TY  - THES
AU  - McKeown, Peter
TI  - Development and Performance of a Fast Simulation Tool for Showers in High Granularity Calorimeters based on Deep Generative Models
IS  - DESY-THESIS-2024-008
PB  - Universität Hamburg
VL  - Dissertation
CY  - Hamburg
M1  - PUBDB-2024-01825
M1  - DESY-THESIS-2024-008
T2  - DESY-THESIS
SP  - 186
PY  - 2024
N1  - Dissertation, Universität Hamburg, 2024
AB  - Modern high energy physics experiments fundamentally rely on large quantities of simulateddata, placing significant demands on the available computational resources. Machine learningmethods based on deep generative models promise to reduce the compute time required tosimulate particle showers in the calorimeter system, which constitutes the most computationallyintensive part of a full detector simulation.This work focuses on the development of a first simulation tool based on deep generativemodels for shower simulation in highly granular calorimeters, and subsequently studies itsperformance in a realistic detector geometry. In order to apply these models in a generalsimulation, they must provide a suitable detector response for particles incident under variousangles to, and at various positions in, the detector. Crucially, the physics performance afterreconstruction must remain high, which is the ultimate target of such a simulator.We initially extend the performant Bounded Information Bottleneck Autoencoder (BIB-AE) to simulate showers from photons with varying incident energy and angle to the surface ofthe electromagnetic calorimeter of the International Large Detector (ILD), before studying thesingle particle performance of the model in terms of key calorimetric observables, both beforeand after reconstruction. We then further extend the model to handle an additional angle ofincidence, as well as taking steps to deal with geometry irregularities in order to allow the useof the model at different positions in the calorimeter.As a next step, we describe a generic library that enables the use of generative models withGeant4 and DD4hep, allowing a full integration into standard software ecosystems used inhigh energy physics. We outline the integration of the BIB-AE into this library, allowing afair benchmark of the computational performance of the model. We then simulate showers atdifferent positions with the model, in order to investigate the effects of performing simulationsin an irregular calorimeter geometry.Finally, we study the performance of the BIB-AE when used to simulate photons fromneutral pion decays in the process e<sup>+</sup>e<sup>−</sup>→ τ<sup>+</sup>τ<sup>−</sup> in terms of key physics observables. We findthat while some deviations from Geant4 occur, they are typically comparable to the MonteCarlo uncertainty, estimated from the performance differences between Geant4 versions.
LB  - PUB:(DE-HGF)3 ; PUB:(DE-HGF)11
DO  - DOI:10.3204/PUBDB-2024-01825
UR  - https://bib-pubdb1.desy.de/record/607309
ER  -