TrackerGNN

Graph neural networks for charged particle tracking with detector hits.

TrackerGNN is a new HEP/ML project focused on charged particle track reconstruction in the high-occupancy environment expected at the HL-LHC. In modern silicon trackers, each collision event can contain many thousands of hits, and the number of possible hit combinations grows rapidly with pileup. This makes tracking a natural setting for graph-based learning: detector hits can be represented as nodes, candidate geometric relationships as edges, and particle trajectories as connected subgraphs.

The project will use the TrackML dataset as a public benchmark for studying graph construction, edge classification, and track building workflows inspired by the Exa.TrkX program.

Physics Motivation

Charged particles bend in the magnetic field of a collider detector and leave localized hits across tracking layers. The reconstruction task is to group those hits into physically valid trajectories while rejecting accidental combinations from other particles and detector noise.

At the HL-LHC, increased pileup will make classical combinatorial tracking more expensive. Graph neural networks offer a way to learn local detector geometry, hit compatibility, and trajectory structure directly from data while keeping the workflow aligned with physics constraints such as detector layers, curvature, and momentum-dependent bending.

Planned ML Pipeline

TrackML events
        |
        v
Detector hits, cells, particles, truth
        |
        +--> Physics and geometry preprocessing
        |
        v
Candidate graph construction
        |
        v
GNN edge classification
        |
        v
Track candidate assembly
        |
        v
Physics evaluation and reconstruction metrics

Planned stages:

• Load TrackML event files and organize hit, truth, and particle tables.
• Build detector-aware hit features using coordinates, layers, modules, and simple helix-inspired geometry.
• Construct candidate hit graphs with local geometric cuts or learned filters.
• Train PyTorch Geometric models for edge classification and message passing.
• Assemble high-confidence edges into track candidates.
• Evaluate tracking efficiency, fake rate, duplicate rate, and physics dependence versus transverse momentum and pseudorapidity.

Current Results

The current development benchmark uses TrackML event000001080. The first sparse graph contains 5000 nodes and 73623 candidate edges, with 765 positive edges. This gives a positive edge fraction of 1.04%, making the task strongly imbalanced.

Model	F1	Average precision
Random Forest baseline	0.9737	0.9959
First EdgeGNN	0.1286	0.0682
EdgeGNN with edge features	0.1278	0.0780
Balanced negative sampling EdgeGNN	0.1263	0.0694

The geometry-based classical baseline is extremely strong on this first sparse graph. The first GNN variants recover many true edges but still produce too many false positives at the default threshold. See docs/current_results.md for interpretation and next steps.

Repository Layout

configs/              Configuration files for experiments and graph building
data/raw/             Original TrackML files
data/processed/       Derived tables, graph datasets, and cached artifacts
docs/                 Notes on detector geometry, datasets, and methods
figures/              Plots and event displays
notebooks/            Exploratory analysis notebooks
reports/              Written analysis outputs
slides/               Presentation material
scripts/              Reproducible command-line workflows
src/trackergnn/       Python package source

Future Work Ideas

• Reproduce a minimal Exa.TrkX-style edge-classification baseline.
• Compare geometric graph construction strategies across detector regions.
• Add track-building from classified edges using connected components or simple graph-walk algorithms.
• Study performance as a function of particle momentum, pseudorapidity, and detector layer occupancy.
• Explore interaction networks, graph attention, and heterogeneous detector graph representations.
• Add scalable training utilities for larger TrackML event samples.