An end-to-end approach to predict physical properties of heterogeneous porous media: Coupling deep learning and physics-based features

Abstract

Digital rock physics (DRP) has become an effective tool to predict the petrophysical properties of rocks and reveal the mass transport mechanisms in porous media. Accurate prediction of the physical properties of heterogeneous rocks based on DRP requires 3D high-resolution (HR) and large-view images. It is, however, extremely challenging to acquire such images since the current imaging technologies cannot resolve the dilemma between the high resolution and large field of view and we often end up with low-resolution (LR) images but with a large field of view or HR images with a small field of view. Moreover, available HR images are limited and always unpaired with accessible LR images, so it is of great difficulty to train a deep learning model based on the limited unpaired images. To address these issues, we used a fast and stabilized generative adversarial network (FastGAN) to synthesize thousands of plausible LR and HR images based on ∼100 unpaired images. Taking the synthetic images as training images, we then utilized a cycle-consistent GAN (CycleGAN) to reconstruct the 3D HR large-scale digital rocks by assimilating the fine-scale structures from 2D HR images into 3D LR images. The accuracy of the proposed method (FastGAN-CycleGAN) is validated by comparing the porosity, pore size distribution, multiple-point correlation, and permeability of the reconstructed digital rocks of shale and carbonate samples with laboratory measurements. The proposed unsupervised approach does not require prior image processing knowledge. Furthermore, it can be also applied to other types of images such as magnetic resonance and fluorescence microscopy images in the future.