Accelerated GNN Tracking
The tracking of charged particles produced in collisions at colliders is a crucial aspect of the science program in the experiments. One of the primary challenges for the HL-LHC is the ability to efficiently, accurately, and rapidly perform tracking in collision events with large interaction pile-up. This project aims to improve charged-particle tracking in the ATLAS and CMS experiments through the use of Geometric Deep Learn methods (particularly Graph Neural Networks (GNNs)) and hardware-based acceleration (currently focused on FPGAs). —
Most current GNN-based approaches to tracking proceed in three distinct stages: graphs are constructed from point cloud of hits in the tracker, the graphs are processed through a GNN to predict a score for each edge (high scores indicate that the edge like belongs to a true particle track, low scores indicate it is a spurious or noise edge), and finally a clustering or graph walk algorithm is used to group the high-scored edges into track candidates. We are studying innovations and optimizations at all three stages of this pipeline. We are also exploring alternate ‘one-shot’ architectures that are trainable end-to-end and go from point-clouds to track candidates with fit parameters in a single pass.
We are also studying two complimentary approaches to accelerate the inference of these GNN tracking pipelines on FPGAs with the goal of assessing feasibility of these algorithms for use at the trigger level at the HL-LHC.. The first uses the OpenCL Framework to optimize a co-processor approach where kernels are initiated on a CPU and executed on the FPGA. The second uses the hls4ml package to translate different components of the GNN pipeline directly into FPGA firmware code.
- Markus Atkinson
- Gage DeZoort
- Javier Duarte
- Abdelrahman Elabd
- Lindsey Gray
- Aneesh Heintz
- Jonathan Kutasov
- Mark Neubauer
- Isobel Ojalvo
- Caitlin Patterson
- Vesal Razavimaleki
- Savannah Thais
- Emily Tsai
- Bei Wang
- 2 Nov 2020 - "Tracking with GNN", Savannah Thais, CMS Tracking POG meeting (CMS internal)
- 21 Oct 2020 - "Accelerated Pixed Detector Tracklet Finding with GNNs on FPGAS", Savannah Thais, 4th Annual Inter-Experiment Machine Learning Workshop
- 30 Sep 2020 - "GNNs for Tracking", Savannah Thais, CMS Machine Learning Forum
- 27 Feb 2020 - "New Algorithms and Computing Architectures for Tracking", Savannah Thais, IRIS-HEP 18 Month Review
- 16 Oct 2019 - "Semantic Segmentation for CMS Pixel Clustering", Savannah Thais, US LHC Users' Association Meeting
- 12 Sep 2019 - "Introduction and Plans", Savannah Thais, IRIS-HEP Institute Retreat
- Charged particle tracking via edge-classifying interaction networks, G. DeZoort, S. Thais, J. Duarte, V. Razavimaleki, M. Atkinson, I. Ojalvo, M. Neubauer and P. Elmer, arXiv 2103.16701 (30 Mar 2021) [1 citation].
- Instance Segmentation GNNs for One-Shot Conformal Tracking at the LHC, S. Thais and G. DeZoort, arXiv 2103.06509 (21 Mar 2021).
- Performance of a geometric deep learning pipeline for HL-LHC particle tracking, X. Ju et. al., Eur.Phys.J.C 81 876 (2021) (11 Mar 2021) [2 citations].
- Accelerated Charged Particle Tracking with Graph Neural Networks on FPGAs, A. Heintz et. al., arXiv 2012.01563 (30 Nov 2020) [9 citations].