Abstract: In order to explore the related properties of porous materials at pore size scale and better apply them to engineering, it is necessary to fully understand the internal pore structure of materials. Because it is difficult to obtain 3D CT images directly, the deep learning method is considered to approximate the real probability distribution of porous materials, and the random number is used for 3D reconstruction. Firstly, the existing binary images are segmented to obtain data sets with appropriate size and sufficient quantity. Then, WGAN (Wasserstein Generative Adversarial Networks) is transformed into a generativeadversarial model that can process three-dimensional data, and the data is generated by using generator andthe random noiseobeying Gaussian normal distribution. Finally, WGAN is trained by generating data and real data. By calculating the two-point correlation function, Minkowski functional and permeability of the generated image, the relevant parameters such as porosity, specific surface area and function curve fitting are investigated, and tested with different data sets. According to the results, the three-dimensional reconstruction model based on WGAN has high accuracy under different input conditions. At the same time, the algorithm also has the characteristics of low time complexity and the generated model can be stored and reused.

Key words: 3D reconstruction')">

3D reconstruction, porous media, Wasserstein distance, generative adversarial networks