Emerging Machine Learning-Based Data Analysis Techniques and Algorithms for Exploiting 4D-STEM Datasets
Hsu-Chih Ni, Renliang Yuan, Jiong Zhang, Jain-Min Zuo · Microscopy and Microanalysis · 2024
Four-dimensional scanning transmission electron microscopy (4D-STEM) is a powerful tool for nano- and atomic-scale materials characterization [1]. Electron scattering signals are collected at every probe position, and extracting material properties from diffraction signals plays a critical role in the success of 4D-STEM as they may not manifest directly. This necessitates the development of dedicated data processing techniques and new algorithms for effective information extraction. Moreover, achieving large field-of-view and high-throughput materials analyses demands efficient data collection and management strategies. In this talk, we introduce a suite of machine learning-based data analysis techniques tailored for exploiting 4D-STEM datasets as outlined in Figure 1. Analysis of nanobeam electron diffraction (NBD) patterns. NBD has long served as a method for measuring lattice strain at nm scale [2]. Here, we introduce the implementation of a convolutional neural network-based approach for strain analysis, enabling high-throughput and high-resolution strain mapping. Furthermore, by leveraging the intensity of diffraction disks in NBD patterns, precise orientation mapping within a single crystal domain can be achieved using artificial neural networks. The synergy of these techniques facilitates a comprehensive understanding of local structural heterogeneity in materials. A nanowire MOSFET device will be used as a case study to demonstrate the effectiveness of these techniques. Diffraction pattern clustering and compressive sensing [3]. By classifying diffraction patterns with dimension reduction algorithms, such as principal component analysis, local microstructural and grain boundary information can be extracted. The concept of clustering can be further extended to compressive sensing [4] with leveraging the sparsity assumption of 4D-STEM datasets, thereby enabling high efficiency and low-dose data collection. An iterative dictionary learning algorithm is developed for 4D-STEM compressive sensing [5]. The algorithm is demonstrated with both synthetic datasets and an experimental dataset which was acquired with random scan 4D-STEM. Additionally, compressive sensing naturally facilitates data compression, as demonstrated by our ability to compress 4D-STEM datasets by a factor of 100 with high fidelity. This technique not only enables data collections on beam-sensitive materials but also reduces the data storage and transmission burden. Beyond NBD patterns. we explore clustering of ronchigrams to reveal atomic-scale structural information. We propose a novel strategy for averaging Ronchigrams across multiple unit cells to elucidate atomic features. By mapping out the similarity between individual and the average ronchigrams, local defects and fluctuations can be revealed. The potential of performing in-line holography [5] with this technique will also be discussed and simulated [7]. Data analysis techniques and data management strategies for 4D-STEM datasets with different sample type and illumination conditions.