Materials Data Science and Informatics (IAS-9)

This research focuses on a different approach to the determination of crystal orientations from Kikuchi patterns, a key component of orientation microscopy methods such as Electron Backscatter Diffraction (EBSD). Traditional Hough transform based indexing, relies on classical image analysis, which may lead to ambiguities at more challenging patterns, e.g. in deformed materials. Relatively recent developments try to use the full information content of Kikuchi patterns by matching acquired patterns with a library of forward simulated patterns or a spherical master pattern. In this trend, this work leverages machine learning (ML) as an alternative method for orientation determination.

A large dataset of ground truth patterns was derived from a forward simulated master pattern that has been computed by the EMSoft package for different electron energies. This dataset closely resembles experimental patterns at a given geometry and machine settings. A modified Xception neural network architecture was trained on these patterns, with the disorientation angle serving as the loss function. The trained model delivers orientation maps in a quaternion parametrization, with reasonably fast and accurate indexing capabilities. The first results demonstrate that the method is able to handle cubic crystal symmetries with a dense sampling in orientation space. This approach serves as a step up to more challenging cases.