Graph Neural Network based Hit Filtering for the Belle II Central Drift Chamber

Sumitted to PubDB: 2026-06-03

Category: Phd Thesis

Tags: TRG CDC Software

Principal Authors Greta Heine
Date 2026-04-15
Belle II Number BELLE2-PTHESIS-2026-022
Abstract The Belle~II experiment operates at high instantaneous luminosity, where an increasing level of beam-induced background poses significant challenges for both the offline and online track-reconstruction algorithms. This work presents the development, implementation, and performance evaluation of a hit-filtering algorithm based on graph neural networks for the central drift chamber of Belle II. The hit filter is designed for application in offline track reconstruction as well as in the Level-1 trigger tracking system, with a targeted deployment on FPGA devices. Applied to offline track reconstruction, the proposed algorithm improves track efficiency in Monte Carlo studies by up to 6.1%, reduces the track fake rate by up to 8.0%, and improves track resolution by up to 7.7% for key physics channels relative to the default filter, while maintaining comparable execution time. An adapted version of the algorithm applied to online triggering improves track efficiency by up to 18% evaluated on dimuon events from late 2025, while satisfying constraints on computing resources, trigger rates, and hardware utilization. In FPGA implementation studies evaluated for a graph size of 820 nodes and 3593 edges, the design utilizes 60.8% of look-up tables, 27.23% of flip-flops, and no DSPs at a 210.92ns end-to-end latency and 128.008MHz system frequency. The design thus satisfies the resource and sub-microsecond latency requirements for integration into the Belle II Level-1 trigger. Furthermore, this work establishes a neural network compression workflow for hardware software co-designed hit filtering in real-time trigger systems, including 4bit weight quantization, pruning, and the BOP metric as a quantitative approximation of the model size, which is applicable to the development of future machine-learning-based trigger designs.
Institute Karlsruhe

Files