Georgia Institute of Technology
Neural Inverse Microstructure Design with Bayesian Scale-Bridging
Abstract
dc:description.abstractThe discovery of novel materials is core to the advancement of countless scientific disciplines, providing a design base which has historically accompanied technological step-changes in capability. At its core, materials exploration hinges upon addressing an array of inverse problems, attempting to uncover latent information regarding the material, its response to external stimuli, and leveraging this information to design materials meeting certain functional target properties. The complex hierarchical materials interactions driving these observable properties are derived from inherently stochastic internal constituent spatial arrangements - the materials' microstructure, and localized behavior at multiple length-scales. These complex mechanisms present several formidable challenges preventing an ability to invert the relationship between a materials’ microstructure, and its resulting properties. A core conceptualization in this work is that materials are the result of stochastic processes, driving uncertainty in their hierarchical response. This theoretical understanding of materials shifts focus from materials design as an optimization problem, to one involving the manipulation of probability densities, admitting the powerful constructs underpinning Bayesian inference, statistical mechanics, and optimal transport. The work presented explores this shifted paradigm, addressing several relevant stochastic inverse problems enabling materials exploration. Specifically, (i) the ability to infer a distribution of constitutive model parameters in the presence of sparse heterogeneous experimental data, (ii) uncovering a probabilistic understanding of lower length-scale constituent response with experimental samples of the microstructure and effective macroscale behavior, and propagating this uncertainty in exploring the microstructure space, and (iii), developing the machinery for enabling the stochastic exploration of the microstructure space, and corresponding space of processing histories for a given target property set. This thesis broadly focuses on the overarching topics of Bayesian scale-bridging and inverse stochastic microstructure design, connecting them in a unified framework for accelerating the discovery of novel microstructures, and presenting the necessary computational methods for its implementation. In taking this probabilistic approach, samples of varied stochastic processes are directly admissible in uncovering latent microstructure information. This information can in turn, be used to augment sparse experimental datasets, providing an anchor for conditional flow-based generative models to perform the larger inversion. This neural parameterization of continuous families of transport maps, lays the foundation for the stochastic exploration of microstructure and processing pathways, providing the designer with the exciting ability to both identify and realize novel structures.
Degree
thesis:*- Level thesis:degree_level
- Doctoral
- Department dc:contributor.department
- Mechanical Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2024
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Generale, Adam P.
- Advisor dc:contributor.advisor
-
- Kalidindi, Surya R.
- Committee members dc:contributor.committeemember
-
- McDowell, David L.
- Joseph, Roshan V.
- Fung, Victor
- Stebner, Aaron P.
- Kumar, Rajesh
Subjects
dc:subject × 10Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/75187