SurfaceScience21 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

O: Fachverband Oberflächenphysik

O 71: Mini-Symposium: Machine learning applications in surface science I

O 71.3: Talk

Wednesday, March 3, 2021, 14:30–14:45, R1

Automatic image evaluation of aberration-corrected HRTEM images of 2D materials. — •Christopher Leist, Haoyuan Qi, and Ute Kaiser — Central Facility for Electron Microscopy,of Electron Microscopy Group of Materials Science, Ulm University, 89081 Ulm, Germany

Aberration-corrected high-resolution transmission electron microscopy (HRTEM) allows for unambiguous elucidation of atomic structures down to sub-Angstrom scale. By determining the positions of each single atom, the distribution and local variation of bond lengths and angles can be evaluated statistically. However, conventional image analysis methods, e.g., handcrafted filter kernels, often requires heavy user supervision and tremendous time cost, posing strong limitations on the data volume for statistical analysis. The incompetence in handling big data volume also incurs the risk of user-induced selection bias, leading to overestimation of low-probability phenomena. Here, we developed a neural network of U-net architecture for automatic analysis of atomic positions in HRTEM images. A combination of networks can be applied to automatically evaluate image series, including automatic exclusion of image regions unusable for evaluation. This method results in large statistics thus reducing the impact of individual errors. The networks are trained with simulated data which reduces user bias and gives a time inexpensive way of generating the required training data. Its implementation on various 2D carbon materials is compared to one another. The distribution of bond angles in CVD graphene, determined by this method, shows excellent agreement with literature.