Learning to Reconstruct 3D Porous Structures from 2D Images via Spatiotemporal Gating Networks

T he three-dimensional (3D) micro structure of porous media serves as a vital bridge between microscopic morphology and macroscopic physical properties, with significant implications for fields such as energy exploration and biomedicine. Due to the inability of existing neural network models to effectively capture inter layer patterns and long-range temporal dependencies, reconstructing 3D structures from limited two-dimensional (2D) images using computational modeling remains a key challenge. To address this, we propose a spatiotemporal convolutional gate ,which integrates an enhanced dynamic memory bank to improve long-range feature extraction, an adaptive forget gate to balance temporal continuity and morphological diversity.Meanwhile,to enhance local spatial feature modeling around pore throats,it combines attention and residual networks. The model was validated using Bernheimer sandstone samples through morphological similarity metrics, multiscale statistical analyses, and physical property simulations, with four model variants confirming its robustness and structural diversity. These results demonstrate the effectiveness of the proposed approach and establish a new 3D porous structure reconstruction method for digital rock modeling in petroleum engineering.