Deep learning techniques for Imaging Air Cherenkov Telescopes

Abstract

Very High Energy (VHE) gamma rays and charged cosmic rays (CCRs) provide an observational window into the acceleration mechanisms of extreme astrophysical environments. One of the major challenges at Imaging Air Cherenkov Telescopes (IACTs) designed to look for VHE gamma rays, is the separation of air showers initiated by CCRs which form a background to gamma ray searches. Two other less well studied problems at IACTs are a) the classification of different primary nuclei among the CCR events and b) identification of anomalous events initiated by Beyond Standard Model particles that could give rise to shower signatures which differ from the standard images of either gamma rays or CCR showers. The problems of categorizing the primary particle that initiates a shower image, or the problem of tagging anomalous shower events in a model independent way, are problems that are well suited to a machine learning (ML) approach. Traditional studies that have explored gamma ray/CCR separation have used a multivariate analysis based on derived shower properties, which contains significantly reduced information about the shower. In our work, we address the problems outlined above by using ML architectures trained on full simulated shower images, as opposed to training on just a few derived shower properties. We illustrate the techniques of binary and multi-category classification using convolutional neural networks, and we also pioneer the use of autoencoders for anomaly detection at VHE gamma ray experiments. As a case study, we apply our techniques to the H.E.S.S. experiment. However, the real strength of the techniques that we broach here in the context of VHE gamma ray observatories, is that these methods can be applied broadly to any other IACT, such as the upcoming Cherenkov Telescope Array (CTA), or can even be suitably adapted to CCR experiments.