Zuzolo, Paolo
(2026)
Mathematical methods in deep learning for geometry processing, [Dissertation thesis], Alma Mater Studiorum Università di Bologna.
Dottorato di ricerca in
Matematica, 38 Ciclo. DOI 10.48676/unibo/amsdottorato/13295.
Documenti full-text disponibili:
![zuzolo_paolo_tesi.pdf [thumbnail of zuzolo_paolo_tesi.pdf]](https://amsdottorato.unibo.it/style/images/fileicons/application_pdf.png) |
Documento PDF (English)
- Richiede un lettore di PDF come Xpdf o Adobe Acrobat Reader
Disponibile con Licenza: Salvo eventuali più ampie autorizzazioni dell'autore, la tesi può essere liberamente consultata e può essere effettuato il salvataggio e la stampa di una copia per fini strettamente personali di studio, di ricerca e di insegnamento, con espresso divieto di qualunque utilizzo direttamente o indirettamente commerciale. Ogni altro diritto sul materiale è riservato.
Download (54MB)
|
Abstract
Geometry processing provides mathematical models and algorithms to represent, analyze, and manipulate geometric data, usually smooth 2–manifolds in \mathbb{R}^3, represented by polygonal meshes. Classical methods rely on discrete differential geometry, variational formulations, and geometric PDEs: curvatures, the Laplace–Beltrami operator, and energy-minimizing flows are used to smooth, denoise, parameterize, and build stable descriptors. Their strength is structural: invariances are encoded explicitly, and numerical discretizations are designed to preserve them, yielding interpretable and reliable methods. Yet classical pipelines lose efficiency and accuracy under adverse meshing, discretization changes, and demands for compact, task-relevant representations. Deep learning has advanced geometry processing by learning task-dependent representations from data, enabling progress in shape analysis. However, transferring learning pipelines to geometric domains is challenging: surfaces are non-Euclidean and irregularly sampled, discretizations vary widely, and learned operators may be sensitive to triangulation. Learning-based methods nonetheless offer a complementary advantage, adapting representations to task objectives and promoting robustness across heterogeneous discretizations. This thesis bridges these paradigms, treating classical geometric models not as competitors to learning, but as principled inductive biases for it. Following Physics-Inspired and Physics-Informed Neural Networks, learning should be guided by governing equations and variational principles.
In geometry processing, the relevant “physics” is geometric—intrinsic PDEs on 2-manifolds and energy functionals encoding regularity, locality, and feature preservation. Since the domain is typically available only as a mesh, we focus on discrete physics-informed formulations, where constraints are imposed via mesh-based differential operators, and graph neural networks provide a natural framework.
The thesis develops mathematical methods for deep learning on manifolds, combining spectral geometry, variational energies, and geometric flows with modern neural architectures.
Contributions include discrete physics-informed neural approaches to Laplace–Beltrami spectral quantities, energy-driven spectral graph convolutions derived from variational formulations, and flow-based multi-scale surface decomposition. The resulting framework aims to deliver expressive yet grounded models: intrinsic, discretization-robust, and interpretable.
Abstract
Geometry processing provides mathematical models and algorithms to represent, analyze, and manipulate geometric data, usually smooth 2–manifolds in \mathbb{R}^3, represented by polygonal meshes. Classical methods rely on discrete differential geometry, variational formulations, and geometric PDEs: curvatures, the Laplace–Beltrami operator, and energy-minimizing flows are used to smooth, denoise, parameterize, and build stable descriptors. Their strength is structural: invariances are encoded explicitly, and numerical discretizations are designed to preserve them, yielding interpretable and reliable methods. Yet classical pipelines lose efficiency and accuracy under adverse meshing, discretization changes, and demands for compact, task-relevant representations. Deep learning has advanced geometry processing by learning task-dependent representations from data, enabling progress in shape analysis. However, transferring learning pipelines to geometric domains is challenging: surfaces are non-Euclidean and irregularly sampled, discretizations vary widely, and learned operators may be sensitive to triangulation. Learning-based methods nonetheless offer a complementary advantage, adapting representations to task objectives and promoting robustness across heterogeneous discretizations. This thesis bridges these paradigms, treating classical geometric models not as competitors to learning, but as principled inductive biases for it. Following Physics-Inspired and Physics-Informed Neural Networks, learning should be guided by governing equations and variational principles.
In geometry processing, the relevant “physics” is geometric—intrinsic PDEs on 2-manifolds and energy functionals encoding regularity, locality, and feature preservation. Since the domain is typically available only as a mesh, we focus on discrete physics-informed formulations, where constraints are imposed via mesh-based differential operators, and graph neural networks provide a natural framework.
The thesis develops mathematical methods for deep learning on manifolds, combining spectral geometry, variational energies, and geometric flows with modern neural architectures.
Contributions include discrete physics-informed neural approaches to Laplace–Beltrami spectral quantities, energy-driven spectral graph convolutions derived from variational formulations, and flow-based multi-scale surface decomposition. The resulting framework aims to deliver expressive yet grounded models: intrinsic, discretization-robust, and interpretable.
Tipologia del documento
Tesi di dottorato
Autore
Zuzolo, Paolo
Supervisore
Co-supervisore
Dottorato di ricerca
Ciclo
38
Coordinatore
Settore disciplinare
Settore concorsuale
Parole chiave
graph neural networks, Laplace Beltrami eigendecomposition, surface processing, physics-informed discrete Neural Network, Normal Total-Variation flow, geometric flow, multi-scale geometric analysis, Finite Volume Method, Gauss map
DOI
10.48676/unibo/amsdottorato/13295
Data di discussione
17 Luglio 2026
URI
Altri metadati
Tipologia del documento
Tesi di dottorato
Autore
Zuzolo, Paolo
Supervisore
Co-supervisore
Dottorato di ricerca
Ciclo
38
Coordinatore
Settore disciplinare
Settore concorsuale
Parole chiave
graph neural networks, Laplace Beltrami eigendecomposition, surface processing, physics-informed discrete Neural Network, Normal Total-Variation flow, geometric flow, multi-scale geometric analysis, Finite Volume Method, Gauss map
DOI
10.48676/unibo/amsdottorato/13295
Data di discussione
17 Luglio 2026
URI
Statistica sui download
Gestione del documento: