Machine Learning Techniques for Signal Classification in Background-Dominant Particle Detectors
Sophia Farrell, Rice University
Rare-event particle detectors are employed for a host of physics and nuclear security applications. These detectors usually share a similar working principle: an experiment searches for a known or hypothesized signal by creating a detectable, sometimes unique response to each particle interaction of interest. These experiments aim to minimize the background rates while maximizing desired signal acceptance. While many physical considerations are made to minimize background, including operating underground, increasing size, and purifying the detector medium, analysis techniques that can enhance signal sensitivity are crucial for both nuclear security and astroparticle physics alike. I present two developments of machine learning for classifying signals in rare event detectors for improved background rejection. One, the classification of scintillation versus ionization responses in XENONnT, a liquid xenon dark matter detector, employs a graphical model to classify signals probabilistically. This simple yet novel method further rejects backgrounds to dark matter and neutrino searches, improving our sensitivity. Two, for the WATCHMAN project, designed to demonstrate capable reactor antineutrino monitoring, machine learning will be shown to aid in otherwise unavoidable cosmogenic background rejection. Both case studies compare human-engineered features to raw-level data as classification inputs, demonstrating the extent to which basic classifiers can learn for themselves, from raw data only, the underlying physics which (we assume) generated the data. In some cases, building interpretable machine learning models enables the ability to understand for ourselves the physics governing a detector’s response.
Abstract Author(s): S. Farrell; A. Higuera, C. Peters (XENON); A. Bernstein, M. Bergevin, R. Wurtz (WATCHMAN).