Combining machine learning algorithms with traditional methods for resolving the atomic-scale dynamic structure of monolayer MoS2 in high-resolution transmission electron microscopy

Abstract High-resolution transmission electron microscopy (HRTEM) promises rapid atomic-scale dynamic structure imaging. Yet, the precision limitations of aberration parameters and the challenge of eliminating aberrations in Cs -corrected transmission electron microscopy constrain resolution. A machine learning algorithm is developed to determine the aberration parameters with higher precision from small, lattice-periodic crystal images. The proposed algorithm is then validated with simulated HRTEM images of graphene and applied to the experimental images of a molybdenum disulfide (MoS 2 ) monolayer with 25 variables (14 aberrations) resolved in wide ranges. Using these measured parameters, the phases of the exit-wave functions are reconstructed for each image in a focal series of MoS 2 monolayers. The images were acquired due to the unexpected movement of the specimen holder. Four-dimensional data extraction reveals time-varying atomic structures and ripple. In particular, the atomic evolution of the sulfur-vacancy point and line defects, as well as the edge structure near the amorphous, is visualized as the resolution has been improved from about 1.75 Å to 0.9 Å. This method can help salvage important transmission electron microscope images and is beneficial for the images obtained from electron microscopes with average stability.