[Paper Review] LCFI Vertex Package

Sonja Hillert|arXiv (Cornell University)|Nov 28, 2008

Particle physics theoretical and experimental studies9 references3 citations

TL;DR

This paper presents the LCFI Vertex Package, a C++ software framework for vertex reconstruction, flavor tagging, and quark charge determination in high-precision vertex detectors for the International Linear Collider (ILC). It introduces ZVMST, a new minimum spanning tree-based vertex finder that improves c-jet tagging purity by up to 5% compared to the established ZVRES method at √s = 91.2 GeV, though with a minor 1.5% degradation in b-jet purity.

ABSTRACT

The LCFIVertex software, developed by the Linear Collider Flavour Identification (LCFI) collaboration and providing tools for vertexing, flavour tagging and quark charge determination for low-mass vertex detectors of high point resolution is presented. Particular emphasis is given to code extensions since the first release in April 2007. A recently developed new vertex finder, ZVMST, and its performance at sqrt(s) = 91.2 GeV are described in more detail.

Motivation & Objective

• To develop a robust, modular software package for vertex reconstruction and flavor tagging in high-precision vertex detectors for the ILC.
• To improve vertex finding performance in complex jet topologies, particularly for b- and c-jets with overlapping or close decay vertices.
• To introduce and evaluate a new vertex finder, ZVMST, based on minimum spanning tree algorithms, as an alternative to the established ZVRES method.
• To assess the impact of improved track-to-vertex assignment on flavor tagging performance using realistic GEANT4-based Monte Carlo simulations.
• To provide a validated, reusable software toolset for detector R&D and Letter of Intent preparation within the ILD and SiD detector concept groups.

Proposed method

• The LCFIVertex package is built on the LCIO event data model and integrated with the Marlin framework for modular, distributed software development.
• The core vertex finder ZVTOP employs three algorithms: ZVRES (vertex resolution-based), ZVKIN (kinematic-based for b-jets), and ZVMST (minimum spanning tree-based for candidate selection).
• ZVMST uses a minimum spanning tree to identify the most promising track combinations for vertex formation, improving robustness in high-multiplicity jets.
• Flavor tagging is performed using 9 dedicated neural networks—3 per vertex count category (1, 2, or ≥3 vertices)—trained on topological observables like vertex mass, decay length, and impact parameter significance.
• Performance is evaluated using a mixed jet sample at √s = 91.2 GeV, comparing ZVMST, ZVRES, and a 'cheater' that uses perfect track-to-vertex assignment from MC truth.
• The study uses GEANT4-based MOKKA simulations with the LDC01_05Sc detector model, 2.8 μm point resolution, and 0.1% X₀ material budget per layer.

Experimental results

Research questions

• RQ1 How does the performance of the new ZVMST vertex finder compare to the established ZVRES algorithm in terms of vertex multiplicity and reconstruction accuracy?
• RQ2 To what extent does ZVMST improve c-jet tagging purity compared to ZVRES, and what is the trade-off in b-jet tagging purity?
• RQ3 How does the quality of track-to-vertex assignment affect the final flavor tagging performance, and is ZVMST superior in this aspect?
• RQ4 Can the LCFIVertex software package serve as a reliable, reusable tool for detector optimization and LoI preparation in ILC detector concepts?
• RQ5 What are the limitations of current flavor tagging networks trained on fast MC (SGV) when applied to realistic GEANT4-based simulations?

Key findings

• ZVMST yields a vertex multiplicity closer to the MC truth than ZVRES, especially at high track multiplicities, with a crossover point at approximately 4 tracks per vertex.
• At √s = 91.2 GeV, ZVMST improves c-jet tagging purity by up to 5% across the efficiency range compared to ZVRES, indicating superior performance in identifying charm jets.
• ZVMST shows a minor degradation in b-jet tagging purity by up to 1.5% compared to ZVRES, suggesting a trade-off in favor of c-jet identification.
• The track-to-vertex assignment performance is similar between ZVMST and ZVRES, with some metrics improved in one and others in the other, indicating no clear overall advantage.
• The current flavor tagging networks, trained on fast MC (SGV), are suboptimal for realistic GEANT4 simulations, and performance differences may be reduced with algorithm-specific network retraining.
• Parameter optimization for ZVMST is preliminary, and further tuning is expected to further improve performance, especially in c-jet tagging.