Hybrid learning models for statistical continuum mechanics

Author(s)
Kelly, Conlain
Editor(s)
Associated Organization(s)
Organizational Unit
Organizational Unit
School of Computational Science and Engineering
School established in May 2010
Supplementary to:
Abstract
This document summarizes my research aimed at creating interpretable and useful machine learning models to assist efforts in materials design. In particular, I focus on the problem of predict- ing local response fields (stresses and strains) over heterogeneous structures subjected to boundary loading conditions, also known as the localization problem. In a sense all of micromechanics con- sists of elasticity plus defect motion; my doctoral work extensively explores the first half of that equation in relation to machine learning. This thesis comprises three papers: an exploratory study applying iterative neural networks to elastic localization, a thermodynamically-informed iterative neural operator which generalizes these ideas to work over a wider range of microstructure classes and loading directions, and an extrapolation study which explores how far neural operators can be taken outside their training distribution by hybridizing them with FFT-based relaxation solvers. All three papers focus on purely elastic deformations, but keep an eye on the long-term goal of mod- eling time-dependent, dissipative deformations. These are contextualized as part of the general stochastic inverse problem known as process-structure-property modeling. Beyond the localization problem, I have been fortunate to collaborate on a number of works which build up different parts of the process-structure-property linkage. Most of these efforts have been published in the theses of Dr. Andreas Robertson and Dr. Adam Generale, so I only pro- vide brief descriptions and summaries for each paper. In particular, I contextualize these works as part of the increasing alignment between the fields of deep learning, numerical methods, and continuum mechanics. The contributions of this thesis are thus twofold: to provide useful deep learning models which allow exploration of the microstructure space, and to advance a shared lan- guage bridging the conceptually-isolated fields of data-driven modeling and statistical continuum mechanics.
Sponsor
Date
2024-12-08
Extent
Resource Type
Text
Resource Subtype
Dissertation
Rights Statement
Rights URI