A review of the use of machine learning to predict the properties of porous media

Abstract

Understanding the properties of porous media is important in a wide range of applications. However, they are often difficult to determine using experimental techniques or fluid simulations. In light of this, methods for predicting properties using machine learning have become popular. A wide range of machine learning techniques have been commonly used for porous media property prediction including the convolutional neural network (CNN) and the graph neural network (GNN). Many early studies represent porous media with 2D images, however, this is often an over-simplification that does not translate well to the properties of 3D porous media. Building 3D models, such as the 3D CNN can lead to more accurate results, but scalability issues are introduced for porous media at the representative elementary volume (REV) scale. In this review, the current state of the art in deep learning for predicting porous media properties, such as permeability of thermal conductivity, is accessed, focusing on both the accuracy and scalability of the models, when considering porous structures of a representative scale. Finally, future areas of research are suggested to achieve more accurate, robust and scalable models, such as couplings of GNNs and CNNs to produce models that exhibit the benefits of both techniques.