EBSD Kikuchi pattern analysis using autoencoder

抄録

Electron backscatter diffraction (EBSD) is a surface analysis method that provides detailed information on the crystal orientation and strain of a solid sample and is applied to the evaluation of the crystal structure of various solid samples. Kikuchi diffraction patterns are generated from each measurement point on the surface of a solid sample depending on the crystal structure and the angle. Main Kikuchi bands in the Kikuchi patterns are generally used for indexing the crystal structures. A pattern matching method based on the simulation of Kikuchi patterns from the type of crystal lattice of the assumed sample is now available from 2022. This method has increased the possibility of indexing complex regions that had been difficult to index in the past. However, it is still difficult to evaluate unknown structures because only crystal structures predicted in advance can be simulated. Therefore, we proposed an analytical method to clarify the crystal structure by multivariate analysis, considering all Kikuchi patterns obtained from all measurement points of the sample as variables [1]. In this study, we aim to extract more detailed crystal structures by applying autoencoder which is an unsupervised learning method based on artificial neural networks. In this study, austenitic stainless steel (SUS304) with dislocations produced by cold working [1] was analyzed using EBSD (Symmetry, Oxford Instruments) with a scanning electron microscope (SU3500, Hitachi High Tech Corp.). The thickness and he grain size of the sample were 100 µm and 50–150 µm, respectively. The accelerating voltage was 20 kV, the step size was 1 µm, and the raster size was 74 × 82 µm². All Kikuchi pattern maps at all measurement points were analyzed at once. The pixels of each Kikuchi pattern map were the variables and the pixels of the measured area were the samples. The original Kikuchi pattern maps were reduced by integrating 8×8 pixels to make the number of the variables smaller than the sample number. The data set was analyzed using sparse autoencoder of Neural Network toolbox of Matlab and our original program coded in Python 3 [2,3]. As a result, the measurement area of the sample was divided into several regions in the features (the middle layer) and a Kikuchi pattern map corresponding to each divided region were extracted in the decoder weights. To obtain clear images for sample regions and Kikuchi pattern maps, large epoch number was required. References 1) S. Aoyagi, D. Hayashi, Y. Murase, N. Miyauchi, A. N. Itakura, e-Journal of Surface Science and Nanotechnology, 21(3) 128-131 (2023). https://doi.org/10.1380/ejssnt.2023-023. 2) S. Aoyagi, K. Matsuda, Rapid Communications in Mass Spectrometry, 37(4), e9445 (2023). https://doi.org/10.1002/rcm.9445 3) S. Aoyagi, D. Hayashi, A. Nagataki, T. Horiba, M. Saito, e-Journal of Surface Science and Nanotechnology, 21(1) 9-16 (2022).