This thesis presents the development and implementation of two innovative jet tagging algorithms designed for the identification of highly boosted H → γγ decays using the ATLAS detector at the Large Hadron Collider (LHC). These algorithms leverage two neural network architectures to enhance the identification process. The code for this work is contained in this repository.
This thesis was completed in three months, completing a piece of work on this scale was not something I had done before so whilst I have attempted to tidy this repo up vs my private repo, it is still quite messy. I recommend reading the thesis, at least the methodology section, before looking at the code.
Additionally, this work required significant compute resources (Nvidia A100) and time (approx. 1 week per NN).
- Performance: Comparable to existing algorithms designed for highly boosted Z → e+e− decays.
- Features:
- Multifunctional, effective in also identifying Z → e+e− decays.
- High classification rates, comparable to another DNN jet tagging algorithm for highly boosted heavy bosons
- Architecture: Utilizes an Adversarial Neural Network for mass-decorrelated classification.
- Noteable results:
- Very slight performance decrease compared to DNN-based tagger.
- Also functional as a dual-use jet tagger.
- Achieves a 27.8% reduction in mutual information between the mass feature and scalar discriminant metric.
- Demonstrates enhanced rejection rates for background (τ)τ-jets.
Read the thesis on the CERN document server.
@thesis{Hey:2878576,
author = "Hey, Nathaniel",
title = "{Identification of highly boosted H → γγ decays with
the ATLAS detector using deep neural networks}",
school = "University of Edinburgh",
year = "2023",
url = "https://cds.cern.ch/record/2878576",
note = "Presented 29 Sep 2023",
}