiments will need to account for and mitigate noise effects.
- Future Work: The results are presented as a reference for future <sup>76</sup>Ge 0νββ experiments, implying that these methods can be directly applied or further refined for actual experimental setups. The significant speed advantage of LightGBM makes it a particularly promising candidate for real-time data analysis in such experiments.
Summary updated