Abstract

Grain Boundary Networks (GBNs) are a high-dimensional feature in polycrystalline microstructural materials that are formed from the crystallographic degrees of freedom. This feature affects multiple material properties such as diffusivity, fracture, elasticity, heat transfer, and others. The link between the crystallography and the material properties is called the structure-property linkage, and allows for simulation and design of materials. While it is possible to use the structure-property linkage to design materials for better performance, the high-dimensional nature of the interconnected grain boundaries preclude the use of many common optimization algorithms for efficiency or local minima considerations. This work shows how the high-dimensional GBN can be understood and optimized using a combination of Human Computation and Harmonic Expansions. The GBN design problem is formulated as a "Game with a Purpose," "Operation: Forge the Deep." This was used to show how and when human players are more efficient and effective at solving the problem than a common global optimization algorithm, through a statistics driven experiment. To understand the differences in resulting GBN solutions, a Harmonic Expansion representation was created to quantitatively differentiate GBN configurations from each other. This representation shows new insights into the similarities and differences created directly from the long range connected nature of GBNs, and describes differences in material properties not seen before with local GBN measures such as Triple Junction Fractions. Finally, the player decisions collected from "Operation: Forge the Deep" were used to generate a GBN Decision Transformer; a machine learning model that predicts optimization decisions like players. This model was able to generate good optimization solutions for GBNs when sufficient training data was available, but was unable to extrapolate to larger structures. However, this model was also capable of inference on multiple different structure-property models, even when only trained on a single model type. This allows for training on one physics model, but use on many different material property predictions. In total, this work helped expand optimization and characterization of long-range GBN structures for use in structure-property linkages.

Degree

PhD

College and Department

Ira A. Fulton College of Engineering; Mechanical Engineering

Rights

https://lib.byu.edu/about/copyright/

Date Submitted

2025-04-15

Document Type

Dissertation

Handle

http://hdl.lib.byu.edu/1877/etd13570

Keywords

microstructures, grain boundary networks, games with a purpose, machine learning

Language

english

Download

Included in

Engineering Commons

Share

COinS