Machine Learning Application for Permeability Estimation of Three-Dimensional Rock Images

Abstract

Estimation of permeability in porous media is fundamental to understanding coupled multi-physics processes critical to various geoscience and environmental applications. Recent emerging machine learning methods with physics-based constraints and/or physical properties can provide a new means to improve computational efficiency while improving machine learning-based prediction by accounting for physical information during training. Here we first used three-dimensional (3D) real rock images to estimate permeability of fractured and porous media using 3D convolutional neural networks (CNNs) coupled with physics-informed pore topology characteristics (e.g., porosity, surface area, connectivity) during the training stage. Training data of permeability were generated using lattice Boltzmann simulations of segmented real rock 3D images. Our preliminary results show that neural network architecture and usage of physical properties strongly impact the accuracy of permeability predictions. In the future we can adjust our methodology to other rock types by choosing the appropriate architecture and proper physical properties, and optimizing the hyperparameters.