Department or Program
Physics and Astronomy
Abstract
Detecting neutrinoless double beta decay relies on extracting weak scintillation signals from noisy silicon photomultiplier (SiPM) readouts. In liquid xenon, radioactive decay produces scintillation photons, which result in voltage spikes on SiPM detectors. These voltage spikes are often obscured by stochastic electrical noise, requiring extensive signal processing before physics analysis can proceed. Although generative denoising methods have shown promise in signal processing, they remain largely unexplored for radiation detection. Here, we evaluate denoising diffusion probabilistic models (DDPMs) for SiPM pulse reconstruction as an alternative to classical Butterworth filtering. We train a DDPM on synthetic clean pulses while injecting standardized empirical detector noise directly into the forward diffusion process. This enables the model to learn a noise distribution matched to the physical detector rather than a generic Gaussian distribution. On empirical data from a Hamamatsu SiPM operating at 32 V bias, only 3 V above the critical operating threshold where classical denoising fails to resolve the first photoelectron peak, denoising improves single-photoelectron peak resolution by 138% for the first peak and 87% for the second peak. We further propose a reformulated diffusion schedule that scales only the empirical noise rather than rescaling the underlying signal, more directly matching the physics of the denoising problem, and demonstrate that a model trained under this shallower schedule achieves comparable resolution improvements with five times fewer diffusion timesteps. We also outline a synthetic-injection validation framework for rigorously characterizing model artifacts in regimes where classical denoising cannot serve as a reliable reference. This demonstrates that DDPMs are a viable denoising framework for detectors with well-characterized noise processes.
Level of Access
Restricted: Embargoed [Open Access After Expiration]
First Advisor
Gillis, Wesley
Date of Graduation
5-2026
Degree Name
Bachelor of Science
Recommended Citation
Gilliam, Sean Michael, "Diffusion-Based Signal Reconstruction for Silicon Photomultiplier Detectors" (2026). Honors Theses. 528.
https://scarab.bates.edu/honorstheses/528
Number of Pages
101