Generative Adversarial Networks for KamLAND-Zen

Robust and sophisticated tools to generate and handle big data have become increasingly important for low-background neutrino experiments searching for extremely rare events. Both simulated data and real detector data play essential roles for such experiments. Aside from traditional methods, such as Monte Carlo simulation, event generation can be supplemented by Generative Adversarial Networks (GANs). In a spherical liquid scintillator detector, such as KamLAND-Zen, the adoption of a canonical GAN model will introduce inevitable deviation from real detector data. Networks that can deal with spherical topology are required to eliminate this deviation. In this work, we will show that the autoencoder can learn the representation for a collection of events coming from a spherical liquid scintillator detector. We will also describe how a GAN network can generate simulated events for a spherical detector with incredible precision, such that the simulated data is indistinguishable from real detector data. The realistic detector simulation provided by a GAN is shown to improve the accuracy of a neural network classifier and enhance data-MC agreement at the same time.