Exploring deep generative modeling approaches with 2D to 3D polycrystalline microstructures generation

Abstract

Microstructure reconstruction has been a significant area of research due to its direct usage in multiscale modeling and micro-mechanical simulations. However generation of 3D microstructures using experimental methods such as FIB-SEM or X-Ray tomography is a tedious process, so alternative approaches have been developed for reconstructing 3D microstructures. In the present work, we attempt to generate high-fidelity 3D microstructures from 2D images using deep generative modeling approaches such as variational autoencoders (VAEs) and generative adversarial networks (GANs). Since obtaining experimental images for training generative models is difficult and expensive, synthetic microstructures generated by DREAM.3D have been employed for this purpose. The generative model predicted 3D microstructures, are statistically compared with their real synthetic counterparts through microstructure-descriptors used previously for characterizing them. The generative models have been trained on the given data to efficiently generate 3D microstructures by employing Unet, conditional variational autoencoders (cVAEs), pix2pixGAN, and Wasserstein conditional generative adversarial networks (WCGANs) frameworks. The trained Unet model resulted in overfitting, while the rest were successful in the generalization of data. The trained pix2pixGAN model has been able to learn the grain structure but is not capable of capturing grain boundaries and the underlying statistics entirely. Postprocessing techniques are necessitated for optimization and the formation of a perfect grain-like structure. Besides an algorithm has been developed as a postprocessing tool to increase the complexity of synthetic microstructures by introducing lamellae.