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

This paper addresses the challenge of reconstructing complex 3D geological microstructures—such as those found in clay and cementitious materials—using only 2D images. Because 3D imaging is often expensive, inaccessible, or technically limited, researchers are turning to deep generative models to simulate these 3D volumes. The authors introduce a new hybrid framework that improves upon existing methods to better capture the intricate details of heterogeneous materials.

The Limitation of Current Methods

In the field of microstructure generation, GAN-based models like SliceGAN have been the standard for creating 3D volumes from 2D slices. While SliceGAN performs well for homogeneous materials that look consistent in all directions, it struggles to accurately represent more complex, heterogeneous microstructures. These materials contain finer, more varied details that traditional GANs often fail to capture, creating a need for more robust generative techniques. The same reasoning question is explored in Extending SMT Solving with Non-Ground Clause..., which adds a research perspective.

A Hybrid Generative Approach

To overcome these limitations, the researchers developed a hybrid model that combines the stability and high-quality output of denoising diffusion models with an adversarial training framework. A key challenge in this domain is the lack of 3D ground truth data to train models. To solve this, the authors replaced the standard denoising loss typically used in diffusion models with an adversarial loss. This modification allows the model to learn from 2D data while maintaining a stable training process.

Performance and Accuracy

The resulting model demonstrates a significant improvement in generating complex microstructures. In their experiments, the authors found that the model produces 3D volumes with minimal "slice artifacts"—the unnatural lines or patterns that can occur when stacking 2D images into a 3D space. Furthermore, the generated structures show close agreement with ground-truth phase fractions and structural descriptors, confirming that the model successfully preserves the physical characteristics of the original materials. The same computer vision question is explored in Geospatial AI, Dataverse Metadata, and the..., which adds a research perspective. as detailed in the full paper on Arxiv