A 3D reconstruction method of porous media based on improved WGAN-GP. (August 2022)

• Record Type: Journal Article
• Authors: Zhang, Ting; Liu, Qingyang; Wang, Xianwu; Ji, Xin; Du, Yi
• Is Part Of: Computers & geosciences. Volume 165(2022)
• Publication Date: 2022-08
• Subjects: Pore structure -- Deep learning -- Training image -- Multiple-point connectivity -- Representative elementary volume; Environmental policy -- Periodicals; 550.5
• DOI: 10.1016/j.cageo.2022.105151
• Languages: English
• ISSNs: 0098-3004

Abstract

Abstract: The reconstruction of porous media is important to the development of petroleum industry, but the accurate characterization of the internal structures of porous media is difficult since these structures cannot be directly described using some formulae or languages. As one of the mainstream technologies for reconstructing porous media, numerical reconstruction technology can reconstruct pore structures similar to the real pore spaces through numerical generation and has the advantages of low cost and good reusability compared to imaging methods. One of the recent variants of generative adversarial network (GAN), Wasserstein GAN with gradient penalty (WGAN-GP), has shown favorable capability of extracting features for generating or reconstructing similar images with training images. Therefore, a 3D reconstruction method of porous media based on an improved WGAN-GP is presented in this paper, in which the original multi-layer perceptron (MLP) in WGAN-GP is replaced by convolutional neural network (CNN) since CNN is composed of deep convolution structures with strong feature learning abilities. The proposed method uses real 3D images as training images and finally generates 3D reconstruction of porous media with the features of training images. Compared with some traditional numerical generation methods and WGAN-GP, this method has certain advantages in terms of reconstruction quality and efficiency. Highlights: IWGAN-GP can save and reuse model parameters for subsequent reconstructions. IWGAN-GP is fast by reusing previous model parameters. IWGAN-GP outperforms traditional methods in the reconstruction of porous media.