eprintid: 35980
rev_number: 14
eprint_status: archive
userid: 8739
dir: disk0/00/03/59/80
datestamp: 2025-01-28 07:22:45
lastmod: 2025-03-07 10:33:50
status_changed: 2025-01-28 07:22:45
type: doctoralThesis
metadata_visibility: show
creators_name: Sorrenson, Peter
title: Free-Form Flows: Generative Models for Scientific Applications
subjects: ddc-530
divisions: i-130300
adv_faculty: af-13
cterms_swd: Physics
cterms_swd: Maschinelles Lernen
abstract: This thesis explores advanced generative modeling techniques with a focus on scientific applications. It addresses two main areas: free-form flows and machine learning applications in particle physics.

Free-form flows are a novel family of generative models that combine the benefits of normalizing flows—exact maximum likelihood training and fast generation—with unrestricted neural network architectures. This approach overcomes the limitations of traditional normalizing flows, allowing for more flexible and domain-specific models.

The thesis also presents a collection of machine learning applications in particle physics, leveraging models with rich representational spaces. These techniques aim to accelerate the processing of vast data streams from the Large Hadron Collider, potentially uncovering new physics beyond the Standard Model.

By advancing both the theoretical foundations and practical implementations of generative models, this work contributes to their increased adoption and impact across diverse scientific disciplines. Applications range from chemistry to particle physics, demonstrating the broad potential of these methods in modeling complex data distributions.
date: 2025
id_scheme: DOI
id_number: 10.11588/heidok.00035980
ppn_swb: 1916509231
own_urn: urn:nbn:de:bsz:16-heidok-359800
date_accepted: 2025-01-22
advisor: HASH(0x55e83af56878)
language: eng
bibsort: SORRENSONPFREEFORMFL
full_text_status: public
place_of_pub: Heidelberg
citation:   Sorrenson, Peter  (2025) Free-Form Flows: Generative Models for Scientific Applications.  [Dissertation]     
document_url: https://archiv.ub.uni-heidelberg.de/volltextserver/35980/1/sorrenson_thesis.pdf