Porous media reconstruction using deep texture synthesis

The field of Digital Rock Physics relies on three-dimensional representations of porous media, acquired using X-ray microtomography (µCT), for extracting macroscopic properties of rocks. The resolution of such representations is limited by the resolution of µCT; higher-resolution information is available, but only with two-dimensional imaging techniques. Porous media reconstruction aims at reconstructing three-dimensional samples of a rock, given only a two-dimensional image, while preserving the rock's macroscopic properties. This work proposes a deep learning solution to this problem that leverages texture networks, which have been recently acclaimed for their impressive results in texture synthesis. The solution learns an implicit representation of the pore structure, which allows it to generate multiple realizations of arbitrary size very rapidly. To validate the reconstruction procedure, we introduce a set of metrics that compare the reconstructed samples with the ground truth. Our results show that the solution successfully preserves the relevant rock properties, both in terms of statistical quantities and physical macroscopic properties such as the rock's permeability. We conclude that texture networks are an effective approach to porous media reconstruction, and they open up interesting research directions in the domain.