Recognition of Permeability from Images with Convolutional Neural Networks

In the study of porous media, image-based analysis immediately emerged as scanning high-resolution images of porous media became available. Pore-scale simulations often incur significant computational costs. The success of image recognition neural networks motivated us to seek fast prediction of porous media properties directly from images. Our steps to validate this concept included (1) generation of synthetic porous media samples, (2) computation of permeability via fluid dynamics simulations, (3) training of convolutional neural networks (CNN) with simulated data, and (4) validations against simulations. Comparison of machine learning predictions and the ground truths from simulations suggests excellent predictive performance across a wide range of porosities and pore geometries, especially for those with dilated pores whose permeability cannot be estimated using the conventional Kozeny-Carman approach. Incorporation of physical parameters (physics-informed CNN) improved the performance of the neural network. CNN-based methods are orders of magnitude faster than direct simulations using lattice Boltzmann. The proposed framework should be applicable to other physical properties of porous media as long as they are solely governed by pore geometry.