2988 Microsc. Microanal. 28 (Suppl 1), 2022
doi:10.1017/S1431927622011199 © Microscopy Society of America 2022
Doing More with Less: Artificial Intelligence Guided Analytics for Electron
Microscopy Applications
Sarah Akers
1
, Marjolein Oostrom
1
, Christina Doty
1
, Matthew Olstza
2
, Derek Hopkins
3
, Kevin Fiedler
4
and Steven R. Spurgeon
2,5*
1.
National Security Directorate, Pacific Northwest National Laboratory, Richland, WA, United States.
2.
Energy and Environment Directorate, Pacific Northwest National Laboratory, Richland, WA, United
States.
3.
Earth and Biological Sciences Directorate, Pacific Northwest National Laboratory, Richland, WA,
United States.
4.
Mathematics and Statistics Department, Washington State University Tri-Cities, Richland, WA,
United States.
5.
Department of Physics, University of Washington, Seattle, WA, United States.
* Corresponding author: steven.spurgeon@pnnl.gov
Scanning transmission electron microscopy (STEM) is an important tool in the study of microstructures
and ultimately a wide range of material and chemical systems. Quantitatively describing these
microstructures with advanced artificial intelligence (AI) techniques has recently become an emerging
and impactful area of research in the materials domain [1]. Sophisticated new models and algorithms are
being developed to automate the task of image characterization, bypassing the need for expensive
manual analysis and enabling the discovery of new and unexpected features within an image. However,
typical AI workflows use large libraries of labeled data to train specialized networks toward specific
tasks, e.g., segmentation of a specific material class. One serious challenge for AI in this area is the lack
of high-quality, well curated training data. Additionally, the need to instantly adapt to changing imaging
conditions, length scales, detectors, aberrations, and materials systems is a top priority for materials
characterization and a challenge for specialized AI.
We explore the topic of few-shot or low-shot learning (FSL) techniques in the context of creating
adaptable AI designed for flexibility in both analytics and data acquisition. FSL, as the name suggests,
uses little to no data in training and nearly eliminates the laborious task of data annotation. We have
observed success with FSL for flexible segmentation of STEM image data, leading to powerful data-
driven opportunities for microstructural characterization for materials discovery and design [2,3,4]. We
will discuss the potential for FSL techniques to draw upon multiple imaging and spectroscopic data
streams simultaneously, in a multimodal framework shown in Figure 1, for improved interpretability and
enhanced latent signature detection [5].
Downloaded from https://academic.oup.com/mam/article/28/S1/2988/6997231 by guest on 25 January 2023