Fast Estimation of Permeability in Sandstones by 3D Convolutional Neural Networks

The motivation of this study is to apply the state-of-the-art artificial intelligence (AI) approach to tackle the challenging permeability estimation problem which is time-consuming and sometimes unreliable. In this study, we applied the tailor-made Convolutional Neural Network (CNN) for fast permeability estimation with incorporated Lattice-Boltzmann Method (LBM) direct permeability simulations. The micro-CT scanned rock images in this study are described, the principles of the LBM direct simulations and CNN set up are introduced, the training and testing process, as well as the validation results, are discussed. Our results demonstrated the designed CNN approach has strong capabilities for estimating permeability directly from input 3D images based on the two real rock samples: Berea sandstone and Bentheimer sandstone. The prediction is extremely fast compared to numerical simulations and experimental measurements. The significance of this integrated approach is the model is able to leverage the powerful feature recognition of CNN combined with numerical simulations and data augmentation techniques for fast physical property estimation directly from different rock images with various levels of complexity. Additionally, the scalability and extensibility of this workflow allow the extraction of various rock-fluid properties and the evaluations at multi-scale.