Hybrid Deep Learning for 3D Reconstruction of Multi-Mineral Porous Media: Integrating U-Net and GAN for Enhanced Segmentation and Texture Preservation

This study proposes a hybrid deep learning approach combining U-Net and Generative Adversarial Network (GAN) architectures for the segmentation and texture-based reconstruction of 3D multi-mineral porous media images. The dataset consists of high-resolution 3D Leopard sandstone images, segmented into four key mineral classes: macro-pores, clay, quartz, and high-density minerals. Our approach leverages the feature extraction capabilities of a ResNet-18 backbone within U-Net, pre-trained specifically for multi-mineral segmentation, which then feeds these detailed features into a GAN framework for image reconstruction. The model effectively bridges segmentation and reconstruction, achieving superior image quality and structural fidelity compared to standalone GAN models by preserving intricate textures and maintaining macroscopic rock structures. Quantitative assessments reveal that the hybrid model yields porosity and absolute permeability values with minimal discrepancies (2.25% and 1.54% error, respectively) compared to actual data. These findings highlight the model's ability to replicate critical geophysical metrics and generate accurate 3D representations. Unlike traditional methods that either focused solely on segmentation or reconstruction, our model uniquely integrates segmentation-driven texture data for image reconstruction, offering a novel solution for geoscientific applications in hydrogeology, petroleum engineering, and environmental science