3D reconstruction of porous media using a batch normalized variational auto-encoder

• Published in: Computational geosciences Vol. 26; no. 5; pp. 1261 - 1278
• Main Authors: Zhang, Ting, Yang, Yi, Zhang, Anqin
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.10.2022; Springer Nature B.V
• Subjects: Coders; Deep learning; Earth and Environmental Science; Earth Sciences; Experimental methods; Feature extraction; Geotechnical Engineering & Applied Earth Sciences; Hydrogeology; Image reconstruction; Mathematical Modeling and Industrial Mathematics; Mathematical models; Methods; Neural networks; Numerical methods; Original Paper; Porous media; Soil Science & Conservation; Teaching methods; Reconstruction; Variational auto-encoder; Squeeze-and-excitation network; 94A08; Pore space; 62 M30; 86A32; 62H11; Batch normalization
• ISSN: 1420-0597; 1573-1499
• DOI: 10.1007/s10596-022-10159-1

The 3D reconstruction of porous media plays a key role in many engineering applications. There are two main methods for the reconstruction of porous media: physical experimental methods and numerical reconstruction methods. The former are usually expensive and restricted by the limited size of experimental samples, while the latter are relatively cost-effective but still suffer from a lengthy processing time and unsatisfactory performance. With the vigorous development of deep learning in recent years, applying deep learning methods to 3D reconstruction of porous media has become an important direction. Variational auto-encoder (VAE) is one of the typical deep learning methods with a strong ability of extracting features from training images (TIs), but it has the problem of posterior collapse, meaning the generated data from the decoder are not related to its input data, i.e. the latent space Z . This paper proposes a VAE model (called SE-FBN-VAE) based on squeeze-and-excitation network (SENet) and fixed batch normalization (FBN) for the reconstruction of porous media. SENet is a simple and efficient channel attention mechanism, which improves the sensitivity of the model to channel characteristics. The application of SENet to VAE can further improve its ability of extracting features from TIs. Batch normalization (BN) is a common method for data normalization in neural networks, which reduces the convergence time of the network. In this paper, BN is slightly modified to solve the problem of posterior collapse of VAE. Compared with some other numerical methods, the effectiveness and practicability of the proposed method are demonstrated.