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

Summary

This method paper introduces DimExDAM, a diffusion–GAN hybrid that synthesizes 3D porous-media microstructures — including heterogeneous Boom Clay, illite, sandstone, and Savonnières carbonate — from as few as three orthogonal 2D training images. Building on Denoising Diffusion GANs, the authors replace the 2D generator with a fully 3D convolutional generator that predicts clean volumes over a short (T=8) denoising chain, and replace the standard denoising MSE loss with an adversarial loss computed on stratified 2D slices, using a single time- and plane-conditioned 2D discriminator instead of SliceGAN's three critics. Six stabilization strategies (spectral normalization, Gaussian-blur curriculum, temperature-scaled tanh, logit/gradient clipping, lazy R1, TTUR) are presented as a coupled recipe. Across five datasets trained with one fixed hyperparameter set, DimExDAM reportedly matches ground-truth porosity, phase fractions, and directional descriptors (chord length, lineal path, two-point correlation, pore size) more closely than SliceGAN, which shows checkerboard artifacts and porosity shifts on 3 of 5 materials; lattice-Boltzmann gas diffusivity and permeability computed on 150 generated volumes per method largely fall within reference ranges. The main acknowledged limitation is reduced between-sample variability: generated volumes cluster tightly around training porosity, which the authors attribute to sparse 2D observations and the generative prior. Remaining weaknesses include the absence of isolated component ablations, a potentially under-tuned SliceGAN comparison, no diffusion-based 2D-to-3D baselines, and no statistical testing.