| ATLAS Slides | |
| Report number | ATL-SOFT-SLIDE-2024-256 |
| Title | Improving Computational Performance of a GNN Track Reconstruction Pipeline for ATLAS |
| Author(s) |
Murnane, Daniel Thomas (Lawrence Berkeley National Lab. (US)) ; Dittmeier, Sebastian (Heidelberg University (DE)) ; Burleson, Jared Dynes (Univ. Illinois at Urbana Champaign (US)) ; Lazar, Alina (Youngstown State University (US)) ; Stark, Jan (Centre National de la Recherche Scientifique (FR)) ; Liu, Ryan (Lawrence Berkeley National Lab. (US)) ; Ju, Xiangyang (Lawrence Berkeley National Lab. (US)) ; Calafiura, Paolo (Lawrence Berkeley National Lab. (US)) ; Vallier, Alexis (Centre National de la Recherche Scientifique (FR)) ; Caillou, Sylvain (Centre National de la Recherche Scientifique (FR)) ; Pham, Minh Tuan (University of Wisconsin Madison (US)) ; Chan, Jay (Lawrence Berkeley National Lab. (US)) ; Neubauer, Mark (Univ. Illinois at Urbana Champaign (US)) ; Torres, Heberth (Centre National de la Recherche Scientifique (FR)) ; Collard, Christophe (Centre National de la Recherche Scientifique (FR)) |
| Corporate author(s) | The ATLAS collaboration |
| Submitted to | 22nd International Workshop on Advanced Computing and Analysis Techniques in Physics Research, Stony Brook, United States, 11 - 15 Mar 2024 |
| Submitted by | [email protected] on 19 Jul 2024 |
| Subject category | Particle Physics - Experiment |
| Accelerator/Facility, Experiment | CERN LHC ; ATLAS |
| Free keywords | Graph Neural Networks ; Inner Tracker Upgrade ; Track Reconstruction ; Tracking ; Machine Learning |
| Abstract | Track reconstruction is an essential element of modern and future collider experiments, including within the ATLAS detector. The HL-LHC upgrade of the ATLAS detector brings an unprecedented tracking challenge, both in terms of number of silicon hit cluster readouts, and throughput required for both high level trigger and offline track reconstruction. Traditional track reconstruction techniques often contain steps that scale combinatorically, which could be ameliorated with deep learning approaches. The GNN4ITk project has been shown to apply geometric deep learning algorithms for tracking to a similar level of physics performance with traditional techniques, while scaling sub-quadratically. In this contribution, we provide details of optimizations that reduce computational cost without significantly affecting physics performance. These include the use of structured pruning, knowledge distillation, simplified and customized convolutional kernels, regional tracking approaches, and GPU-optimized graph segmentation techniques. |
Record created 2024-07-19, last modified 2024-07-20
PDF