Spherical-angular dark field imaging and sensitive microstructural phase clustering with unsupervised machine learning.

Electron backscatter diffraction is a widely used technique for nano- to micro-scale analysis of crystal structure and orientation. Backscatter patterns produced by an alloy solid solution matrix and its ordered superlattice exhibit only extremely subtle differences, due to the inelastic scattering that precedes coherent diffraction. We show that unsupervised machine learning (with principal component analysis, non-negative matrix factorisation, and an autoencoder neural network) is well suited to fine feature extraction and superlattice/matrix classification. Remapping cluster average patterns onto the diffraction sphere lets us compare Kikuchi band profiles to dynamical simulations, confirm the superlattice stoichiometry, and facilitate virtual imaging with a spherical solid angle aperture. This pipeline now enables unparalleled mapping of exquisite crystallographic detail from a wide range of materials within the scanning electron microscope.

Advancing characterisation with statistics from correlative electron diffraction and X-ray spectroscopy, in the scanning electron microscope.

The routine and unique determination of minor phases in microstructures is critical to materials science. In metallurgy alone, applications include alloy and process development and the understanding of degradation in service. We develop a correlative method, exploring superalloy microstructures, which are examined in the scanning electron microscope (SEM) using simultaneous energy dispersive X-ray spectroscopy (EDS) and electron backscatter diffraction (EBSD). This is performed at an appropriate length scale for characterisation of carbide phases' shape, size, location, and distribution. EDS and EBSD data are generated using two different physical processes, but each provide a signature of the material interacting with the incoming electron beam. Recent advances in post-processing, driven by 'big data' approaches, include use of principal component analysis (PCA). Components are subsequently characterised to assign labels to a mapped region. To provide physically meaningful signals, the principal components may be rotated to control the distribution of variance. In this work, we develop this method further through a weighted PCA approach. We use the EDS and EBSD signals concurrently, thereby labelling each region using both EDS (chemistry) and EBSD (crystal structure) information. This provides a new method of amplifying signal-to-noise for very small phases in mapped regions, especially where the EDS or EBSD signal is not unique enough alone for classification.