Abstract
Abstract
Accurate representations of porous microstructures are essential for studying transport and effective properties in geomaterials such as clay, carbonate, and sandstone. Deep generative models provide a means of producing additional statistically representative microstructures, but different model families can differ substantially in their ability to reproduce morphology, population variability, and transport-relevant connectivity. In this work, we present a comparative benchmark of three generative-model families for porous microstructure synthesis: a Wasserstein generative adversarial network with gradient penalty (WGAN-GP), denoising diffusion probabilistic models (DDPMs), and flow matching models. The benchmark considers six microstructure datasets in two dimensionally consistent settings: two-dimensional models trained on 2D images and supervised three-dimensional models trained on 3D volumes. The comparison therefore addresses 2D-to-2D and 3D-to-3D synthesis rather than dimensional expansion from 2D to 3D. Model performance is assessed using visual comparison, phase statistics, and microstructure-relevant structural descriptors, with effective diffusivity additionally evaluated for the three-dimensional benchmark. Across the considered datasets, all three model families reproduce important aspects of the reference microstructures, although their relative performance depends on the quantity being assessed. Flow matching is comparatively consistent across the reported structural measures and, in three dimensions, reproduces the reference porosity--diffusivity relation most consistently. It also requires substantially fewer sequential sampling evaluations than the 1,000-step DDPM configuration considered here, while WGAN-GP retains the advantage of single-pass generation. Overall, the results identify flow matching as a competitive and computationally attractive approach for porous-microstructure synthesis, providing a favorable quality--cost compromise without universally outperforming the alternative models across every material and metric.