# Track Seeding and Labelling with Embedded-space Graph Neural Networks

**Authors:** Nicholas Choma, Daniel Murnane, Xiangyang Ju, Paolo Calafiura, Sean Conlon, Steven Farrell,  Prabhat, Giuseppe Cerati, Lindsey Gray, Thomas Klijnsma, Jim Kowalkowski, Panagiotis Spentzouris, Jean-Roch Vlimant, Maria Spiropulu, Adam Aurisano, V Hewes, Aristeidis Tsaris, Kazuhiro Terao, Tracy Usher  
**arXiv:** [2007.00149](https://arxiv.org/abs/2007.00149)  
**Date:** 2020-06-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2007.00149v1/

## Key points

- The <Organization>Exa.TrkX project</Organization> is focused on handling the large-scale data from the HL-LHC.
- Graph neural networks (GNNs) have been identified as a promising solution for particle track reconstruction.
- GNNs process event data as a graph, connecting track measurements and candidate line segments between hits.
- Message-passing GNNs have shown success in predicting doublet likelihood, leading to further exploration of advanced architectures.
- Advancements in graph construction and embedded representations have enabled fully learned end-to-end track finding.
- Construction of graphs from learned representations with non-linear metric structures improves clustering and neighborhood queries.
- The framework integrates with traditional clustering pipelines and GNN approaches, offering versatility in application.
- Embedded graphs feed into high-accuracy doublet and triplet classifiers or can be used as an end-to-end track classifier through clustering in an embedded space.
- Post-processing techniques informed by detector physics knowledge enhance performance significantly.

## Summary

The <Organization>Exa.TrkX project</Organization> is leading the way in tackling the immense challenge of handling the unprecedented scale of data from the HL-LHC. Through extensive exploration of various machine learning methods for particle track reconstruction, the project has identified graph neural networks (GNNs) as a highly promising solution. These GNNs process event data as a graph, connecting track measurements (represented by nodes) with candidate line segments between hits (represented by edges). By incorporating detector information into nodes and edges, GNNs can effectively propagate embedded parameters throughout the graph to predict node-, edge-, and graph-level observables. In previous work, message-passing GNNs have shown success in predicting doublet likelihood, prompting further investigation into state-of-the-art architectures for this task. The Exa.TrkX project has also made advancements in graph construction and embedded representations to achieve fully learned end-to-end track finding. This effort has resulted in a suite of extensions to the original model that demonstrate promising results for hitgraph classification. Additionally, progress has been made in constructing graphs from learned representations that incorporate non-linear metric structures. This enhancement enables efficient clustering and neighborhood queries of data points, ultimately improving performance. The framework seamlessly integrates with traditional clustering pipelines and GNN approaches, providing versatility in its application. Embedded graphs generated through these methods feed into high-accuracy doublet and triplet classifiers or can be utilized as an end-to-end track classifier through clustering in an embedded space. Post-processing techniques informed by knowledge of detector physics further enhance performance. Numerical results on the TrackML particle tracking challenge dataset highlight the effectiveness of this framework in both seeding and track finding tasks. With contributions from a diverse team of researchers including Nicholas Choma, Daniel Murnane, Xiangyang Ju, Paolo Calafiura, Sean Conlon, Steven Farrell, Prabhat, Giuseppe Cerati, Lindsey Gray, Thomas Klijnsma, Jim Kowalkowski, Panagiotis Spentzouris Jean-Roch Vlimant Maria Spiropulu Adam Aurisano V Hewes Aristeidis Tsaris Kazuhiro Terao Tracy Usher among others , this research represents a significant step forward in leveraging machine learning for particle track reconstruction in high-energy physics experiments.