Book/Dissertation / PhD Thesis

PUBDB-2024-05063

http://join2-wiki.gsi.de/foswiki/pub/Main/Artwork/join2_logo100x88.png Generative Modelling in High-Energy Physics

Käch, B. (Corresponding author)DESY* ; Melzer-Pellmann, I. (Thesis advisor)DESY* ; Kasieczka, G. (Thesis advisor)Extern*

2024
Verlag Deutsches Elektronen-Synchrotron DESY Hamburg

This record in other databases:  

Please use a persistent id in citations: urn:nbn:de:gbv:18-ediss-119360  doi:10.3204/PUBDB-2024-05063

Report No.: DESY-THESIS-2024-011

Abstract: Monte Carlo simulations are an important tool in high-energy physics, e.g. to test the predictions of theory models or to infer a priori unknown parameters of the models. However, these simulations demand a substantial amount of computational resources. Thus, this thesis explores the viability of neural-network-based generative models for the CMS experiment at the LHC. The upcoming upgrades of the LHC further challenge the available computing budget of the collaboration in the near future, as these upgrades are expected to substantially increase the number of recorded collisions, necessitating a corresponding expansion in Monte Carlo simulation. For the CMS experiment, the simulation of a single event currently requires approximately two minutes. However, the required time is further expected to at least double, due to upgrades of the CMS detector. This increase is mainly owed to the upgrade of the endcap calorimeters, where the resulting number of channels that need to be simulated will be significantly higher.First, studies on the high-energy physics community $\rm{JetNet}$ dataset are extensively discussed, and the performance of different generative models is compared. An attention-based information aggregation, which scales linearly with the number of particles in terms of computational complexity, is proposed. Not only does this lead to state-of-the-art results on the \texttt{JetNet} datasets, but also promising results on the CaloChallenge. Finally, the viability of an end-to-end generation approach is studied in a search for Supersymmetry. The semi-leptonic decay of gluinos, produced via pair production, to neutralinos with an intermediate chargino in the decay chain is investigated. In this search, three a priori unknown parameters need to be scanned, which correspond to the masses of the superpartners. Typically, the mass of the intermediate particle is not scanned, since it is not feasible to generate Monte Carlo simulated data for all parameter combinations. The consequences of not scanning the chargino mass when a neural-network-based classifier is used to identify a signal pure region of the phase space for the statistical inference are investigated. Instead of fixing the mass to an arbitrary value, it is explored whether synthetic data from a generative model, which transforms the distributions from different values of the chargino mass into another, improves the statistical significance of the search. In this study, the required integrated luminosity to reach a similar statistical significance is reduced by $20\pm12\%$.

Note: Dissertation, University of Hamburg, 2024

---

Contributing Institute(s):

1. LHC/CMS Experiment (CMS)

Research Program(s):

1. 611 - Fundamental Particles and Forces (POF4-611) (POF4-611)
2. PHGS, VH-GS-500 - PIER Helmholtz Graduate School (2015_IFV-VH-GS-500) (2015_IFV-VH-GS-500)

Experiment(s):

1. LHC: CMS

Appears in the scientific report 2024

---

Database coverage:
;

---