Parametric Simulation of Electron Backscatter Diffraction Patterns through Generative Models

Abstract Recently, discriminative machine learning (ML) models have been widely used to predict various attributes from Electron Backscatter Diffraction (EBSD) patterns. However, there has never been any generative model developed for EBSD pattern simulation. On one hand, the training of generative models is much harder than that of discriminative ones; On the other hand, numerous variables affecting EBSD pattern formation make the input space high-dimensional and its relationship with the distribution of backscattered electrons complicated. In this study, combining two famous generative models, we propose a framework (EBSD-CVAE/GAN) with great flexibility and scalability to realize parametric simulation of EBSD patterns. Compared with the frequently used forward model, EBSD-CVAE/GAN can take variables more than just orientation and generate corresponding EBSD patterns in a single run. The accuracy and quality of generated patterns are evaluated through multiple methods. The model developed does not only summarize a distribution of backscattered electrons at a higher level, but also offers a new idea of mitigating data scarcity in this field.