3D-PMRNN: Reconstructing three-dimensional porous media from the two-dimensional image with recurrent neural network

Abstract

The accurate reconstruction of three-dimensional(3D) structure of porous media is crucial for predicting their physical or mechanical properties. Reconstructing the 3D structure from its reference two-dimensional(2D) image is an intractable inverse problem. Although many traditional methods have been proposed to address this problem, their low efficiency and poor performance limit their application. Recently, the deep learning(DL) based methods have attracted widespread attention. However, the existing DL-based methods suffer from some serious defects, including heavy demand for training samples, unstable training and high GPU memory requirement. In this study, a recurrent neural network(RNN) based model, namely 3D-PMRNN, has been proposed to overcome these weaknesses. To the best of our knowledge, it is the first time that RNN model has been applied to solve the 2D-TO-3D reconstruction problem. Benefiting from this innovative architecture, the model just requires one 3D training sample at least and reconstructs the 3D structure layer-by-layer. Furthermore, the proposed model can reconstruct a larger-scale realization(2563 or larger) due to the novel network architecture, less demand for GPU memory and stable training process. The reconstruction efficiency has also been greatly improved compared to the traditional methods. Three experiments are carried out on isotropic and anisotropic porous media to verify the model’s performance on accuracy, diversity and generalization. Experimental results indicate that the synthetic realizations have good agreement with the testing target in terms of visual observation and quantitative comparison.