The LEGEND experiment investigates the nature of neutrinos by searching for the potential neutrinoless double beta decay. This search uses high-purity germanium detectors enriched in 76 Ge. These detectors measure events as pulse shapes of excellent quality, which are analyzed to reconstruct the event’s energy and reject background-induced events.
In this thesis, pulses are analyzed with a deep neural network to determine the location of their respective events. The deep neural network used for the analysis consists of a long short-term memory model supervised by an attention mechanism to extract information from an event’s pulse. A fully connected classifier processes this extracted information to reconstruct the event’s position. To train and test the neural network, pulses are simulated from events at random positions using the LegendGeSim software. These simulated pulses occur as ideal charge pulses or as realistic pulses, which are similar to the pulses measured in the LEGEND experiment.
The experiments of this work consist of two main parts: The analysis of ideal simulated pulses and experiments with realistic simulated pulses. Both parts start by optimizing the deep neural network’s architecture, which was achieved using Bayesian optimization. Afterward, the position reconstruction performance of a trained neural network was analyzed. The final part of the analysis of ideal pulses consisted of experiments on the training consistency of different neural networks. The experiments on realistic simulated pulses were concluded by investigating the dependency of the position reconstruction potential on the rise time of the charge-sensitive amplifier.