Research on Digital Core Reconstruction of Tight Sandstone Based on Deep Learning

Three-dimensional digital core reconstruction methods based on deep learning have emerged as a promising approach to address the limitations of traditional physics experiment-based and numerical computation-based methods. This study focuses on tight sandstone digital core reconstruction using generative adversarial networks (GANs). A multi-scale feature fusion network is proposed, which reduces reconstruction loss by 23.57%, and a loss decomposition strategy is adopted that increases training speed by 38.57%. Comparative experiments among DCGAN, LS-GAN, and WGAN demonstrate that DCGAN cannot effectively complete the 3D digital core reconstruction task, while both LS-GAN and WGAN produce high-quality reconstructions. Compared with LS-GAN, WGAN improves reconstruction efficiency by 27.14% and reduces final reconstruction loss by 11.17%. Quantitative evaluation using SSIM, PSNR, and porosity–permeability correlation confirms that WGAN-generated digital cores exhibit the highest structural and physical fidelity. Pore network analysis further reveals that WGAN best preserves the topological connectivity of the original rock. Statistical reproducibility tests across five different random seeds demonstrate the robustness of the proposed method. The proposed approaches provide an efficient and reliable solution for tight sandstone digital core reconstruction in geoscience applications.