Supervised and Unsupervised approaches for Electron Microscopy Data Analysis

Mary Scott
University of California, Berkeley (UC Berkeley)

Recently, materials science has undergone a data science revolution. With the increasing application of advanced computational methods for analysis of experimental data streams, the development of advanced algorithms for data distillation is an important theme in modern materials science research. Electron microscopy is the characterization method of choice to observe the atomic-scale and microstructural local features within materials that play a critical role in material performance. With resolution that can be deeply sub-Angstrom, a single image from a high-resolution electron microscope can measure atomic positions, defects, and strain. Furthermore, advances in high frame rate electron detection generate datasets consisting of millions of diffraction patterns- an approach that enables multimodal analysis from the same dataset to create maps of crystal orientation, strain, and more.

The increasing ability to perform high throughput electron microscopy has created opportunity for large scale nanomaterial studies alongside a need for robust, automated analysis. Advances in machine learning and computer vision have made high accuracy automated image interpretation possible. While widely applied to natural images, this approach is only recently being applied to atomic resolution electron microscopy images. Therefore, it is desirable to establish how to best implement machine learning approaches for scientific imaging data analysis. When combined with existing automatic image acquisition protocols, machine learning is now a viable option to close the materials design loop and incorporate electron microscopy into high-throughput materials design and synthesis.

Back to Workshop IV: Multi-Modal Imaging with Deep Learning and Modeling