Reconstruction Of Porous Media Based On Generative Adversarial Network

The three-dimensional micro-structure of rock can be obtained by the X-ray Computed Tomography(CT)technology in the Digital Rock application,and the CT images can be used for 3D modeling and analysis of the physical properties of rocks.However,due to the high complexity of rock internal geometry,the traditional pore structure modeling methods have been unable to meet the demand for high-quality rock microscopic images in modern rock physics analysis.Generative Adversarial Network(GAN)is a deep learning framework proposed in the field of artificial intelligence in 2014,which can effectively construct a mapping from random probability distribution(Mean distribution,Gaussian distribution)to real data distribution.In this paper,GAN is innovatively applied to the study of the reconstruction of porous media such as rocks.Based on limited or damaged porous media image samples,two-dimensional or three-dimensional microscopic images with similar characteristics to real samples are generated from randomly distributed data to realize the reconstruction of porous media model.The advantages of this method are as follows: 1.The reconstructed porous media image is relatively realistic;2.The speed of reconstruction is fast;3.A large number of non-repetitive porous media images can be generated.In this paper,the deep learning method for the reconstruction of porous media images using three-dimensional porous media images is first studied.Some improvements are made on the basic GAN structure.A ternary GAN structure is designed by combining Wasserstein GAN and three-dimensional Deep Convolutional Generative Adversarial Network(3DCGAN)network.Secondly,considering that two-dimensional images are more economical with less computing resources being occupied,this paper proposes a method for reconstruction of three-dimensional porous media images using only two-dimensional porous media images.We compare this method with the reconstruction of porous media images using three-dimensional images.Then,the problem of image repair of damaged porous media is studied by combining the GAN with the automatic encoder.Based on the network structure of pix2 pix network,an algorithm model for repairing damaged porousmedia images into corresponding complete images is developed,and compared with traditional image repair algorithms in this paper.Finally,the paper presents the numerical results of physical properties using the generated three-dimensional porous media model.The experimental results show that the porous media model generated by this method is consistent with the real porous media model in physical and statistical characteristics.The GAN effectively expands the data set of the porous media model and can be better applied to the study of deep learning methods to predict the physical characteristics of porous media samples.The results show that compared with the traditional 3D model reconstruction method,the porous media reconstruction method based on GAN has the characteristics of real and non-replicating,fast speed,stable training and consistent physical characteristics with real porous media.It fully demonstrates the application potential of deep learning methods in interdisciplinary fields such as porous media modeling and analysis.