Diffusion Model-Based Generation of Three-Dimensional Multiphase Pore-Scale Images

We propose a diffusion model-based machine learning method for generating three-dimensional images of both the pore space of rocks and the fluid phases within it. This approach overcomes the limitations of current methods, which are restricted to generating only the pore space. Our reconstructed images accurately reproduce multiphase fluid pore-scale details in water-wet Bentheimer sandstone, matching experimental images in terms of two-point correlation, porosity, and fluid flow parameters. This method outperforms generative adversarial networks with a broader and more accurate parameter range. By enabling the generation of multiphase fluid pore-scale images of any size subject to computational constraints, this machine learning technique provides researchers with a powerful tool to understand fluid distribution and movement in porous materials without the need for costly experiments or complex simulations. This approach has wide-ranging potential applications, including carbon dioxide and underground hydrogen storage, the design of electrolyzers, and fuel cells.