Advancing atomic resolution electron tomography: missing data retrieval and nonlinearity correction

Abstract

Electron tomography is a powerful tool revealing the three-dimensional (3D) structural details of materials beyond 2D imaging. Because the annular dark field scanning transmission electron microscopy (ADF-STEM) produces image contrast proportional to a mass thickness, ADF-STEM images can be directly used for electron tomography. Aided by the development of aberration correction in electron microscopy and reconstruction algorithm, the resolution of the 3D reconstruction reached atomic resolution, which enables us to determine the 3D atomic structure of various nanomaterials. Therefore, ADF-STEM based atomic resolution electron tomography (AET) became a prototype of the technique, providing a new insight to understand the structure-properties of nanomaterials at a most fundamental level.

In this dissertation, we present the AET methods for correctly missing data retrieval and nonlinearity correction, and their applications. First, we introduce the neural network-assisted AET method and provide experimental results of a Pt nanoparticle. In a conventional ADF-STEM based tomography, the full tilt angle measurement is not achievable due to sample geometry and tilt limitation as known missing wedge problem. The missing wedge problem results in undesired artifacts and resolution reduction along the missing wedge direction, which hinders the precise determination of the 3D surface/interface atomic structure. The neural network-assisted AET method can resolve the missing wedge problem by retrieving the missing information. Using Pt nanoparticles as a proof-of-principle, we demonstrate our deep learning-based methods to reliably determine the 3D surface atomic structure of the nanoparticle with precision of 15 pm. Second, we directly apply the deep learning-based AET method to a dumbbell-shaped Pt nanoparticle formed by the coalescence of two single-crystal monomers. Using the neural network-assisted AET, we determined the full 3D atomic structure of a Pt nanodumbbell at the 9.9 pm precision, where we found the double twin boundary at the interface. The formation of the double twin boundary was governed by the diffusion of interfacial atoms. At the interface, substantial anisotropy and disorder were also observed. It suggests that the coalescence process includes not only a simple atomic diffusion process but also complex dynamic processes like plastic deformation or structure restructuring. Also, precise determination of the 3D surface structure can yield the full 3D strain tensor, which allows direct calculation of the oxygen reduction reaction activity at the surface.

However, conventional ADF-STEM based tomography inherently suffers from several issues including high electron dose requirement, poor contrast for light elements, and artifacts from image nonlinearity. We also propose a tomography method called MultiSlice electron tomography (MSET) based on a tilt series of 4D-STEM dataset. Using metallic nanoparticles as model systems, our simulation shows that the MSET algorithm can not only significantly reduce the nonlinear artifacts from the multiple scattering but also improve sensitivity for low-Z element detection under low-dose experimental conditions.

Further, we expect that the deep learning-based AET method and the MSET method in the near future would be combined, and the combined method can be applied to beam-sensitive nanomaterials whose 3D surface/interface atomic structures have never been fully revealed.