Exploring the role of deep learning for particle tracking in high energy physics

Mayur Mudigonda, Dustin Anderson, Jean-Roch Vilmant, Josh Bendavid, Maria Spiropoulou, Stephan Zheng, Aristeidis Tsaris, Giuseppe Cerati, Jim Kowalkowski, Lindsey Gray, Panagiotis Spentzouris, Steve Farrell, Jesse Livezey, Prabhat, Paolo Calafiura

Feb 17, 2017 (modified: Feb 18, 2017) ICLR 2017 workshop submission readers: everyone
  • Abstract: Tracking particles in a collider is a challenging problem due to collisions, imperfections in sensors and the nonlinear trajectories of particles in a magnetic field. Presently, the algorithms employed to track particles are best suited to capture linear dynamics. We believe that incremental optimization of current LHC (Large Halidron collider) tracking algorithms has reached the point of diminishing returns. These algorithms will not be able to cope with the 10-100x increase in HL-LHC (high luminosity) data rates anticipated to exceed O(100) GB/s by 2025, without large investments in computing hardware and software development or without severely curtailing the physics reach of HL-LHC experiments. An optimized particle tracking algorithm that scales linearly with LHC luminosity (or events detected), rather than quadratically or worse, may lead by itself to an order of magnitude improvement in the track processing throughput without affecting the track identification performance, hence maintaining the physics performance intact. Here, we present preliminary results comparing traditional Kalman filtering based methods for tracking versus an LSTM approach. We find that an LSTM based solution does not outperform a Kalman fiter based solution, arguing for exploring ways to encode apriori information.
  • TL;DR: Particle tracking in LHC meets DL. Preliminary work, presenting the challenges in the field and exploring the role of learning based schemes for these problems.
  • Conflicts: lbl.gov, berkeley.edu, caltech.edu, fnal.gov