Prediction of effective diffusivity of porous media using deep learning method based on sample structure information self-amplification

Effective diffusivity is one of the basic transport coefficients used to describe the mass transport capability of a porous medium. In this study, a deep learning method based on a convolutional neural network (CNN) with sample structure information self-amplification is proposed to predict the effective diffusivity of a porous medium, which is considerably influenced by the morphological and topological parameters of the porous medium. In this method, the geometric structures of three-dimensional (3D) porous media are reproduced via a stochastic reconstruction method. Datasets of the effective diffusivities of the reconstructed porous media were first established by the pore-scale lattice Boltzmann method (LBM) simulation. A large number of geometric structures of 3D porous media are obtained using the proposed sample structure information self-amplification approach. The 3D geometric structure information and corresponding effective diffusivities are directionally applied to a CNN for training and prediction. The effective diffusivities of media with porosities ranging from 0.48 to 0.58 are employed as training datasets, and the effective diffusivities of media with a broader porosity range of 0.39 to 0.79 are predicted by CNN. The CNN model can achieve a fast and accurate prediction of the effective diffusivity. The relative error between the CNN and LBM is 0.026%–8.95% with porosities ranging from 0.39 to 0.79. For a typical case with a porosity of 0.5, the computation time required by the CNN model is only 3 × 10−4 h, while the computation time for the same case is 16.96 h using the LBM. These findings indicate that the proposed deep learning method has a powerful learning ability; it is time-saving, provides accurate predictions, and can serve as a promising and powerful tool to predict the transport coefficients of complex porous media.