Abstract: Permeability, which represents the ability to transmit fluid, is the key parameter of porous media. The finite volume method (FVM) and lattice Boltzmann method (LBM) for calculating permeability have common shortcomings, i.e., long computational time. To address this difficulty, a deep-learning approach is proposed for rapidly predicting the permeability of porous media in this paper. First, 40 real porous media images are obtained using X-ray micro-computed tomography, and 400 realizations of porous media images are generated by the synthetic approach. Afterward, direct pore-scale modeling with the FVM is used to compute the permeability of porous media. A dataset including 440 porous media images is obtained and 90% of the samples are used for training, 10% are used for testing. A deep learning network is built for estimating permeability. Based on the trained model, satisfactory predictions of the permeability are achieved with an accuracy of ±14% in the testing dataset. It can be concluded that the trained deep-learning network takes tens of milliseconds to predict the permeability of one sample, about 10 000 times faster than FVM. The deep-learning approach provides a new way to calculate permeability from the pore microstructure of porous media, and it has the potential to facilitate the understanding of the relation between pore microstructure and permeability.

Key words: porous media, permeability, deep learning