MCDDIM: A Fast Multi-Conditional Diffusion Model for 3D Porous Microstructure Reconstruction

The stochastic reconstruction of 3D porous microstructures is crucial for understanding and designing advanced building materials. While denoising diffusion models have shown promise in generating high-quality reconstructions, their reverse sampling process is notoriously time-consuming. To address this, we propose a multi-conditional denoising diffusion implicit model (MCDDIM). MCDDIM builds upon the denoising diffusion implicit model (DDIM) framework to significantly accelerate generation by employing a non-Markovian reverse process with a reduced number of sampling steps. More importantly, it integrates four key physical property conditions—porosity, pore radius, throat radius, and coordination number—as constraints to guide and control the reconstruction, ensuring accuracy despite the acceleration. We validate MCDDIM on a porous material dataset (FinneyPack) and compare it against state-of-the-art methods. Evaluations across pore space, morphology, connectivity, and permeability demonstrate that MCDDIM achieves reconstruction quality comparable to the much slower MCDDPM, while drastically reducing the generation time. This makes MCDDIM an efficient and effective tool for 3D porous microstructure reconstruction.