A Machine-Learning Approach to Characterization of Amorphous Materials with EELS-SI and 4D-STEM

Recently, amorphous materials are used more and more in many industries, which are associated especially with various semiconductor devices, e.g., phase change memory, light-emitting diodes, etc. Accordingly, the importance of the property characterization of amorphous materials has become greater, but it is challenging to study atomic and electronic structures of amorphous materials for a very small area. Furthermore, in-situ experiments are essential to investigate phase changes of amorphous materials in actual devices because their physical properties are sensitively influenced by their atomic structures. Transmission electron microscopes (TEM) satisfy those demands, i.e., TEMs are appropriate for obtaining information on both atomic and electronic structures of materials with electron diffraction and electron energy loss spectroscopy (EELS) respectively. In addition, by combining those techniques with scanning TEM (STEM), i.e., four-dimensional STEM (4D-STEM) and STEM-EELS, we can investigate the physical properties of materials for each scanning position (~ 1 nm2) with a convergent electron beam. However, since the amount of information is tremendous in EELS-SI or 4D-STEM datasets, it is still difficult to retrieve meaningful information on amorphous materials from those datasets, which may contain severe noise caused by low-dose in-situ TEM experiments.<br/>In this research, a machine-learning approach is proposed to identify where distinct local structures are in amorphous materials and what they represent. Linear and nonlinear dimensionality reduction (DR) methods are used to reduce the dimensions of EELS or diffraction patterns (DP) into a much lower-dimensional and information-condensed space. In this space, distinct features implied in the dataset should appear as data clusters. Then, by applying a density-based clustering algorithm to the dimensionality-reduced dataset, we can achieve reasonable separation of the data clusters. In addition, the representative EELS or DP can be obtained by averaging all of the data in each cluster. Also, we can observe the spatial distribution of each cluster in the scanning area. As a result, we can identify locally-distinct atomic and electronic structures of amorphous materials simultaneously for the same scanning area. Hence, we can correlate the atomic and electronic structures with their physical properties easier than before. This approach describes a truly data-driven analysis because it does not require any physical model or background; thus, it can be applied to any hyperspectral dataset.