AI/ML Analytics Suite for the Automated Interpretation of Scanning Transmission Electron Microscopy Data

With the rise of automated experimentation, the topic of few-shot and semi-supervised learning techniques are increasingly relevant in the context of creating adaptable AI designed for flexibility in both analytics and data acquisition. We have developed a platform with a custom central controller that automatically utilizes these techniques for both quantitative characterization and adaptive sampling in scanning transmission electron microscopes (STEM). We previously developed a flexible, semi-supervised few-shot machine learning approach for segmentation of STEM images that uses a small support set of canonical examples across a material system to detect and quantify discrete regions. We expand our previously developed few-shot approach to include multiple modalities by incorporating energy dispersive X-ray spectroscopy (EDS) data, which allows both lattice structure and elemental composition to be taken into account during segmentation. We also demonstrate methods for unsupervised clustering using similarity graphs and semi-supervising clustering to automate support set generation. High-throughput characterization though rapid and accurate image classification and unsupervised microstructural feature mapping take us one step closer to truly autonomous microscope platforms.