Crystal orientation and deformation state analysis from Kikuchi patterns via pattern reconstruction aided deep Siamese network

Abstract

Fast and accurate analysis of orientations and deformation states from Kikuchi patterns is vital to understand the crystalline materials’ micro- and macro-mechanical behavior. Here, a pattern reconstruction-aided Siamese network is developed to extract the crystal orientations and deformation states from Kikuchi patterns. We demonstrate the pattern reconstruction technique is essential for accurate predictions at large-area EBSD scans since the remapping operation enables the Siamese network to consider changes in experimental setup parameters dynamically. Moreover, the reconstruction technique serves as a data augmentation method that expands 342,225 raw patterns to 5,475,600 reconstructed patterns. The resulting Siamese network achieves unprecedented accuracy and robustness over a wide range of experimental setup parameters including pattern center, camera elevation and exposure time. Compared to the cross-correlation based high-resolution EBSD (HR-EBSD) analysis, it shows improved robustness against large rotations and achieves accurate measurements even if the rotation angle is as large as 25°. Further investigation also indicates the Siamese network trained on pure aluminum can be generalized at copper, nickel and 316 stainless steel without retraining, suggesting broad and desirable applicability.