Exploring Different GNN Models For Hit-Cleanup At Belle II

This thesis presents a GCN as a drop-in replacement for the existing IN-based model used in hit clean-up for the CDC of the Belle II experiment.

We evaluate the proposed GCN model on simulated mu+mu-(gamma) and BBar events overlaid with the nominal Phase 3 background. Due to the larger number of background hits, this setting serves as a future estimate of detector conditions, as the background levels will rise with the increased luminosity. We compare it against existing methods, including IN approaches and MVA-based approaches. While the GCN underperforms in terms of track fitting efficiency and hit efficiency, it demonstrates slightly improved background suppression over the MVA, characterized by a reduced number of extra CDC hits. These are the hits left after track construction that are not associated with any fitted tracks. The GCN also shows competitive track fake rates and track clone rates. Although the computational expense of downstream tracking scales quadratically with the number of CDC hits, the GCN’s ability to reduce these extra hits may offer runtime benefits. However, this marginal gain comes at a significant cost in track and hit efficiency compared to the MVA.

Our results highlight key design insights into GCN architectures and provide a foundation for future development of hybrid or edge-aware GNN models for tracking applications in high-occupancy environments.