New Physics searches at the LHC using Anomaly Detection

Collider experiments provide a powerful means of exploring high-energy physics and identifying potential signatures of new physics. This study contributes to these efforts in a model-agnostic way, using semi-supervised learning approaches for anomaly detection. Various searches for new physics focus on fully hadronic final states, with jets serving as probes for potential signals.

This work specifically targets anomaly detection at the jet level. Each jet is represented as a graph, with nodes corresponding to its hadronic constituents. Simulated datasets of Dark Jets events, framed within a dark matter model, serve as the benchmark signal, where a heavy vector boson Z' mediator connects a Standard Model quark pair with a pair of dark quarks.

These quarks then shower and hadronize, producing dark jets. The background consists of QCD dijet events. The objective is to extract a vector embedding that maps high-dimensional graph information into a low-dimensional vector using convolution and pooling mechanisms. This embedding serves as input to an AD method, such as DeepSVDDs and Autoencoders, allowing for jet prediction and classification based on anomaly scores. Performance comparisons are conducted against baseline deep learning approaches.