Speaker
Description
Accurate three-dimensional (3D) representations of porous microstructures are essential for predicting transport, mechanical, and reactive behavior in natural and engineered porous media. However, acquiring 3D datasets remains costly, technically demanding, and often infeasible for fragile or fine-grained materials such as clay-based systems. Recent deep generative approaches attempt to infer 3D structures from two-dimensional (2D) images, yet existing methods face important limitations. Classical reconstruction algorithms rely on low-order statistics and struggle with heterogeneous media, while Generative Adversarial Network (GAN)-based models, such as SliceGAN, exhibit unstable training and difficulties reproducing complex multi-phase textures. Diffusion models, although promising, typically require full 3D training data or incur high computational cost.
This work introduces Dimensionality Expansion Diffusion Adversarial Model (DimExDAM), a hybrid generative framework designed specifically for 2D-to-3D microstructure generation using minimal training data. The approach integrates a 3D diffusion-based generator with a single 2D adversarial discriminator. Instead of using a conventional denoising loss, the method employs an adversarial objective computed on orthogonal slices, allowing the model to learn structural consistency without access to 3D ground truth. This formulation stabilizes training, mitigates vanishing-gradient issues common in multi-critic GAN architectures, and reduces sampling redundancy typically observed in diffusion-based reconstruction.
We evaluate DimExDAM on porous materials with increasing structural complexity, including clay, carbonate, and sandstone datasets. Generated volumes are assessed using phase fraction agreement, directional connectivity measures, and structural descriptors relevant to porous media characterization. The model demonstrates: (i) consistent recovery of anisotropic features, (ii) minimal slice artefacts compared with SliceGAN, and (iii) strong statistical alignment with reference descriptors while requiring as little as one 2D training image per orientation. Training exhibits smoother convergence behavior than traditional GAN approaches and avoids the heavy dependence on full 3D volumes inherent to other diffusion frameworks.
The results indicate that DimExDAM provides a robust pathway toward data-efficient 3D reconstruction of complex porous microstructures, enabling realistic synthetic datasets for simulation. Ongoing work explores conditioning strategies and physics-informed priors to further integrate transport-relevant constraints into the generative process.
| Country | Belgium |
|---|---|
| Student Awards | I would like to submit this presentation into both awards |
| Acceptance of the Terms & Conditions | Click here to agree |
