Deep-Learning-Assisted Rapid STEM and Its Application to Tomography Observation

Abstract

Scanning transmission electron microscopy (STEM) is more suitable for visualizing the internal structure of thick samples compared to conventional transmission electron microscopy whose resolution is limited by the chromatic aberration of the imaging lens system, and it is often used for three-dimensional structural analysis using electron tomography. However, STEM image quality is seriously degraded by noise and artifacts, especially when pursuing rapid imaging on the order of milliseconds per frame or faster. In this paper, we report that deep learning-based denoising is effective for rapid STEM imaging, and can be applicable to rapid STEM tomography. By acquiring tilt-series images in just 5 seconds, the three-dimensional dislocation arrangement in a thick (300 nm) steel sample can be determined with sufficient accuracy. This method has enormous potential on improving in-situ or operando observation of samples in relatively thick media including liquid cells.