EP-IT Data Science Seminars

Graph Neural Networks for High Luminosity Track Reconstruction

by Daniel Thomas Murnane (Lawrence Berkeley National Lab. (US))

Europe/Zurich
503/1-001 - Council Chamber (CERN)

503/1-001 - Council Chamber

CERN

162
Show room on map
Description

With the upgrade to HL-LHC, traditional algorithms in the event analysis pipeline may struggle to scale to meet throughput requirements, due to the density of detector data and incompatibility with modern heterogeneous parallelism. A promising alternative path is emerging, by treating detector data as a graph-like structure and applying Graph Neural Networks (GNNs) to learn a representation of the underlying physics. Subsequently, there has been a flurry of progress in understanding which GNN architectures are best suited to which stage of analysis - track reconstruction, jet tagging, seeding, event generation, and end-to-end analysis. 

While GNNs clearly bring a new way of handling HEP analysis, they can be cumbersome and expensive, and technologies are still being developed to handle them effectively. The Exatrkx Project is a collective effort across 10 institutions to study and validate ML approaches to tracking in HEP experiments, including ATLAS and DUNE. I will discuss how we are bringing the speed and accuracy of GNNs to these challenging datasets. In particular, highly GPU-optimized graph construction from O(100k)-sized point clouds and graph manipulation libraries are combined with state-of-the-art distributed ML training and inference techniques to deliver sub-second track reconstruction on high-luminosity datasets. This talk will cover many of these ideas, as well as ways that symmetry and representation-learning are included in the GNN models, and the progress being made on integrating graph techniques with existing tracking frameworks, like ACTS.
 

Daniel Murnane is a postdoctoral researcher at Lawrence Berkeley Lab. He is currently working with the Exatrkx project, a collaboration of US institutions that develops AI/ML techniques for track reconstruction, targeting exascale computing resources. Daniel's research focuses on graph neural networks (GNNs) for high energy physics problems, including building physics-informed & symmetry-aware GNNs for highly efficient performance. He received his PhD from the University of Adelaide, where he studied the phenomenology and fine-tuning of Composite Higgs models. 

Organised by

M. Girone, M. Elsing, L. Moneta, M. Pierini

Coffee will be served at 10h30

Videoconference
EP/IT Data Science Seminar
Zoom Meeting ID
98545267593
Description
EP/IT Data Science seminar
Host
Lorenzo Moneta
Alternative hosts
Thomas Nik Bazl Fard, EP Seminars and Colloquia, Maurizio Pierini, Caroline Cazenoves, Maria Girone, Markus Elsing, Pascal Pignereau
Passcode
97200142
Useful links
Join via phone
Zoom URL