Computational Performance of the ATLAS ITk GNN Track Reconstruction Pipeline

ATLAS Collaboration

Report

PUBDB-2024-07907

Computational Performance of the ATLAS ITk GNN Track Reconstruction Pipeline

ATLAS Collaboration (Collaboration author)DESY*

2024

9 pp. (2024) [10.3204/PUBDB-2024-07907]

This record in other databases:

Please use a persistent id in citations: doi:10.3204/PUBDB-2024-07907

Report No.: ATL-PHYS-PUB-2024-018

Abstract: The ATLAS event reconstruction chain is projected to increase dramatically in computational cost with the upgrade to the HL-LHC. A particularly expensive step in this chain is track finding, where energy deposits in the inner tracker (ITk) are grouped into subsets of track candidates, which can then be fitted and provided for downstream tasks. In an effort to reduce execution times and harness accelerator hardware such as GPUs, for both offline and online purposes, machine learning approaches are being developed for track finding. A first functional implementation of a graph neural network-based track pattern reconstruction for ITk has been developed, with competitive physics performance compared with traditional methods. This document describes a variety of improvements to the algorithmic implementations and machine learning models, to significantly decrease execution times from minutes to hundreds of milliseconds.

Keyword(s): ITk ; Track Reconstruction ; Graph Neural Network ; Timing ; Computational Performance ; Upgrade ; TRACKING ; FUTURE

Note: All figures including auxiliary figures are available at https://atlas.web.cern.ch/Atlas/GROUPS/PHYSICS/PUBNOTES/ATL-PHYS-PUB-2024-018. The authors list may be incomplete!

Contributing Institute(s):