A New Identification Technique for High Energy Photons from Energy Signature in the Electromagnetic Calorimeter using Convolutional Neural Networks

Harilal, Abhirami (2018) A New Identification Technique for High Energy Photons from Energy Signature in the Electromagnetic Calorimeter using Convolutional Neural Networks. Masters thesis, Indian Institute of Science Education and Research Kolkata.

PDF (MS dissertation of Abhirami Harilal (13MS103)) 13MS103.pdf - Submitted Version Restricted to Repository staff only Download (7MB)

Abstract

This thesis aims to address an important problem in experimental high energy physicsto distinguish between the signature of a real photon from fake photons in the electromagnetic calorimeter (ECAL) in a particle detector. We try to distinguish prompt photon signals produced from hard collision of hadrons, from two kinds of fake signals - photons from bremsstrahlung of beam halo muons, and photons from the decay of neutral pions. We use their spatial energy deposition patterns in the ECAL crystals and employ novel methods in machine learning like Convolutional Neural Networks (CNN) which have a remarkable track record in solving spatial pattern recognition problems, along with Artificial Neural Networks (ANN), and compare their performances with conventional shower shape methods used. The analysis is done on data simulated from a flat geometry and a cylindrical geometry adaptation of the CMS ECAL with tracker. For the beam halo separation on a 10 GeV sample with a flat geometry, we achieve a background rejection of 99.99% for a signal efficiency of 99.9% using a CNN, while for the π⁰ - ɣ separation on a 10 GeV sample the maximum background rejection we got was 86.185% for a signal efficiency of 90% using a CNN. For a cylindrical geometry, beam halo separation by CNN gave a background rejection of 99.87% for a signal efficiency of 99.9% and for π⁰ - ɣ separation a maximum background rejection of 91.9% was obtained for a signal efficiency of 90% which is an improvement on the existing results.

Actions (login required)

View Item