Deep Graph Anomaly Detection with Contrastive Learning for New Physics Searches

This project proposed to develop a graph-based unsupervised Anomaly Detection model for identifying new physics at the LHC, combining contrastive learning with graph autoencoder reconstruction. While combining these two signals has proven effective for general graph anomaly detection, it has not yet been applied to particle collision data. Building on prior GENIE work on graph variational autoencoders, the project will explore graph construction strategies, GNN architectures, and contrastive learning objectives tailored to HEP data. Deliverables include a trained graph-based anomaly detection model, benchmark results with comparison to convolutional autoencoder baselines, a comprehensive report, documented codebase, and a blog post.