Data-efficient generation of pore-scale microstructures for rock-on-chip design.

Abstract

We present a unified digital-to-experimental workflow that advances the integration of single-image diffusion-based rock generation with connectivity conditioning, scalable texture synthesis, and microfluidic experimentation, enabling statistically realistic generated images to be converted into hydraulically functional, fabrication-ready porous media. For each reference rock image across multiple lithologies, a separate SinDiffusion model is trained to generate statistically consistent pore-scale realizations that preserve key features of the input structure. Quantitative evaluation using intensity statistics, porosity, Minkowski functionals, connectivity metrics, and pore-shape eccentricity confirms preservation of multi-scale morphological characteristics. The generated images are further processed through percolation-constrained thresholding and texture synthesis to produce fabrication-ready designs with controlled connectivity and arbitrary geometries. Additional geometric and flow-property analyses across the four workflow stages show that the final designs retain comparable structural and transport-relevant characteristics while satisfying microfluidic fabrication requirements. The digital layouts are translated into rock-on-chip microfluidics <i>via</i> maskless photolithography and used for CO₂ drying and salt precipitation experiments. Homogeneous and artificially fractured sandstone-like configurations demonstrate the workflow's ability to resolve structure-dependent drying and precipitation patterns, with fractures promoting localized deposition and delayed clogging. Experiments on multiple realizations generated from the same input image show a consistent qualitative sequence of CO₂ displacement, brine depletion, and salt accumulation, while capturing realization-specific variability in phase evolution.