Applications of Machine Learning in Crystallographic Orientation Determination

Being placed at the heart of the materials paradigm, characterization connects up the components surrounding and helps explain how they interact with each other. Through extrinsic excitation, it helps set up a mapping function between the response signal and corresponding material attributes. Among all factors that could affect material’s performance, texture is of high importance as anisotropy is prevalent in materials’ microstructure. This makes the study of grain orientations indispensable. During the past nearly three decades, electron back-scatter diffraction (EBSD) in scanning electron microscope (SEM) has become a mainstream microstructure characterization technique for the study of grain orientations and texture of crystalline metallurgical and geological materials. This thesis work conducts systematic research into the applications of machine learning algorithms in crystallographic orientation determination and simulation. The objective is to overcome the limitations of conventional orientation indexing methods, as well as study the distribution of back-scattered electrons that form EBSD patterns. Specifically, for indexing of EBSD patterns, an end-to-end convolutional neural network, and a hybrid framework composed of convolutional neural network and dictionary indexing module are presented. Furthermore, a generative model is proposed to realize parametric simulation of EBSD patterns. Model behavior and performance are analyzed through multiple metrics and compared with other accepted methods in the field. Besides, through visualization, the model interpretability is investigated. We hope this study will promote the integration of machine learning into materials characterization.