Unsupervised machine learning for automated crystal orientation mapping on noisy 4D-STEM data

Abstract

Four-dimensional scanning transmission electron microscopy (4D-STEM) is a high-throughput measurement used to comprehensively acquire diffraction patterns across a region of interest in a specimen. While the measurement is useful for identifying crystallographic microstructures, the data analysis requires significant effort owing to the complexity and high dimensionality of the data. Furthermore, the analysis of low signal-to-noise data from beam-sensitive materials such as polymers is particularly challenging. This study proposes an automatic analysis pipeline that leverages denoising and unsupervised crystallographic analysis for 4D-STEM data. The denoising method is based on a self-supervised approach that eliminates the need for clean training images. The developed method is extended to 4D-STEM data by utilizing diffraction patterns observed at neighboring probe positions. The subsequent unsupervised analysis is designed to be invariant to the crystal in-plane rotations, enabling the effective identification of crystal components. The efficacy of the proposed pipeline is evaluated using both synthetic and experimental 4D-STEM data. Quantitative evaluation for synthetic data across various noise levels demonstrates the robustness of our approach. The results from the experimental data show good agreement with expert analysis in the previous study; furthermore, the pipeline identifies minor crystallographic components that were previously overlooked. These results demonstrate that the pipeline enables robust and automated analysis with significantly reduced analysis cost.