Pejman Tahmasebi
Generating porous media structures that honor target physical properties is a fundamental challenge. Available methods are either computationally expensive or they often do not account for physically relevant statistical properties when porous media are generated. In this paper, a new framework is proposed that combines generative machine learning models with ensemble-based assimilation to enable property-constrained porous media generation. A machine learning model is first used to map binary pore structures into a lower-dimensional space, where optimization is performed to minimize the mismatch between simulated and target properties such as porosity, permeability, tortuosity, specific surface area, and characteristic length scales. Instead of directly matching pore structures, the proposed approach steers the generative process through iterative ensemble updates in lower-dimension space, guided by forward-modeled features derived from generated samples. The methodology accommodates multiple uncertain and noisy targets and balances them using feature-specific weighting and trust levels. The performance of this workflow is demonstrated using reliably produces pore-scale realizations that are consistent with given physical constraints in two examples where a wide range of target properties and a specific value are of interest to examine how the proposed method can reproduce the uncertainty space in these two distinct situations. The results indicate that in both cases the proposed method can reproduce the target properties in a reasonable time when all computations are conducted in a representative latent dimension, accompanied by rigorous statistical optimization.