A Deep Learning Approach to Predicting Sandstone Permeability

Abstract

Sandstone permeability is a critical parameter in reservoir engineering, influencing the extraction efficiency of hydrocarbons. Traditional methods of permeability estimation are time-consuming and require extensive laboratory measurements. In this study, we propose a deep learning model to predict sandstone permeability from petrophysical data. The model leverages a convolutional neural network (CNN) architecture to capture the complex relationships between input features and permeability, demonstrating significant improvements in prediction accuracy compared to conventional methods. This study enhances our previous work on predicting grid-level dynamics in porous media using a data-driven approach. We developed a deep learning surrogate model that accurately predicts permeability regardless of grain density or shape, significantly reducing computational time. High-fidelity simulations for 2D porous media with varying circular grains were used to train the model. The model's robustness was tested on different grain angularities and elongations, not included in the training data. We employed a deep convolutional neural network with ResNet structures to capture context and ensure precise localization.