Machine Learning Applications for the NEWS-G Dark Matter Search Experiment

Abstract

The NEWS-G collaboration searches for light dark matter using Spherical Proportional Counter (SPC) detectors in low-background environments. Currently, the collaboration is working to publish results from data collected in 2019 and is commissioning a new experiment at SNOLAB. This thesis outlines an avenue to incorporate machine learning techniques within the NEWS-G analysis process. Two neural network architectures are designed and tested: a deep learning convolutional autoencoder for removing electronic noise from recorded detector signals, and a modified single output version of the former to predict pulse-specific physics-based characteristics. The noise-removing model is found to be statistically more effective than traditional noise-removal methods, and both models provide benefits for energy measurements, primary electron counting, and event identification. Limit calculations were performed for a hypothetical dark matter search experiment with and without each model for event identification, resulting in more restrictive limits when machine learning was incorporated.