Speaker
Description
In this talk, I present a convolutional autoencoder trained to denoise pulses from a p-type point contact high-purity germanium (HPGe) detector. HPGe detectors are frequently used in the search for rare event interactions, such as neutrinoless double-beta decay, due to their intrinsic purity and excellent energy resolutions. However, electronic noise can make the identification of signal events challenging, especially at low energies. An effective denoising algorithm could help in identifying signals which would otherwise be too noisy to distinguish from backgrounds. It could also improve the measurements of pulse amplitudes.
I first describe the implementation details of the denoising autoencoder. This includes the motivation behind the network architecture and three training procedures, one of which requires simulations of the detector and two that require only noisy detector data. Results on both detector simulations and real data from an $^{241}$Am source show that the autoencoder is more effective than traditional denoising methods, does not distort the pulses, and can be used to improve the energy resolution under various circumstances.
Furthermore, the methodology described here can be easily extended to work with technologies other than HPGe detectors. For example, our group is beginning to apply these methods to spherical proportional counters and bubble chambers with promising results. Our group is also exploring the use of the latent representation from the encoder for other tasks including multi-site event discrimination and peak finding. This deep learning-based approach for denoising pulses is thus broadly applicable to the particle astrophysics community and beyond.
