One-Shot Neural Inference for Latent Space Clustering in GNN-Based Track Finding at Belle II

This thesis proposes a one-shot neural clustering algorithm for object condensation pipelines. The algorithm groups an unordered input point cloud of variable count with latent space representations of the input hits into a variable number of clusters and differentiates between signal and noise. Implementation and evaluation is done in the graph neural network based track-finding pipeline for the Belle II central drift chamber. In a first
step, the clustering pipeline and machine learning architecture are presented, combining key solutions from different computer vision pipelines. Next, the required hyperparameters are fitted using simulated detector samples, showing a good generalization across vastly differing event topologies. Then, the clustering is evaluated on a diverse set of
challenging events, leveraging a realistic detector simulation combined with high levels of detector noise measured in actual collisions. Additional validation is performed directly on measured collision data with high levels of detector noise, showing real-world applicability and robustness. The evaluation shows a significant improvement in track charge efficiency
and clone rate, while retaining a similar fake rate. When benchmarked against state of the art classical clustering algorithms, the one-shot neural inference provides a significant speedup without sacrificing clustering quality.