Reconstruction of three-dimension digital rock guided by prior information with a combination of InfoGAN and style-based GAN

Abstract

In digital rock physics, the study of physical parameters and flow characteristics of reservoirs requires a wealth of three-dimension digital rock samples. However, traditional physical methods of obtaining digital rock are expensive, and numerical reconstruction method cannot obtain a reasonable pore structure for complex rock. Recently, generative adversarial networks (GANs) have proven to be a successful method for reconstructing pore-scale models, but the reconstructed large-size digital rocks look unreasonable due to the lack of sufficient consideration of multi-scale information fusion and it takes a lot of computational resources and time to build the model. Hence, we proposed a combination of InfoGAN and style-based GAN guided by prior information (CISGAN) to reconstruct more controllable and reasonable digital rock models using small set of samples. Porosity distribution as prior information is added into latent space to control pore distribution, and a classifier Q is set in discriminator to ensure the porosity is limited to reasonable range. Multi-scale information is applied each layer by style transfer and used to optimize the background information, pore structure, and micro information of model for multi-scale fusion to produce more reasonable and natural digital rock. Simple-structure sandstone and complex-structure carbonate are implemented to test the reconstruction ability of network. The result shows that synthetic sample has high consistency on pore-throat geometry and connectivity, and the CISGAN can produce natural and user-specified digital rock samples. And the CISGAN will provide more reasonable and various type samples for intelligent parameter prediction of digital rock.