Improving Pulse Shape Simulations with Generative Adversarial Machine Learning

The Large Enriched Germanium Experiment for Neutrinoless double-beta Decay (LEGEND) collaboration is searching for neutrinoless double-beta 0νββ decay in ⁷⁶Ge using modular arrays of germanium detectors enriched in the isotope. Candidate 0νββ events happen at a single site in the germanium detector. Pulse shape simulations to model the movement of charge carriers in the detectors are key to accurately modelling requirements that can reject background from multi-site and surface events. However, a series of corrections based on physical first principles are needed to correctly generate an accurate detector response. We present a novel neural network model called Cyclic Positional U-Net (CPU-Net) that enables translations of simulated pulses such that they are indistinguishable from actual detector pulses. The model uses a CycleGAN framework to develop an Ad-hoc Translation Network (ATN) for such translations with high precision and low latency. Using an HPGe detector, we demonstrate that the model correctly reproduces critical pulse shape reconstruction parameters, and that its data-driven nature enables generalization to multiple detectors and operating conditions without detector-wise model tuning.