Can a machine learn to peak through the noise?: Implementation of a Deep Neural Network for signal denoising in the search for the Majorana neutrino.

We evaluate the implementation of a Denoising Autoencoder deep neural network (DNN) to remove the noise from a Silicon Photomultiplier (SiPM) detector in a prototype liquid Xenon experiment at UMass Amherst. This is part of the nEXO collaboration's work toward fully characterizing the SiPM's properties and maximizing the energy resolution of the next search for a neutrinoless double beta decay. We expect to improve the current single photon event (SPE) analysis process by substituting the existent baseline correction and frequency filters with the DNN. The right combination of training hyperparameters, together with a large enough, high-quality training dataset, seems to promise a cleaner analysis output. The training and validation data consists of synthetic waveforms generated using Monte Carlo methods. We extracted real signals to produce a pulse template and noise baselines to produce a large set of clean and corresponding noised waveforms with a varying number of pulses of different amplitudes. Preliminary tests show improved SPE analysis for some datasets. However, we still need to improve the retention of peak amplitude in lower signal/noise cases. A successful implementation of this model could help reduce uncertainties in the experimental measurements, as well as the amount of time and training required for new researchers to analyze data from this prototype cell. This would increase the efficiency of the team and lay the groundwork for other teams and experiments to apply similar DNNs.