High-Throughput In Situ Electron Microscopy with Self-Supervised Machine-Learning Denoising Framework

In situ transmission electron microscopy (TEM), including liquid cell TEM, cryo-TEM, time-series scanning TEM (STEM), and more, gains attention as one of the main analytical methods in materials science because it allows real-time observations of materials at sub-nanometer scale resolution. The in situ EM offers an opportunity to directly observe diverse classes of chemical reactions of materials nucleation, crystallization, to diverse types of phase transformations. There are also many examples where in situ TEM directly reveals important reactions in energy conversion and storage systems. Special efforts have been made to liquid cell EM as a reliable method for monitoring electrochemical process of materials. With this approach, direct TEM observation successfully reveals how materials respond to incoming molecules and electrochemical/potential cycles. Recently, quantification of the e-beam radiolysis is increasingly important for reliable in situ TEM observations because systems of interest are readily susceptible for radiolysis degradation during imaging, reducing the reliability of the obtained data. In addition, many of samples are beam-sensitive. To overcome this limitation, one of the realistic solutions in experiment is to minimize the electron dose rate to suppress unwanted radiolytic reactions. However, minimizing the dose rate inevitably compromises achievable contrast resolution. We introduce a new methodology for quantitative in situ TEM using machine learning-based data denoising. In this approach, a novel self-supervised machine learning framework can handle different imaging conditions in diverse in situ TEM. This approach does not require ground truth datasets, allowing us to obtain high-quality images with only experimental datasets. Using this self-supervised machine learning framework, we were able to successfully denoise liquid cell EM observations, based on silicon nitride liquid cell TEM and high-resolution graphene liquid cell TEM, time-series STEM, and cryo-TEM.