Fast and Scalable Digital Rock Reconstruction using Spatially Assembled Generative Adversarial Neural Networks

Because of expensive, labor-intensive, and time-consuming data collection process, geomaterials with complex geometry and topology such as shales and carbonate rocks, are typically characterized with sparse field data samples. Accordingly, generating arbitrary size of 3D rock structures with topological and connectivity features similar to original samples at a low computation cost is one of the key tasks for realistic geomaterial reconstruction. Recently, generative adversarial neural networks (GANs) have demonstrated remarkable results over traditional multi-point geostatistics and has opened up new possibilities for realistic rock reconstruction with reliable uncertainty quantification. However, GANs are often limited by considerable computational costs for high-dimensional image reconstruction applications. In this presentation, we propose an efficient and scalable method called the spatially assembled GANs (SA-GANs) where a generator reconstructs arbitrary size of geomaterial images and a discriminator evaluates their spatial characteristics by assembling smaller image patches from the reconstructed images. The performance of the SA-GANs was tested by reconstructing widely-used 2D and 3D geomaterials. It is shown that SA-GANs can generate the arbitrary size of statistical realizations with similar connectivity and structural properties to training samples, and even a single training sample can result in multiple size of realizations that standard GANs fail to achieve. The computational time is significantly improved compared to those from standard GANs approaches, especially for 3D geomaterial reconstruction examples.

Sandia National Laboratories is a multi-mission laboratory managed and operated by National Technology & Engineering Solutions of Sandia, LLC, a wholly owned subsidiary of Honeywell International, Inc., for the U.S. Department of Energy's National Nuclear Security Administration under contract DE-NA0003525.