Autonomous Acquisition in Scanning Transmission Electron Microscopy for Dynamic Sampling of Material Heterogeneities

Abstract

Scanning transmission electron microscopy (STEM) spectrum imaging techniques, including 4D-STEM and electron energy loss spectroscopy, are widely applied to provide complex information about the local structures and properties of materials. However, the long acquisition time and size of the diffractive and spectral datasets involved in these techniques create significant challenges in preventing electron beam-induced sample changes, specimen drift, and redundant data acquisition. In this work, we present an autonomous data acquisition methodology that allows the sampling rate to be dynamically adjusted through on-the-fly assessments of incoming data. Based on programmable level of interest metrics, the method iteratively collects diffractive and spectral data while varying the local sampling rate to target the highest interest areas and avoid redundant data collection. A hardware platform built around a customized field-programmable gate array is used to implement the dynamic sampling strategy through high-speed synchronization between the scan coils, pixelated electron detector, and high-speed beam blanker. Machine learning methods are deployed to perform unsupervised clustering of acquired spectral data and inpaint the data at unsampled locations. Our development constitutes a flexible STEM platform that can be tailored for a wide range of experimental goals and enables efficient exploration and feature detection in heterogeneous materials.