Estimating rock permeability and porosity using machine learning

Abstract

Investigating the permeability and porosity of rocks is essential for petroleum production, CO2 geological storage and etc. The permeability is one of the most basic physical properties that evaluate the ease of flow of fluid in a porous object for understanding the behavior of fluids such as water, oil, and CO2 in rocks. Laboratory experiments using excavated rock for permeability estimation is very time and cost-consuming. In recent years, estimation of physical properties of rock using the deep learning technology has received a lot of attention. In this study, we have developed a method for efficiently predicting the physical properties of permeability using the convolutional neural networks (CNN). Four types of convolutional neural networks have been tested: basic CNN, VGG, GoogLeNet, and Resnet. These networks are trained using the dataset of directly calculated permeability values based on flow simulation and the corresponding raw CT images. By comparing the prediction accuracy of each network, we found that the Resnet model showed the best performance.