Latent Space Clustering for Tracking in Belle II

One major task for the Belle II experiment is Track Finding, where hits are clustered into tracks presumably originating from the same particle. The CAT Finder (CDC AI Tracking) proposed by Lea Reuter et. al. poses an alternative to the current basf2 baseline for track finding in Belle II. The CAT Finder utilizes a Graph Neural Network (GNN) to project hits from the Central Drift Chamber (CDC) into a latent space, where they are subsequently clustered.

The goal of this work is to compare the implementation of multiple clustering algorithms in the CAT Finder pipeline in terms of the resulting track charge fitting efficiency. The current CAT Finder uses a simple clustering algorithm similar to K-Means. The results show that despite its simple implementation, its difficult to surpass the current CAT Finder implementation when it comes to the track charge fitting efficiency.