Indexing High-Noise Electron Backscatter Diffraction Patterns Using Convolutional Neural Network and Transfer Learning

Abstract

Convolutional neural network (CNN)-based indexing of electron backscatter diffraction (EBSD) patterns has recently been proposed as an alternative to commercially available indexing methods, due to the issues of the traditional Hough transform-based indexing method in indexing experimental high-noise EBSD patterns and the high time consumption of the dictionary indexing (DI) method in indexing EBSD patterns of low-symmetry crystal system. However, the robustness of the network to noise, as well as its generalization ability, has remained a major challenge. In this work, we proposed an efficient and general CNN for the rapid and autonomous indexing of crystal orientations in experimental high-noise EBSD patterns of various crystal structures by transfer learning. The robustness of the network to noise was analyzed by collecting multiple sets of Kikuchi patterns of Ni metal with different strains, and the generalization ability of the network was evaluated by indexing other materials (Mg, Fe, Al, ferritic austenitic dual-phase steel) with different crystal structures. We also conducted a sensitivity analysis of the network to explore the stability of the model and visualized some feature maps of the model to reveal the key features that determine the crystal orientation. Ultimately, the model kept a mean disorientation error in 0.18° when indexing 500 randomly selected simulated EBSD patterns of Ni and a mean disorientation error in 1.35° when indexing experimental EBSD patterns of other materials in the presence of 0~40% strain. The results show that our method can index experimental high-noise EBSD patterns and can be applied to index the crystal orientation of other crystal structures (BCC, FCC, BCC+FCC，HCP) without changing the simulated dataset, where the crystal structure of Ni in the simulated dataset is FCC.

Keywords: Electron backscatter diffraction, crystal orientation indexing, convolutional neural network, transfer learning