Minireview on Porous Media and Microstructure Reconstruction Using Machine Learning Techniques: Recent Advances and Outlook

Abstract

To accurately model and analyze various phenomena in porous media, it is necessary to have three-dimensional (3D) microstructure samples that genuinely represent the key characteristics of the material being studied. Although modern imaging methods such as X-ray computer tomography have enabled the extraction of 3D images of pore space, obtaining 3D microstructure images remains a costly and time-consuming process. Therefore, it is becoming more crucial to use image reconstruction techniques to create new realizations of microstructures. Machine learning (ML) techniques have been increasingly used to generate porous microstructures in recent years. This article reviews some of the most promising studies that have introduced such ML-based reconstruction techniques. We categorize the existing approaches into different types while highlighting their key characteristics, advantages, and disadvantages. Furthermore, this paper provides information on various methods for evaluating the performance of the reconstruction algorithms. We also discuss the current research status of ML-assisted porous media reconstruction in energy-related applications, namely, energy storage systems and digital rock reconstruction. Despite their success and rapid progress over the past decade, ML-based approaches, particularly generative deep neural networks, still suffer from issues that impede their further applications. The paper elaborates on some of the most important limitations and current challenges of these approaches while indicating potential research areas for future studies.