Abstract: This study develops a deep learning model that is capable of reconstructing the three-dimensional (3D) microstructure of porous materials from limited and even a single two-dimensional (2D) image. This model addresses the time-consuming, money-consuming and labor-consuming challenges of image sampling in traditional experimental imaging methods. A deep convolutional generative adversarial network architecture for 2D and 3D reconstructions are constructed by combining a deep convolutional neural network with a loss function featuring gradient penalty coefficients. Compared with the traditional stochastic optimization method, this method shows advantages in terms of accuracy and efficiency. Moreover, based on the 3D reconstructed microstructural samples, the permeability, effective thermal conductivity and normalized diffusion coefficient of porous materials are successfully predicted. The results demonstrate that the deep learning framework can effectively restore the statistical characteristics and macroscopic transport properties of the target structures.