Variational Autoencoders for Jet Simulation

Abstract

We introduce a novel variational autoencoder (VAE) architecture that can generate realistic and diverse high energy physics events. The model we propose utilizes several techniques from VAE literature in order to simulate high fidelity jet images. In addition to demonstrating the model's ability to produce high fidelity jet images through various assessments, we also demonstrate its ability to control the events it generates from the latent space. This can be potentially useful for other tasks such as jet tagging, where we can test how well jet taggers can classify signal from background for events generated by the VAE. We test this idea by seeing the signal efficiency vs background rejection for different types of jet images produced by our model. We compare our VAE with generative adversarial networks (GAN) in several ways, most notably in speed. The architecture we propose is ultimately a fast, stable, and easy-to-train deep generative model that demonstrates the potential of VAEs in simulating high energy physics events.