Neural Network Methods for Improving Signal Processing in High-Purity Germanium Detectors for Rare Event Searches

This thesis introduces a hybrid convolutional Transformer-autoencoder for self-supervised denoising of high-purity germanium p-type point contact detector signals. Faithful extraction of information from noisy signals is critical to the sensitivity of neutrinoless double-beta decay experiments, and deep learning-based denoising offers a complementary approach to established noise-reduction techniques. The model is trained using the Noise2Noise method, which requires only pairs of independently noisy real data with no need for clean simulated data as targets. Evaluated against the baseline convolutional autoencoder from previous work in our group, it achieves improved denoising performance, signal reconstruction accuracy, and energy resolution, and lowers the effective energy threshold of a detector. Two downstream applications of the trained model are explored. The pretrained encoder is repurposed via transfer learning for drift time estimation, and denoising as a preprocessing step is shown to improve drift time measurements at low energies. The full chain from drift time estimation to charge trapping correction is demonstrated on real detector data. The model's latent representation is also used for unsupervised pulse-shape clustering, where it resolves finer distinctions between waveform populations than clustering on raw data, with potential applications to data cleaning, especially in low-energy analyses. While the methods developed in this work are applied to high-purity germanium p-type point contact detectors, they are broadly applicable to other detector technologies and can be used to improve signal processing in other rare event searches. More generally, the approach is applicable to one-dimensional waveforms in any domain where paired noisy observations are available, with no need for clean ground truth. This work is thus relevant both within and beyond the particle physics community.

Department dc:contributor.department

Physics, Engineering Physics and Astronomy

Year dc:date.issued

2026

Author dc:creator

• Ye, Tianai

Advisor dc:contributor.supervisor

• Martin, Ryan

• Particle Astrophysics
• Neutrino Physics
• Denoising
• Machine Learning
• Deep Learning
• Neural Networks
• Self-Supervised Learning
• Transfer Learning
• Germanium Detectors
• Reconstruction

Statement dc:rights

• Attribution-NonCommercial-NoDerivatives 4.0 International

Licence dc:rights.uri

• http://creativecommons.org/licenses/by-nc-nd/4.0/

Language dc:language.iso

eng

Handle dc:identifier.uri

https://hdl.handle.net/1974/36536

OAI identifier oai:identifier

oai:queensu.scholaris.ca:1974/36536