On the use of clustering workflows for automated microstructure segmentation of analytical STEM datasets

This study considers the issue of automated segmentation of scanning transmission electron microscopy (STEM) datasets using unsupervised machine learning approaches. To this end, a systematic comparison of two clustering workflows that had been established in previous literature was performed on two distinct material systems—an experimentally acquired Co₂FeSi Heusler alloy and a simulated Au-matrix and Al₂Cu precipitate. The cluster outputs were evaluated using a variety of unsupervised clustering metrics measuring separation and cohesion. It was found that the cluster output of a variational autoencoder (VAE) performed better compared to a more conventional latent transformation via Uniform Manifold Approximation & Projection (UMAP) on 4D-STEM data alone. However, the UMAP workflow applied to merged 4D-STEM and STEM-energy dispersive x-ray (STEM-EDX) data produced the best cluster output overall, indicating that the correlated information provides beneficial constraints to the latent space. A potential general workflow for analyzing merged datasets to identify structural-composition changes across different material systems is proposed.