Deep Learning for Accurate Prediction of Physical Properties of Heterogeneous Digital Rocks
Digital rock physics (DRP), an approach to predicting physical properties of porous media based on images, has been utilized in various fields, such as geosciences, environmental engineering, and civil engineering. Accurate prediction of physical properties in DRP depends on high-resolution (HR) and large-view 3D images. However, it is of great difficulty to obtain such images since current imaging techniques strike a balance between the image resolution and field of view. Moreover, few HR 2D SEM images are only available in real scenarios, and these HR images are not matched with low-resolution (LR) CT images. To alleviate these issues, we proposed a hybrid unsupervised deep learning method (FastGAN-CycleGAN) to fuse multiscale structure information from HR 2D SEM images for reconstructing HR and large-scale 3D images based on a few unpaired training images. The generated HR 3D images not only keep the morphological features of minerals and pores but also resolve the microstructures and fine details which are nonexistent in original LR images. The comparison reveals the physical properties of the reconstructed HR 3D images by deep learning are closer to the experimental data than the LR images and HR images from bicubic interpolation. The study opens a new avenue for construction of the multiscale HR porous media, characterization of complex rocks and accurate prediction of physical properties.
Files not available for this dataset
This contains only metadata