Graph-Building and Input Feature Analysis for Edge Classification in the Central Drift Chamber at Belle II

This thesis investigates components of a Graph Neural Network (GNN)-based pipeline addressing challenges of current tracking algorithms at the Belle II experiment, such as highly displaced vertices and escalating beam backgrounds. The thesis examines input features for real-time pattern recognition algorithms, systematic graph-building in the Central Drift Chamber (CDC), which serves as the primary tracking detector, and the use of the Interaction Network (IN) for edge classification and background clean-up.

The study finds that Analog-to-Digital Converter (ADC) count and Time-to-Digital Converter (TDC) count, representing deposited energy and associated timing information in a CDC cell, provide orthogonal discrimination power, making them both valuable for distinguishing signal from background.

Different patterns for graph-building in the CDC are analyzed to find graphs that effectively encapsulate crucial information about signal particle tracks of a simulated Inelastic Dark Matter with a Dark Higgs model. Metrics are introduced to aid in balancing between capturing essential edges connecting signal hits and excluding those associated with background.

The graphs are evaluated, employing the IN as a classifier for graph edges. This process allows for the identification of signal hits and the execution of a background clean-up. The clean-up task yields a promising result, correctly identifying up to (80.6 ± 0.4) % of signal hits in the CDC while maintaining a purity of (67.4 ± 0.4) % in the hit selection. An initial analysis towards a real-time implementation is also conducted, aligning the input feature resolution with the anticipated resolutions at the Level 1 Trigger (L1 Trigger) stage.

This thesis provides encouraging evidence that a GNN-based pipeline offers a viable solution to the challenges posed.