Abstract

We use Wasserstein Generative Adversarial Networks to learn and integrate multi-scale features in segmented three-dimensional images of porous materials, enabling the dependable generation of large-scale representations of complex porous media. A Laplacian pyramid generator is introduced which creates pore-space features with a hierarchy of spatial scales. Feature statistics mixing regularization enhances the ability of the model generation to reliably maintain multi-scale pore-space features of images by increasing diversity. The method is tested on a variety of X-ray images of porous rocks. The generated images can be of any size -- cm-scale ten-billion-cell images are generated to demonstrate the power of the approach -- which have two-point correlation functions, porosity, permeability, Euler characteristic, curvature, and specific surface area close to the training datasets. The images demonstrate a considerable improvement over previously- published studies using generative adversarial networks.