3D Grain Shape Generation in Polycrystals Using Generative Adversarial Networks

This paper presents a generative adversarial network (GAN) capable of producing realistic microstructure morphology features and demonstrates its capabilities on dataset crystalline titanium grain shapes. Alongside this, we present an approach to train deep learning networks understand material-specific descriptor features, such as shapes, based existing conceptual relationships with established spaces, functional object A style-based GAN Wasserstein loss, called M-GAN, was first trained recognize distributions from function objects in the ShapeNet then applied morphologies 3D crystallographic Ti–6Al–4V. Evaluation feature recognition showed comparable or better performance than state-of-the-art voxel-based approaches. When experimental data, M-GAN generated those seen quantitative comparison moment invariant that grains were similar shape structure ground truth, but scale invariance learned led difficulty distinguishing between physical small spatial resolution artifacts. The implications M-GAN’s are discussed, well extensibility this other material characteristics related morphology.