Multi-scale reconstruction of porous media based on progressively growing generative adversarial networks

The fluid flow and heat transfer in porous media are not only related to the physical properties of the fluid itself, but also related to the structure and distribution of the pore space of porous media. The reconstruction of porous media has become a prerequisite for the research and analysis of the microscopic pore structure. At present, numerical simulation methods are widely used in the field of porous media reconstruction, which can achieve reconstructed results similar to the real pore structure, but generally the whole process is CPU-intensive and time-consuming. With the widespread application of deep learning in many scientific fields, although generative adversarial network (GAN), as a branch of deep learning generative models, has the strong ability of feature extraction and prediction, it is challenged by the large CPU/memory consumption in training. One solution is to perform multi-scale reconstruction, which has been used in the variants of GAN. Based on multi-scale reconstruction, smaller-scale training samples can be used to reconstruct larger-scale samples with the benefits of a much faster speed than traditional numerical simulation methods and lower burdens on CPU/memory. Hence, this paper proposes a progressively growing multi-scale GAN model for the reconstruction of porous media. Compared with some variants of GAN and traditional numerical simulation methods, the effectiveness and practicability of our method are proved.