Digital Rock Reconstruction Using Wasserstein GANs with Gradient Penalty

Abstract Due to the scarcity and vulnerability of physical rock samples, digital rock reconstruction plays an important role in the numerical study of reservoir rock properties and fluid flow behaviors. With the rapid development of deep learning technologies, generative adversarial networks (GANs) have become a promising alternative to reconstruct complex pore structures. Numerous GAN models have been applied in this field, but many of them suffer the unstable training issue. In this study, we apply the Wasserstin GAN with gradient penalty, also known as the WGAN-GP network, to reconstruct Berea sandstone and Ketton limestone. Unlike many other GANs using the Jesnen-Shannon divergence, the WGAN-GP network exhibits a stable training performance by using the Wasserstin distance to measure the difference between generated and real data distributions. Moreover, the generated image quality correlates with the discriminator loss. This provides us an indicator of the training state instead of frequently subjective assessments in the training of deep convolutional GAN (DCGAN) based models. An integrated framework is presented to automate the entire workflow, including training set generation, network setup, model training and synthetic rock validation. Numerical results show that the WGAN-GP network accurately reconstructs both Berea sandstone and Ketton limestone in terms of two-point correlation and morphological properties.