Machine Learning Denoising of nEXO's Time Project Chamber using an Unsupervised Variational Autoencoder

The nEXO experiment will use a single-phase liquid xenon time projection chamber (TPC) to search for neutrinoless double beta decays of 136Xe. As a single-phase TPC, nEXO will observe drift electrons using sensing electrodes, reducing complexity but increasing electronics noise relative to the two-phase approach. While nEXO is robust against this noise at the 2.5 MeV signal energy of interest and its multi-variate signal/background separation, advanced mitigations of this noise will enhance nEXO’s scientific reach by improving its energy resolution and topological background discrimination.

This talk presents the preliminary results of a method for denoising nEXO’s charge sensors using unsupervised machine learning. Because nEXO’s charge signal depends only on the position and shape of a single charge deposit, the charge waveforms can be described using a physically-motivated small number of dimensions. Using a variational autoencoder, these signals can be encoded into this low-dimensional latent space and then decoded, naturally suppressing the random noise which cannot be described in so few dimensions.

This approach can operate fully unsupervised on experimental output without requiring simulations to train the autoencoder.