PM-ARNN: 2D-TO-3D reconstruction paradigm for microstructure of porous media via adversarial recurrent neural network

The availability of high-quality 3D microstructures is an essential prerequisite for simulating and studying transport processes and physical properties of porous media. Such numerical simulation analysis can provide basic data for oil and gas exploration and thus improve the extraction efficiency of reservoirs. Considering the deficiencies of 3D microscopy imaging, including its high cost, complicated operation, and the contradiction between resolution and field of view, a new method to obtain a large number of synthetic 3D structures from the knowledge of 2D images cheaply and efficiently is required. This study proposed a recurrent neural network (RNN)-based generative model incorporating a generative adversarial network (GAN) to address the 2D-TO-3D reconstruction problem. The hybrid generative model can learn the 2D morphological features of the pores and the spatial relationship between adjacent layers through a series of 2D slices. Upon training the model, it receives a 2D slice as the input, and the corresponding full 3D structures can be recovered layer-by-layer. Further, to overcome the inherent defects of the layer-by-layer generative model, an adversarial training strategy incorporating an RNN was proposed to better perceive the 3D structure of pores and further enhance the reconstruction ability of adjacent layers. Consequently, visual and statistical comparisons were performed to evaluate the effectiveness of the proposed model. Reconstruction experiments were conducted on homogeneous and heterogeneous porous media, as well as fractured cores, to validate the accuracy, diversity, and stability of the proposed model. The results confirmed the ability of the proposed model to reproduce 3D realizations that precisely preserved the statistical characteristics, morphological features, and long-distance connectivity based on the 2D input image.