Generation of highly realistic microstructural images of alloys from limited data with a style-based generative adversarial network
G. Lambard, Kazuhiko Yamazaki, Masahiko Demura
Abstract
In materials science, the amount of observational data is often limited by operating protocols that require a high level of expertise, often machine-dependent, developed for a time-consuming integration of valuable data. Scanning electron microscopy (SEM) is one of those methodologies of characterisation for which the number of observations of a given material is limited to just a few images. In the present study, we present the possibility to artificially inflate the size of SEM image datasets from a limited ([Formula: see text] of images) to a virtually unbounded number thanks to a generative adversarial network (GAN). For this purpose, we use one of the latest developments in GAN architectures and training methodologies, the StyleGAN2 with adaptive discriminator augmentation (ADA), to generate a diversity of high-quality SEM images of [Formula: see text] pixels. Overall, coarse and fine microstructural details are successfully reproduced when training a StyleGAN2 with ADA from scratch on at most 3000 SEM images, and interpolations between microstructures are performed without significant modifications to the training protocol when applied to natural images.