A state-of-the-art conditional denoising diffusion probabilistic model (DDPM) for reconstructing three-dimensional porous media from limited surface images.
AB-CDM leverages recent advances in diffusion models to achieve high-fidelity 3D porous media reconstructions guided by 2D surface observations. This framework enables detailed recovery of internal pore architectures, which is essential for digital rock analysis, inverse design, and related applications in geosciences and materials science.
-
Time-step Embedding: Encodes the diffusion process for improved sampling quality and model stability.
-
Spatial Conditional Embedding: Injects surface image features along principal directions, modulated by learnable exponential attenuation, ensuring spatially consistent guidance throughout the generation process.
-
Conditional Residual Blocks: Enhance feature propagation and flexibility; optional attention mechanisms improve representation of complex microstructures.
-
U-Net Backbone: Employs hierarchical feature extraction with skip connections for efficient, multi-scale learning.
-
Physics-driven Evaluation: Supports permeability prediction for rigorous validation of generated porous structures.
- Training:
python MainCondition.py - Inference:
python MainCondition_eval.py
- Datasets and Pretrained Models: Public datasets and pretrained models are available via Zenodo (DOI: 10.5281/zenodo.15617273).
The code structure was inspired by "https://github.com/zoubohao/DenoisingDiffusionProbabilityModel-ddpm-/tree/main."