From Traditional Approaches to Single-Image Generative Models for Stationary and Nonstationary Porous Media

Summary The generation of representative porous media images is a critical step for training machine learning (ML) models in applications where experimental data are scarce and expensive to obtain. Traditional approaches, such as process-based simulations, object-based methods, and multiple-point statistics (MPS), have been widely used for this purpose, but they all come with some limitations. In this paper, we explore the potential of a single-image-based ML, a generative adversarial network (GAN) model capable of learning from a single training image (TI), for porous media modeling. We test the performance of this method on a variety of systems, including binary pore structures, grayscale porous media, and nonstationary cases such as fracture networks and multifacies systems. The results show that the produced realizations are visually and statistically consistent with the TIs, yielding pore and solid size distributions, two-point correlation functions, connectivity measures, and permeability values comparable with those from state-of-the-art methods. More importantly, this method is able to handle nonstationary systems without the need for auxiliary data, a major limitation of conventional approaches.