A novel generative adversarial networks based multi-scale reconstruction method for porous rocks

The traditional reconstruction methods for numerical rock models, such as simulated annealing reconstruction method, have disadvantages, such as unclear details of the generated structure and the need of prior functions. Therefore, this paper attempts to introduce GANs-based techniques to reconstruct numerical porous rock models. The introduction of GANs-based techniques can solve the problem of requiring prior functions before reconstruction and can improve the clarity and richness of the generated reconstruction models in terms of details. First, compression and computer tomography tests are conducted to obtain the necessary parameters. Then, the generative adversarial network (GAN) method is introduced to propose the novel multi-scale reconstruction method. Later, the GAN reconstruction method is used to generate multi-scale structures of rocks. After, the equivalence in statistics between the reference and reconstructed model is verified by the two-point probability distribution function. The equivalence in topology between the reference and reconstructed model is verified by the modified skeleton algorithm, and the equivalence in mechanical property between the reference and reconstructed model is verified by the numerical results. The verifications also show that this proposed novel multi-scale reconstruction method has great potential in engineering applications.