Ensemble Models for Calorimeter Simulations

Abstract

Foreseen increasing demand for simulations of particle transport through detectors in High Energy Physics motivated the search for faster alternatives to Monte Carlo-based simulations. Deep learning approaches provide promising results in terms of speed up and accuracy, among which generative adversarial networks (GANs) appear to be particularly successful in reproducing realistic detector data. However, the GANs tend to suffer from different issues such as not reproducing the full variability of the training data, missing modes problem, and unstable convergence. Various ensemble techniques applied to image generation proved that these issues can be moderated either by deploying multiple generators or multiple discriminators. This work follows a development of a GAN with two-dimensional convolutions that reproduces 3D images of an electromagnetic calorimeter. We build on top of this model and construct an ensemble of generators. With each new generator, the ensemble shows better agreement with the Monte Carlo images in terms of shower shapes and the sampling fraction.