Advancing Long-Lived Particle Detection in the CMS Muon System Using ParticleNet-Based Dynamic Graph Neural Networks

Since the inception of the CERN Large Hadron Collider (LHC), the Standard Model (SM) has been substantially validated, yet several key questions remain unanswered, such as the origins of neutrino mass, the matter-antimatter imbalance, and the nature of dark matter. Extensions to the SM—including supersymmetry, hidden valley, and little Higgs models—suggest the existence of new long-lived particles (LLPs) potentially producible at the LHC. These LLPs are hypothesized to generate detectable decay cascades within the tracker, calorimeters, and muon subdetectors of the Compact Muon Solenoid (CMS).

Our research focuses on enhancing the detection capabilities of the CMS muon system for LLPs by analyzing hit clusters resulting from these decay showers. We have developed a practical and efficient classifier based on ParticleNet, a dynamic graph convolutional neural network architecture tailored for particle physics applications. This classifier processes the spatial and relational information within the recorded hit clusters more effectively than traditional methods.

By representing the hits as nodes in a dynamic graph, with their spatial relationships forming the edges, our network architecture captures the complex interaction patterns inherent in LLP events. The model is trained on centrally produced CMS Monte Carlo samples, enabling it to distinguish LLP signals from background noise with improved accuracy.

The application of ParticleNet-based dynamic graph neural networks has shown significant promise in advancing the search for new physics beyond the SM. Our findings indicate enhanced classification performance in detecting LLP events within the muon system. This work represents a step forward in leveraging advanced machine learning techniques for particle physics, potentially contributing to new discoveries in the field.