Back to search

Boston University

Some statistical contributions to physics-informed neural networks and energy-based models

Abstract

dc:description.abstract

This dissertation presents several statistical contributions to the modeling of high-dimensional systems, focusing specifically on Physics-Informed Neural Networks (PINNs) and Energy-Based Models (EBMs). In the first part, the research examines Bayesian Physics-Informed Neural Networks (BPINNs), which integrate physical constraints directly into deep learning frameworks. This work investigates how BPINNs infer unknown parameters from noisy or incomplete observational data for PDE-based models, providing a detailed theoretical analysis of the posterior distribution’s convergence properties. Particularly, it demonstrates that, as the amount of data increases, the posterior distribution of parameters approaches a Gaussian distribution centered at the true parameters. Additionally, the study explores BPINNs in scenarios involving functional parameters that vary across space or time, enhancing understanding of their potential in modeling complex physical phenomena. In the second part, the dissertation investigates EBMs, a flexible class of statistical models that define distributions through energy functions, which are often computationally challenging due to intractable normalizing constants. This research explores short-run Markov Chain Monte Carlo (MCMC) methods by Nijkamp et al. (2019), where short, randomly initialized Markov chains approximate the gradient required for model fitting. It formalizes the theoretical connections between these short-run MCMC methods and kernel-based maximum mean discrepancy (MMD) estimations, specifically highlighting that the kernel used corresponds to the neural tangent kernel. The proposed framework is validated through practical examples, including applications to high-dimensional Gaussian graphical models and diffusion-based image generation, demonstrating improved computational efficiency and practical effectiveness compared to traditional methods.

Degree

thesis:*
Name thesis:degree_name
Doctor of Philosophy
Level thesis:degree_level
doctoral
Discipline thesis:degree_discipline
Mathematics & Statistics
Grantor
Boston University
Year dc:date.issued
2025

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Sun, Yi
Advisor dc:contributor.advisor
  • Atchadé, Yves
ORCID dc:identifier.orcid

Subjects

dc:subject × 7

Rights

dc:rights
Statement dc:rights
  • Attribution 4.0 International
Language dc:language.iso
en_US

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/2144/53347

Chain of custody

source
Harvested from
Boston University
Base URL
open.bu.edu/oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Sun, Yi. Some statistical contributions to physics-informed neural networks and energy-based models. doctoral thesis, Boston University, 2025. https://hdl.handle.net/2144/53347