SPA-Net, a New Perspective for Parton-Jet Assignment?

Ta-Wei Ho^1*, Michael James Fenton², Alexander Shmakov³, Shih-Chieh Hsu⁴, Daniel Whiteson²

¹Department of Physics, National Tsing Hua University, Hsinchu City, Taiwan
²Department of Physics and Astronomy, University of California Irvine, California, USA
³Department of Computer Science, University of California Irvine, California, USA
⁴Department of Physics and Astronomy, University of Washington, Washington, USA

* Presenter:Ta-Wei Ho, email:davidho@gapp.nthu.edu.tw

The top quarks produced by pp collision in Large Hadron Collider (LHC), have a very complicated process still can't be well-classified today. In this project, we present a novel approach to the “all hadronic decay” process of Top quarks base on the neural networks with attention mechanism, we call it “Symmetry Preserving Attention Networks”(SPA-Net). This network identifies the decay products of each quarks unambiguously and without combinatorial explosion. This approach performs an outstanding result compare to the existing state-of-the-art method. Our network can correctly assign all hadronic decay in 93.0% of 6 jets, 87.8% of 7 jets, and 82.6% of  8 jets event respectively. This work is submitted to PRL (arXiv: 2010.09206)

Keywords: Machine Learning, Parton-Jet assignment, Top-pair production, Attention network