Year: 2026
Type: article
Abstract: The reconstruction of three-dimensional (3D) multiphase microstructures is essential for understanding the physical properties of porous materials. In this study, we evaluate the performance of Denoising Diffusion Probabilistic Models (DDPMs) and Generative Adversarial Networks (GANs), specifically WGAN-GP and iWGAN, in generating 3D representations of multiphase materials with varying degrees of heterogeneity: homogeneous illite clay, heterogeneous Boom Clay, and highly complex concrete. Our findings demonstrate that DDPMs outperform GANs in capturing the intricate spatial statistics and multiphase structures of these materials. While GAN-generated samples exhibit mode collapse and structures that do not resemble the ground truth, DDPMs produce microstructures that better preserve phase distributions and morphological characteristics. These results highlight the potential of diffusion models for realistic 3D microstructure synthesis, paving the way for improved simulations in subsurface and construction material applications.
Source: Computational Geosciences
Institutions: Belgian Nuclear Research Centre, UCLouvain
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W4415345814