The aim of this project is to develop and test end-to-end machine learning methods for reconstruction and identification of hadronically decaying tau lepton, while also providing a thouroughly validated and tested dataset for evaluating the performances of said algorithms.
Tau leptons can decay both leptonically and hadronically, however only hadronic decays are targeted with this project:
The dataset contains 2 signal samples (ZH->Ztautau and Z->tautau) and one background sample (Z->qq). While the validation plots can be reproduced with this script, here is a selection of these:
The generator-level hadronically decaying tau visible transverse momentum:
The jet substructure of two neutral-hadronless decay modes:
The results of these studies have been divided across two separate papers, with the first one covering tau identification and the latter covering both kinematic and decay mode reconstruction.
"Tau lepton identification and reconstruction: a new frontier for jet-tagging ML algorithms"
[Published in: Comput.Phys.Commun. 298 (2024) 109095]
In this paper, we studied the performance of state-of-the-art methods and compared them with the ML architectures initially designed for jet-tagging.
"A unified machine learning approach for reconstructing hadronically decaying tau leptons"
Here we demonstrated how three different types of models with a varying degree of expressiveness and priors can be employed for hadronically decaying tau kinematic reconstruction and decay mode reconstruction.
