Pour vous authentifier, privilégiez eduGAIN / To authenticate, prefer eduGAINeu

Thèses

Aishik Ghosh "Simulation of the ATLAS Electromagnetic Calorimeter using Generative Adversarial Networks and Offshell Higgs Boson Coupling Measurements in the Four-Lepton Decay Channel at the LHC" (Pôle PHE)

Europe/Paris
Description
Lien de connexion / Link :
https://ijclab.zoom.us/j/92944485401?pwd=MzJoRjhnN2E2SkNQd1FhZ2JrUE5jUT09
ID de réunion : 929 4448 5401
Mot de passe : 889426
Merci de laisser fermés camera et micro de votre ordinateur pendant la soutenance.
Please keep your camera and microphone switched off throughout the defense. 
 
 
Title of Defense

Simulation of the ATLAS Electromagnetic Calorimeter using Generative Adversarial Networks and Offshell Higgs Boson Coupling Measurements in the Four-Lepton Decay Channel at the LHC

 
Abstract
 
Accurate simulations of particle showers consume the largest fraction of CPU time on the CERN computing grid. A Generative Adversarial Network, where one Generative Network is trained to fool two Adversarial Networks, is investigated as a scalable solution for fast simulation of the ATLAS calorimeter. Steps have been taken to inject physics knowledge into the training procedure of the network. For the first time, the integration into the ATLAS software allows a fair comparison with more traditional hand designed fast simulation algorithms. The synthesised showers show good agreement to showers from full detector simulation using Geant4 and the technique is more memory efficient than the traditional algorithm.

Quantum interference renders the notion of ’signal’ and ‘background’ classes ill-defined, forcing us to re-asses traditional analysis optimisation strategies. We perform a feasibility study to adapt the recently developed 'Likelihood-Free Inference based Madminer’ methods for an offshell Higgs couplings measurement in the four lepton decay channel in the context of the ATLAS experiment on the Large Hadron Collider at CERN. We show that such a physics-aware algorithm performs better than traditional approaches because it can automatically re-optimise itself over the full range of phase space under study.
 
 
Organisé par

Membres du jury :
• David Rousseau IJCLab CNRS, Directeur de thèse
• Marie Hélène Schune IJCLab, CNRS, Présidente du jury
• Isabelle Wingerter-Seez, LAPP-Annecy, CNRS, rapporteur
• Maurizio Pierini, CERN, rapporteur
• Tilman Plehn, professeur Université Heidelberg
• Glen Cowan professeur RHUL, London
• Danilo Rezende, Google Deepmind