Mineral Component Reconstruction Network: Deep Generative Adversarial Network for 2D-3D Multimineral Component Reconstruction of Large-Sized Rock Microcomputed Tomography Images

Summary Mineral composition information obtained from high-quality real 3D rock microcomputed tomography (CT) images is crucial for understanding physical properties, such as seepage and thermodynamic characteristics. Traditional 2D-3D rock micro-CT image reconstruction is based on binary images and cannot reproduce 3D mineral composition. However, reconstructing 3D mineral components is challenging because of the complex spatial distribution of mineral-rich cores, making it difficult to directly fit them with neural networks. In this study, we reformulated the complex 2D-3D multimineral rock reconstruction problem as an alternative task and proposed a four-stage mineral component reconstruction network (MCRNet). The MCRNet comprises a mineral segmentation module (multithreshold Otsu), two-phase 2D-3D dimension promotion module [generative adversarial neural network (GAN) dictionary-based superdimension (GANDB-SD) algorithm], two-phase to multiphase 3D-3D image translation module (improved pix2pix), and mineral-component image fusion module (MC-Fusion GAN). Channel attention (CA) and spatial attention (SA) mechanisms enhance layer-to-layer connections in dimension promotion and the importance of the inner content in image translation. The MC-Fusion GAN loss function was designed to maintain both the physical mechanisms and content detail similarity. Qualitative and quantitative evaluations demonstrated the effectiveness of the proposed MCRNet in the 2D-3D multimineral reconstruction of large rock micro-CT images. In addition, seepage and thermodynamic analyses confirmed that compared with binary-core 2D-3D reconstructions, the 3D structures generated by MCRNet are more accurate for physical property experiments. This study enhanced 3D mineral reconstruction for improved rock characterization, seepage and thermodynamic analysis, and mineral resource assessment.