1 of 12

Kinematic Analysis of TrackML Challenge Submissions

By Emily Ramey

Supervisors: Moritz Kiehn and Andreas Salzburger

Summer 2018

2 of 12

The TrackML Challenge

  • Goal is to use machine learning to improve the speed and accuracy of collision analysis
  • Participants are given a sample event simulated with the ATLAS detector
    • This event contains each detector hit as well as the true tracks for each particle
  • They must then use their algorithm to reconstruct a different event with only the hit data, not the truth file

3 of 12

Kinematic Analysis of Submissions

  • Participants are given a score based on how well they reconstruct the event
  • However, the raw score does not tell the full story
  • Further analysis is needed to have an idea of a submission’s true performance
  • Two kinds of kinematic analysis addressed:
    • Momentum dependence
    • Angular dependence

4 of 12

Momentum analysis (p┴)

  • High-pass filter on p┴, ending at 1 GeV
  • Scores get better with more particles excluded

Absolute score

Relative score (% original)

5 of 12

Angular analysis (η)

  • η is the pseudorapidity, related to the angle of a particle with respect to the beam axis
  • Three different kinds of η analysis
    • Start with the full angular region and decrease the bounds
    • Start with none of the region and expand the bounds
    • Analyze discrete sections of the region separately

6 of 12

Exclusive Region

7 of 12

p┴ vs. η analysis

  • Each η region is broken down into hits of different p┴
  • The result is a 2D image of which η and p┴ values are best for a certain submission

8 of 12

Future work

  • Optimize/parallelize code to minimize the memory-time tradeoff
  • Analyze submissions with only primary or secondary particles
  • Analyze submissions with other parameters (efficiency, time, etc.)

9 of 12

Thanks for watching!

Special thanks to my advisors, Moritz Kiehn and Andreas Salzburger, and to the NSF REU program, the CERN Studentships in Summer program for their support. Additional thanks to Steven Goldfarb, Junjie Zhu, Jennifer Roloff, Myron Campbell, and the University of Michigan program.

10 of 12

My Summer Experience

11 of 12

Credits

Sources:

Images:

12 of 12

Corrections to the simulation