From the highest-resolution electron ptychography to micrometer-scale strain mapping, four-dimensional scanning transmission electron microscopy (4D-STEM) has demonstrated significant advancements in the study of materials. However, the processing of large 4D data still remains challenging and heavily relies on experts. In this talk, I will present the utilization of machine learning methods to assist 4D-STEM data processing. For mapping deformations in materials, we utilized a hierarchical unsupervised learning workflow to cluster nanobeam electron diffraction patterns in 4D datasets. This approach provides an initial analysis of the 4D data and uncover essential features in the sample even without a prior knowledge. For example, using this method, we have uncovered strain and ripples in two-dimensional (2D) lateral heterojunctions, which has been previously reported. Applying this approach on a novel 2D ferroelectric SnSe sample, we have identified ferroelectric domains and super domains. Based on the results, a subsequent quantitative analysis of the same 4D data provides details in lattice structure at different domains and domain boundaries. In addition, we also developed a streamlined data-processing workflow for electron ptychography. Electron ptychography is capable of deep sub-angstrom resolution reconstructions but requires careful selection of multiple reconstruction parameters. We demonstrate a scheme for automatic parameter tuning by implementing Bayesian optimization with Gaussian processes. This approach is able to find high-resolution reconstructions after exploring only 1% of the entire parameter space. With minimal prior knowledge, the approach significantly improves the efficiency of 4D data processing and provides high quality ptychography reconstructions that agree with the experts.
Copyright. All Rights Reserved.