Machine learning approach for top pair reconstruction

This repository provides files used to create and execute the neural network used in the thesis of ref. [1] for reconstruction of top quark pairs from proton-proton collisions. The code is using the Pepper framework [2].

Installation

Required libraries are Pepper, Tensorflow (GPU installation recommended), and Scikit-hep's vector library.

Usage

1. Run the Processor (a Pepper processor) using process_reco.py. The required Pepper configuration file can be any Pepper configuration file for pp -> ttbar analyses.
2. Run prepare_data_reco.py, which creates normalized values for a range of observables the network is working with and saved them into HDF5 files.
3. Run model_bcls.py, which trains a network, whose purpose is to identify the correct pairing of bottom quarks.
4. Run model_tt.py, which trains a network to reconstruct the top quarks. It uses the network from the previous step.
5. The files plot_validation_bcls.py and plot_validation.py execute the networks on data and create plots to judge their performance.

References

[1] Rübenach, J. (2023) Search for heavy Higgs bosons in conjunction with neural-network-driven reconstruction and upgrade of the Fast Beam Condition Monitor at the CMS experiment. CERN-THESIS-2023-066
[2] Pepper contributors. Pepper - ParticlE Physics ProcEssoR. 2023. URL: https://gitlab.cern.ch/pepper/