Abstract Quantifying fluid–rock interactions within the lithosphere is vital for both geological processes and applications such as storage and geothermal energy development. Mineral replacement reactions generate transient pore networks that enhance fluid flow, yet many pores become isolated once reactions are completed, reducing pore connectivity. The transient nature of reaction‐induced porosity, arising from dynamic dissolution and precipitation processes, makes it challenging to quantify and parametrize the evolution of pore structure. Here, we train a generative model, StyleGAN2‐ADA, on time‐resolved synchrotron computed microtomography (μ‐CT) data of KBr‐KCl replacement to evaluate whether such models can learn pore‐scale evolution and whether the resulting latent representations can be transferred, in a hypothesis‐generating manner, to reconstruct pore connectivity in a natural system lacking continuous observations due to experimental limitations. Our results show that the proposed generative framework can accurately reconstruct the porosity evolution in KBr‐KCl system. Applying the trained model on the natural sample results in controlled increase in pore connectivity and permeability consistent with existing permeability‐depth relations within the crust. The integration of synchrotron X‐ray imaging with generative AI thus provides a novel quantitative framework for systematically editing pore‐space morphology and topology and for investigating structure‐transport relationships in reactive geological systems under experimentally constrained conditions.
Amiri et al. (Mon,) studied this question.