Particle Tracks Reconstruction via Graph Neural Networks
TrackForge is a from scratch Graph Neural Network (GNN) pipeline for reconstructing particle tracks in high energy physics (HEP) detectors. Hits become nodes, candidate connections become edges, and the model learns to classify which pairs belong to teh same helical trajectory.
Built by a CS major with a Physics minor, this project sits at the intersection of machine learning and particle physics. The goal is not to beat state of the art systems but to deeply understand every component, while documenting the journey (including dead ends).
It is designed to be accessible for others with similar backgrounds who find HEP intersting, and want to start out.
| Academic | Technical | Softwareengineering |
|---|---|---|
| Deepen GNN understanding | GNN + PyG pipeline on synthetic data | Improve Python Proficiency |
| Deepen particle physics intuition | Hit pairing / edge classification | Modular design for easy experimentation |
| Explore CUDA / performance-critical code |
This is not an attempt to compete with production track reconstruction systems (e.g., those used at LHC experiments). I deliberately avoided deep literature dives early on to develop my own intuition and solutions first.
I do not have formal physics qualifications (yet). Assumptions are clearly stated with reasoning. Performance is modest, but learning is massive.
Graph Encoders
- Simple GCN
- Simple GAT
Graph Attention
- Simple MLP
Everything is configurable via YAML files for rapid eexperimentation.
After cloning and cd-ing, you can just run make run to generate a synthetic dataset and start the model training
and evaluation.
Note that the generated dataset will be used for further runs. You can delete it with make clear-synthetic.
Regarding the dependecies, just try to run the model with small config parameters, and install what you are being
expected to.