Back to search

Georgia Institute of Technology

Neural networks for solving hyperbolic conservation laws

Abstract

dc:description.abstract

Conservation laws in the form of partial differential equations (PDEs) are fundamental to our understanding of physical phenomena in both nature and engineering. When applied to problems involving wave propagation, such as gas dynamics, these equations are inherently hyperbolic. Their corresponding weak solutions often contain discontinuities due to the intrinsic nonlinearity of the equations. To address these challenges, high-resolution schemes—such as TVD, MUSCL, ENO, and WENO—have been developed. These methods employ adaptive stencils and nonlinear limiters to achieve high accuracy without compromising stability. Despite significant advancements, high-fidelity simulations remain computationally expensive and time-consuming, limiting their application to complex, high-dimensional problems. Recently, machine learning and deep learning techniques have emerged as promising alternatives by constructing surrogate models to overcome these limitations. However, existing surrogate models—including data-fitting approaches, hierarchical models, and reduced-order models (ROMs)—often require large datasets and substantial computational resources for training. Moreover, they continue to struggle with accurately resolving flow fields that involve abrupt changes or discontinuities. This PhD thesis aims to develop a novel neural network-based ROM, known as Neural Networks with Local Converging Inputs (NNLCI), to address these challenges. NNLCI is a groundbreaking technique that significantly reduces both the complexity and training cost of solving Euler’s equations. By leveraging the local domain of dependence, NNLCI extracts critical features from two converging low-fidelity local patches to predict high-fidelity solutions at desired spatiotemporal locations. This localized learning approach offers several key advantages: a compact and simple neural network architecture, minimal training data requirements, and accurate predictions for both smooth and discontinuous solutions. Notably, NNLCI can be trained on one set of computational domains and deployed on different domains, making it highly suitable for design optimization. NNLCI’s efficiency and robustness are demonstrated by solving one-dimensional and two-dimensional Riemann problems. Furthermore, when combined with a novel interpolation method, NNLCI will be demonstrated to effectively solve supersonic CFD problems on hybrid unstructured meshes. This advancement in computational science and engineering highlights NNLCI’s potential to tackle increasingly complex flow scenarios and improve the efficiency of engineering design optimization.

Degree

thesis:*
Level thesis:degree_level
Doctoral
Department dc:contributor.department
Mechanical Engineering
Grantor dc:publisher
Georgia Institute of Technology
Year dc:date.issued
2025

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Huang, Haoxiang
Advisors dc:contributor.advisor
  • Lieuwen, Timothy C.
  • Yang, Vigor
  • Liu, Yingjie
Committee members dc:contributor.committeemember
  • Oefelein, Joseph
  • Chen, Yongxin

Subjects

dc:subject × 6

Rights

Language dc:language.iso
en_US

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/1853/82060

Chain of custody

source
Harvested from
Georgia Tech
Base URL
repository.gatech.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Huang, Haoxiang. Neural networks for solving hyperbolic conservation laws. Doctoral thesis, Georgia Institute of Technology, 2025. https://hdl.handle.net/1853/82060