Physical properties of rocks such as permeability, relatively permeability, and dispersion coefficient are of critical importance for prediction of subsurface flow and transport. Advances in microscopic imaging have made it possible to obtain the pore-scale microstructure of rock samples via SEM, TEM, or CT scan at low costs and fast turn-around time. Image-based reconstructions have also opened the door for obtaining rock characteristics by using physics-based pore-scale simulations. Unfortunately, such simulations are expensive, and as a result there are generally far more images than that can be simulated with current simulation technology and computational resources. A scheme for fast and accurate evaluation of physical properties of rock samples is urgently needed. In this work, we propose a computational framework and a workflow to evaluate permeability of rock samples by combining physics-based simulations and novel machine learning techniques. The computational framework fully utilizes the data provided by the lattice Boltzmann simulations while encoding the fundamental physics of flow in porous media into the architecture of the machine-learning model. The proposed framework is extensively evaluated on samples of a wide range of porosities and surface area ratios. Comparison of machine learning results and the ground truths suggest excellent predictive performance across all regimes, ranging from largely homogeneous pore structures to those with significant heterogeneities. The computational costs are reduced by several orders of magnitude compared to traditional simulations, demonstrating a great potential of the proposed framework.
|Acceptance of Terms and Conditions||Click here to agree|