Optimizing a Neural Network Approach to Pileup Rejection for CUPID*

The CUORE (Cryogenic Underground Observatory for Rare Events) experiment is searching for neutrinoless double beta decay at the Gran Sasso National Laboratory. Neutrinos are currently treated as Dirac fermions, particles distinct from their antiparticle, making this decay forbidden in the Standard Model. However, observation of neutrinoless double beta decay would indicate neutrinos are Majorana fermions, particles that are their own antiparticle. CUORE currently sets the best lower limit on the half-life for neutrinoless double beta decay in Tellurium-130. CUPID (CUORE Upgrade with Particle Identification) is an experiment in development that will look for neutrinoless double beta decay in Molybdenum-100 at much higher sensitivity.

One issue that arises from using ¹⁰⁰Mo as a source is its unusually high two-neutrino double beta decay rate. Due to the relatively slow response time of the detectors, two events that occur near simultaneously result in a “pileup” event, where the total sum of energy produced by the two constituent pulses can create a signal within the predicted region of interest of neutrinoless double beta decay. Neural Networks have demonstrated promise in this task, especially with convolutional neural networks (CNNs). However, higher efficiencies may be found in alternative model architectures and data representations. A thorough exploration of these alternative approaches is essential to determining the ideal pileup rejection method.