Using conditional variational autoencoders to generate images from atmospheric Cherenkov telescopes

TL;DR

This paper demonstrates that conditional variational autoencoders can efficiently generate realistic Cherenkov telescope images of gamma-ray events, potentially reducing reliance on computationally expensive Monte Carlo simulations.

Contribution

It introduces a novel application of conditional variational autoencoders to generate Cherenkov telescope images conditioned on event size, improving simulation efficiency.

Findings

Generated images are classified as gamma with 98.4% accuracy.

The generated images match the size distribution of Monte Carlo images.

The size of generated images has an average error of 0.33.

Abstract

High-energy particles hitting the upper atmosphere of the Earth produce extensive air showers that can be detected from the ground level using imaging atmospheric Cherenkov telescopes. The images recorded by Cherenkov telescopes can be analyzed to separate gamma-ray events from the background hadron events. Many of the methods of analysis require simulation of massive amounts of events and the corresponding images by the Monte Carlo method. However, Monte Carlo simulation is computationally expensive. The data simulated by the Monte Carlo method can be augmented by images generated using faster machine learning methods such as generative adversarial networks or conditional variational autoencoders. We use a conditional variational autoencoder to generate images of gamma events from a Cherenkov telescope of the TAIGA experiment. The variational autoencoder is trained on a set of Monte Carlo events with the image size, or the sum of the amplitudes of the pixels, used as the conditional parameter. We used the trained variational autoencoder to generate new images with the same distribution of the conditional parameter as the size distribution of the Monte Carlo-simulated images of gamma events. The generated images are similar to the Monte Carlo images: a classifier neural network trained on gamma and proton events assigns them the average gamma score 0.984, with less than 3% of the events being assigned the gamma score below 0.999. At the same time, the sizes of the generated images do not match the conditional parameter used in their generation, with the average error 0.33.