Deep learning pipeline for statistical quantification of amorphous two-dimensional materials and dynamic tracking

Abstract text (incl. figure legends and references)

Recent years have witnessed the rise of amorphous 2D (a-2D) materials. The combination of atomic thinness and lack of long-range order has endowed a-2D materials with unique characteristics, such as ultra-high uniformity, excellent mechanical and chemical stability, and abundance in catalytic-active sites. Owing to the atomic thinness of a-2D materials, direct observation of atomic arrangements can be well achieved using aberration-corrected high-resolution transmission electron microscopy (HRTEM) (1). However, extracting structural information from HRTEM images of a-2D materials is a nontrivial work because every individual atom in the HRTEM image (Fig. 1a) needs to be pinpointed. Moreover, research on reaction dynamics, such as amorphous-crystalline phase transition within an a-2D material, requires analysis of every frame from the time-resolved image series, leading to an exponential increase in the workload.

In this work, we developed a U-net-based neural network (NN) for comprehensive image analysis in a-2D materials. HRTEM has been conducted on the Cc/Cs-corrected microscope operated at 80 kV, resolving single atoms with strong contrast (2). The NN is capable of atom coordinate identification and segmentation of evaluation-suitable regions with high precision. This has been achieved by training the NN exclusively on simulated images. A semi-random atomic model with porous and bilayer areas led to the successful recognition of complex attributes in experimental images, including disorder, micropore, and surface contaminations. The NN"s robustness against intensity inhomogeneity was further enhanced by introducing intensity gradients in the training datasets. Implementation on the experimental images of monolayer a-carbon and a-polymer provided multifold statistical datasets and direct visualization of the short-range-ordered structures. Although designed for a-materials, our deep learning pipeline is readily generalizable to crystalline samples, enabling the identification of local defects and grain boundaries. We envisage that automated quantification of short-range order will bring new insights into the structural understanding of a-2D systems, laying the foundation for the establishment of structure-property correlation in this rising class of materials. The deep learning approach may also pave the way for probing dynamic processes with efficiency and precision, e.g., amorphous-crystalline transition, amorphization, nucleation and crystal growth.

Acknowledgement

This research is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – 424798828; 492191310 and the European Union's Horizon2020 research and innovation program (No. 881603, GrapheneCore3). We thank Dr. Xue Liu for providing the a-2D carbon monolayer samples.

Reference

P. Huang et al. Science, 342, 224, 2013 M. Linck et al. Phys. Rev. Lett. 117, 76101, 2016

Fig. 1. Challenges of automated image analysis. a, HRTEM image of monolayer a-carbon. b-d, Cropped out images from the boxed regions in (a), showing areas of disorder (b), micropore (c), and surface contamination (d).

Fig. 2. Automated image evaluation via NN. a, 80 kV HRTEM image of monolayer a-carbon. b, Segmentation applied to (a) for removal of non-monolayer regions (black) and micropores (purple). c, Enlarged image from the boxed region in (a). d, NN mapping of polygons. e-g, Statistical histogram of bond angles, lengths, and polygon frequency. Scale bar: 2 nm (a); 1 nm (c)