Generating Heterogeneous 3D Geological Microstructures from 2D Images via a Stable Diffusion-Adversarial Model

TL;DR

The paper presents a new method that generates detailed 3D geological microstructures from 2D images using a hybrid model combining diffusion-based generation with adversarial training. By replacing the typical denoising objective with an adversarial loss, the model achieves stable training and produces high-quality 3D reconstructions without requiring 3D ground truth data. The generated structures accurately reflect phase distributions and key structural features.

Why It Matters

This approach enables accurate 3D microstructure modeling for materials like clay and cement without the need for costly or technically challenging 3D imaging, which is crucial for applications in geotechnical engineering and long-term waste containment. It expands the utility of deep generative models in materials science by handling complex heterogeneous systems more effectively than prior methods.

Key Findings

• The hybrid diffusion-adversarial framework enables stable training without access to 3D ground truth data.
• Generated 3D microstructures exhibit minimal slice-related artifacts and match real phase fractions and structural descriptors closely.
• The model outperforms existing GAN-based approaches in capturing fine-scale heterogeneity in complex geological materials.