Reconstruction of Three-Dimensional Porous Media: Statistical or Deep Learning Approach?

Review of Data Science Trends and Issues in Porous Media Research With a Focus on Image‐Based Techniques

“…This could be due to the inherent randomness of the porous micro-structures that demands for heuristic models and out-of-the-box solutions, such as ML (Meng & Li, 2018). The mainstream ML techniques that have been used in porous material research can be categorized as artificial neural networks (Akratos et al, 2009;Singh et al, 2011;ANNs), deep and convolution neural networks (DNNs and CNNs, respectively; Alqahtani et al, 2018;Santos et al, 2020;Wu et al, 2018), generative adversarial neural networks (GANs; Mosser et al, 2017Mosser et al, , 2018Shams et al, 2020), Bayesian (Mondal et al, 2010), ensemble learning (Al-Juboori & Datta, 2019;Nekouei & Sartoli, 2019), support vector machines (SVMs; Wang, Tian, Yao, & Yu, 2020), self-organizing maps (SOMs; Balam et al, 2018), and Gaussian processes (Crevillen-Garcia et al, 2017). Although, ANN is a general term to address many sorts of trainable network of nodes with any level of complexity in the structures, it is often used to refer to the shallow ANNs which is applicable in the present review, too.…”

“…The objective function is described in Equation and is optimized using the gradient‐descent‐based method (Mosser et al., 2017). Owing to adversarial battle between and , the GAN has an unstable learning curve and trial and error could be demanded (Mosser et al., 2018). During the training, although tries to label any received sample correctly, at the end of training, learns how to create synthetic realizations to fool (Mosser et al., 2017).…”

“…In addition to visual inspection, different morphological, statistical and dynamical parameters can be calculated to evaluate the performance of reconstruction process. The comparison between Minkowski Functional, grain and pore chord length, auto-correlation function, porosity, specific surface area, permeability and velocity distributions of original and reconstructed porous media are common in the literature (Goodfellow et al, 2016;Mosser et al, 2017Mosser et al, , 2018. It is noteworthy that the fully convolutional nature of GAN enables the machine to generate images with desirable sizes, even larger than training dataset at a reasonable computational cost.…”