Machine Learning-Driven 3D Sectioning and Analysis in Electron Microscopy

Transmission electron microscopy (TEM) is pivotal in determining atomic-scale structures in materials science. Two primary methods for 3D sectioning in electron microscopy are electron tomography and multi-slice ptychography. While electron tomography is powerful, it often falls short with beam-sensitive nanomaterials due to the long acquisition time required for numerous tilt series images. Multi-slice ptychography, on the other hand, uses iterative algorithms to find probe-specimen interactions in samples, offering improved resolution but poor depth accuracy. We introduce a machine learning-driven approach to electron tomography without acquisition of tilt series images, specifically tailored for twisted bilayer transition metal dichalcogenides (TMDCs). This technique reconstructs high-resolution 3D images from defocused diffraction patterns obtained via scanning transmission electron microscopy, similar to multi-slice ptychography. By integrating machine learning, our method surpasses traditional multi-slice ptychography and electron tomography, enhancing both in-plane resolution and depth accuracy. This advancement in atomic resolution tomography significantly improves the structural determination of a wide range of beam-sensitive nanomaterials.