Distance-Weighted Graph Neural Networks on FPGAs for Real-Time Particle Reconstruction in High Energy Physics

Frontiers in Big Data

Mia Liu¹  Kevin Pedro¹  Sergo Jindariani¹  Nhan Tran²  Edward Kreinar³  Zhenbin Wu⁴  Giuseppe Di Guglielmo⁵  Jennifer Ngadiuba⁶  Javier Duarte⁷  Gianluca Cerminara⁸  Jan Kieseler⁸  Abhijay Gupta⁸  Maurizio Pierini⁸  Sioni Summers⁸  Gerrit Van Onsem⁸  Marcel Rieger⁸  Vladimir Loncar⁹  Yutaro Iiyama^1,10  Philip Harris^1,11  Dylan Rankin^1,11  Shah Rukh Qasim^1,12  Kinga Anna Wozniak^1,13 

[1] 0Department of Physics and Astronomy, Purdue university, West Lafayette, IL, United States;1Department of Electrical and Computer Engineering, Northwestern University, Evanston, IL, United States;2HawkEye360, Herndon, VA, United States;3Department of Physics, University of Illinois at Chicago, Chicago, IL, United States;Department of Computer Science, Columbia University, New York, NY, United States;Department of Physics, Math and Astronomy, California Institute of Technology, Pasadena, CA, United States;Department of Physics, University of California, San Diego, San Diego, CA, United States;Experimental Physics Department, European Organization for Nuclear Research (CERN), Geneva, Switzerland;Institute of Physics Belgrade, Belgrade, Serbia;International Center for Elementary Particle Physics, University of Tokyo, Tokyo, Japan;Laboratory for Nuclear Science, Massachusetts Institute of Technology, Cambridge, MA, United States;Manchester Metropolitan University, Manchester, United Kingdom;University of Vienna, Vienna, Austria;

关键词: deep learning;    field-programmable gate arrays;    fast inference;    graph network;    imaging calorimeter;   

DOI  :  10.3389/fdata.2020.598927

来源: DOAJ

【 摘 要 】

Graph neural networks have been shown to achieve excellent performance for several crucial tasks in particle physics, such as charged particle tracking, jet tagging, and clustering. An important domain for the application of these networks is the FGPA-based first layer of real-time data filtering at the CERN Large Hadron Collider, which has strict latency and resource constraints. We discuss how to design distance-weighted graph networks that can be executed with a latency of less than one μs on an FPGA. To do so, we consider a representative task associated to particle reconstruction and identification in a next-generation calorimeter operating at a particle collider. We use a graph network architecture developed for such purposes, and apply additional simplifications to match the computing constraints of Level-1 trigger systems, including weight quantization. Using the hls4ml library, we convert the compressed models into firmware to be implemented on an FPGA. Performance of the synthesized models is presented both in terms of inference accuracy and resource usage.

【 授权许可】

Unknown