Frontiers in Big Data (Jan 2021)
Distance-Weighted Graph Neural Networks on FPGAs for Real-Time Particle Reconstruction in High Energy Physics
- Yutaro Iiyama,
- Gianluca Cerminara,
- Abhijay Gupta,
- Jan Kieseler,
- Vladimir Loncar,
- Vladimir Loncar,
- Maurizio Pierini,
- Shah Rukh Qasim,
- Shah Rukh Qasim,
- Marcel Rieger,
- Sioni Summers,
- Gerrit Van Onsem,
- Kinga Anna Wozniak,
- Kinga Anna Wozniak,
- Jennifer Ngadiuba,
- Giuseppe Di Guglielmo,
- Javier Duarte,
- Philip Harris,
- Dylan Rankin,
- Sergo Jindariani,
- Mia Liu,
- Kevin Pedro,
- Nhan Tran,
- Nhan Tran,
- Edward Kreinar,
- Zhenbin Wu
Affiliations
- Yutaro Iiyama
- International Center for Elementary Particle Physics, University of Tokyo, Tokyo, Japan
- Gianluca Cerminara
- Experimental Physics Department, European Organization for Nuclear Research (CERN), Geneva, Switzerland
- Abhijay Gupta
- Experimental Physics Department, European Organization for Nuclear Research (CERN), Geneva, Switzerland
- Jan Kieseler
- Experimental Physics Department, European Organization for Nuclear Research (CERN), Geneva, Switzerland
- Vladimir Loncar
- Experimental Physics Department, European Organization for Nuclear Research (CERN), Geneva, Switzerland
- Vladimir Loncar
- Institute of Physics Belgrade, Belgrade, Serbia
- Maurizio Pierini
- Experimental Physics Department, European Organization for Nuclear Research (CERN), Geneva, Switzerland
- Shah Rukh Qasim
- Experimental Physics Department, European Organization for Nuclear Research (CERN), Geneva, Switzerland
- Shah Rukh Qasim
- Manchester Metropolitan University, Manchester, United Kingdom
- Marcel Rieger
- Experimental Physics Department, European Organization for Nuclear Research (CERN), Geneva, Switzerland
- Sioni Summers
- Experimental Physics Department, European Organization for Nuclear Research (CERN), Geneva, Switzerland
- Gerrit Van Onsem
- Experimental Physics Department, European Organization for Nuclear Research (CERN), Geneva, Switzerland
- Kinga Anna Wozniak
- Experimental Physics Department, European Organization for Nuclear Research (CERN), Geneva, Switzerland
- Kinga Anna Wozniak
- University of Vienna, Vienna, Austria
- Jennifer Ngadiuba
- Department of Physics, Math and Astronomy, California Institute of Technology, Pasadena, CA, United States
- Giuseppe Di Guglielmo
- Department of Computer Science, Columbia University, New York, NY, United States
- Javier Duarte
- Department of Physics, University of California, San Diego, San Diego, CA, United States
- Philip Harris
- Laboratory for Nuclear Science, Massachusetts Institute of Technology, Cambridge, MA, United States
- Dylan Rankin
- Laboratory for Nuclear Science, Massachusetts Institute of Technology, Cambridge, MA, United States
- Sergo Jindariani
- 0Department of Physics and Astronomy, Purdue university, West Lafayette, IL, United States
- Mia Liu
- 0Department of Physics and Astronomy, Purdue university, West Lafayette, IL, United States
- Kevin Pedro
- 0Department of Physics and Astronomy, Purdue university, West Lafayette, IL, United States
- Nhan Tran
- 0Department of Physics and Astronomy, Purdue university, West Lafayette, IL, United States
- Nhan Tran
- 1Department of Electrical and Computer Engineering, Northwestern University, Evanston, IL, United States
- Edward Kreinar
- 2HawkEye360, Herndon, VA, United States
- Zhenbin Wu
- 3Department of Physics, University of Illinois at Chicago, Chicago, IL, United States
- DOI
- https://doi.org/10.3389/fdata.2020.598927
- Journal volume & issue
-
Vol. 3
Abstract
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.
Keywords
