Research Of Digital Core Reconstruction Based On Generative Adversarial Networks

There are many methods for digital core reconstruction,such as numerical reconstruction,physical experiment,and hybrid modeling.However,these methods have the disadvantages of long training time and complicated implementation process.Generative adversarial networks(GANs),as the most popular artificial intelligence models in the current image generation field,have excellent image production capabilities,such as excellent performance in the generation of people's images and fingerprints.Therefore,based on the sample production capacity of GANs,this paper proposes a novel digital core reconstruction method.First,the convolutional neural network is used as a generative network to learn the distribution of real digital core samples.Then,the convolutional neural network is constructed as a discriminative network to distinguish reconstructed from real digital core.The generative network tries to deceive the discriminative network,and the discriminative network distinguishes reconstructed digital core as much as possible.Through the confrontation training method,the realistic digital core is finally reconstructed.To analyze the quality of digital cores produced,this paper uses a two-point covariance function,the Fréchet distance and the Kernel distance to evaluate the ability to generate real digital core samples.The results show that the covariance function can test the similarity between real and generated shale samples;the Fréchet distance and the Kernel distance quantitatively measure the difference between the Gaussian distribution of the generated shale sample and the real shale sample.The smaller the value,the higher the quality of the generated shale sample.In addition,to solve the irregular texture problem of GANs in the process of generating shale samples,this paper proposes to use the Three-value-segmentation shale image pre-processed as training samples,and successfully compares the morphological characteristics of real and reconstructed digital core,such as organic matter component ratio and specific surface area.The research in this paper shows that using powerful deep learning methods,GANs can efficiently reconstruct high-quality digital cores.Compared with the multi-point geostatistical method,the new method does not require a priori inference of the probability distribution associated with the training data,and does not need to give hard data constraints in advance,and can directly use noise to reconstruct the digital core of shale.In addition,the new method is simpler,the model training time is shorter,and the trained model can also achieve unlimited production of new samples.Therefore,compared with the existing method,the new method has more potential.The main work of this article includes the following parts:The first part mainly introduces the research background and research status of digital core.The second part mainly introduces the traditional digital core reconstruction methods,such as physical experiment method,numerical reconstruction method and hybrid modeling method.The third part mainly introduces the basic concept,development and realization principle of generative adversarial networks.The fourth part mainly introduces the WGAN-based shale image reconstruction method,related details and algorithm training process,and uses the covariance function and the Fréchet distance,the Kernel distance to measure the similarity between reconstructed and real shale sample,verifying the efficiency of the algorithm.The fifth part mainly introduces the Three-value-segmentation image generation method based on the generated adversarial networks,including the preprocessing flow such as the Three-value-segmentation shale sample,the implementation details of the algorithm,and compares the morphological characteristics of the organic matter ratio or specific surface area between real and reconstructed digital core,we solved the problem of irregular texture in the samples generated by GANs.