Analysis of Interpretable Data Representations for 4D-STEM Using Unsupervised Learning

Abstract

Understanding the structure of materials is crucial for engineering devices and materials with enhanced performance. Four-dimensional scanning transmission electron microscopy (4D-STEM) is capable of mapping nanometer-scale local crystallographic structure over micron-scale field of views. However, 4D-STEM datasets can contain tens of thousands of images from a wide variety of material structures, making it difficult to automate detection and classification of structures. Traditional automated analysis pipelines for 4D-STEM focus on supervised approaches, which require prior knowledge of the material structure and cannot describe anomalous or deviant structures. In this article, a pipeline for engineering 4D-STEM feature representations for unsupervised clustering using non-negative matrix factorization (NMF) is introduced. Each feature is evaluated using NMF and results are presented for both simulated and experimental data. It is shown that some data representations more reliably identify overlapping grains. Additionally, real space refinement is applied to identify spatially distinct sample regions, allowing for size and shape analysis to be performed. This work lays the foundation for improved analysis of nanoscale structural features in materials that deviate from expected crystallographic arrangement using 4D-STEM.

Imaging the Surface/Interface Morphologies Evolution of Silicon Anodes Using in Situ/Operando Electron Microscopy

“…Numerous factors, including material design and synthesis, enhanced ex situ and in situ characterization, simulation, and actual performance testing as feedback for greater understanding, influence the development of silicon-based battery materials (Figure a,b). Additionally, the combination of big data analysis with machine learning is a potent one that has shown to be useful in the field of materials science (Figure c). − …”

Versatile domain mapping of scanning electron nanobeam diffraction datasets utilising variational autoencoders

“…An alternative contemporary approach is to use the workflow outlined in Bruefach et al 21 . This approach performs three methods of feature extraction on the dataset to represent the diffraction information and then, for each, an iterative NMF decomposition 41 to produce domain maps.…”

“…More information on the specifics are available in the manuscript 21 . As before, the workflow is applied to both the raw thin foil Al alloy and the precipitate-masked thin foil Al alloy datasets.…”

“…In this section, the outcome of the workflow presented in this paper is compared to some other more established approaches for processing 4D-STEM datasets. The techniques in question are a direct NMF of the sample data (without the use of a VAE or the hull clustering technique) and a feature extraction-based approach 21 utilising Py4DSTEM 40 for much of the data processing. For the sake of simplicity, the techniques are compared using the same single dataset, the precipitate on thin Al foil, but the technique is applied to both the raw dataset and also the precipitate-masked dataset to see how this pre-processing step affects performance.…”

“…This work reports a structurally agnostic, data-driven workflow for domain mapping, which requires no prior knowledge of possible crystal structures and is equally as effective for identifying amorphous regions with little to no crystalline diffraction character. Techniques based on feature extraction with non-negative matrix factorisation (NMF) 21 and hierarchical k-means clustering to group similar diffraction signal 22 have recently been reported. The workflow, discussed within, leverages a variational autoencoder (VAE) to identify the variance and similarity within the diffraction data.…”