A Physics-Informed Machine Learning Solution to the Pileup Problem in Cryogenic Bolometers

Cryogenic bolometry is a technique used for particle detection, measuring the energy of incident particles by measuring the change in the temperature of the material, taking advantage of the extremely low temperatures at which the bolometers are cooled. Examples of this technology are the CUORE and CUPID experiments searching for neutrinoless double beta decay. One problem that cryogenic bolometers such as CUPID face is the relatively slow response time of the detectors, leading to the possibility that two near-simultaneous events can appear to be a single event. This occurrence, called “pileup”, can present an irreducible source of background for the process of interest.

Neural Networks have demonstrated promise in rejecting these pileup events, especially Convolutional Neural Networks (CNNs). While current attempts convolve over the original pulses in one dimension, using the phase portrait of the pulse may better illuminate the differentiating features of pileup events, allowing a 2-dimensional CNN to extract these features. In addition, Physics-Informed Neural Networks (PINNs) present an opportunity to integrate existing models of the detector dynamics into the network architecture, balancing the learning capabilities of Neural Networks and the physical constraints defined by current models to improve the network's learning capabilities.