Generating airshower images for the VERITAS telescopes with conditional Generative Adversarial Networks

TL;DR

This paper introduces a conditional GAN-based method to rapidly generate realistic airshower images for VERITAS telescopes, reducing reliance on costly simulations and aiding machine learning training.

Contribution

The study presents a novel cGAN approach to synthesize VERITAS airshower images conditioned on class labels, enabling fast and scalable data augmentation for astrophysical research.

Findings

cGANs can accurately replicate shower morphologies based on class conditions

The method generalizes signals through interpolation in class and latent spaces

Over one million shower events can be generated in under a minute

Abstract

VERITAS (Very Energetic Radiation Imaging Telescope Array System) is the current-generation array comprising four 12-meter optical ground-based Imaging Atmospheric Cherenkov Telescopes (IACTs). Its primary goal is to indirectly observe gamma-ray emissions from the most violent astrophysical sources in the universe. Recent advancements in Machine Learning (ML) have sparked interest in utilizing neural networks (NNs) to directly infer properties from IACT images. However, the current training data for these NNs is generated through computationally expensive Monte Carlo (MC) simulation methods. This study presents a simulation method that employs conditional Generative Adversarial Networks (cGANs) to synthesize additional VERITAS data to facilitate training future NNs. In this test-of-concept study, we condition the GANs on five classes of simulated camera images consisting of circular muon showers and gamma-ray shower images in the first, second, third, and fourth quadrants of the camera. Our results demonstrate that by casting training data as time series, cGANs can 1) replicate shower morphologies based on the input class vectors and 2) generalize additional signals through interpolation in both the class and latent spaces. Leveraging GPUs strength, our method can synthesize novel signals at an impressive speed, generating over $10^6$ shower events in less than a minute.