Convolution and Graph-based Deep Learning Approaches for Gamma/Hadron Separation in Imaging Atmospheric Cherenkov Telescopes

TL;DR

The paper tackles $γ$-ray vs hadron separation in Imaging Atmospheric Cherenkov Telescopes and proposes a CNN-GNN framework that jointly processes per-telescope images and graph-structured event context. It introduces two enhanced variants—one incorporating reconstructed core information—and a Hillas-parameter-based GNN baseline to benchmark performance. On simulated H.E.S.S. data, the CNN-GNN models improve discrimination over Hillas-based methods, with the Strong CNN-GNN achieving an additional ~9% reduction in false positives at a fixed threshold due to the reconstructed core position, suggesting better generalization to real observations. The work highlights the value of permutation-equivariant, graph-based architectures for cross-telescope information integration and outlines future steps to test on real data and include richer contextual features like zenith/azimuth to bridge simulation-reality gaps.

Abstract

The identification of $γ$-rays from the predominant hadronic-background is a key aspect in their ground-based detection using Imaging Atmospheric Cherenkov Telescopes (IACTs). While current methods are limited in their ability to exploit correlations in complex data, deep learning-based models offer a promising alternative by directly leveraging image-level information. However, several challenges involving the robustness and applicability of such models remain. Designing model architectures with inductive biases relevant for the task can help mitigate the problem. Three such deep learning-based models are proposed, trained, and evaluated on simulated data: (1) a hybrid convolutional and graph neural network model (CNN-GNN) using both image and graph data; (2) an enhanced CNN-GNN variant that incorporates additional reconstructed information within the graph construction; and (3) a graph neural network (GNN) model using image moments serving as a baseline. The new combined convolution and graph-based approach demonstrates improved performance over traditional methods, and the inclusion of reconstructed information offers further potential in generalization capabilities on real observational data.

Figures (4)

• Figure 1: Schematic visualization of the image processing pipeline of images in an event for input into a convolutional neural network.
• Figure 2: Sketch illustrating the shift in coordinate frames with the new origin centered at the reconstructed core position of an event. The orange nodes indicate four IACT telescopes with the black lines indicating the edges of the graph structure. The blue cross marks the reconstructed impact point of the primary particle. This transformation is used for the inclusion of reconstructed information in the Strong CNN-GNN approach.
• Figure 3: Receiver Operating Characteristic (ROC) curves for the three models, with the Area Under Curve (AUC) values indicated in the legend.
• Figure 4: Permutation Feature Importance scores for the Hillas GNN and Fast CNN-GNN (split-training) models. A higher value indicates greater importance of that feature in the model's performance.