Reaction-Induced Porosity During Fluid-Mineral Interaction: From 4D Synchrotron Imaging to AI-Driven Permeability Reconstruction

Fluid-rock interactions drive critical lithospheric processes and industrial applications including CO₂ storage and geothermal energy extraction. In deep crystalline crust where static permeability is negligible and rocks do not deform, fluids primarily exploit transient pathways created through chemical reactions with minerals in disequilibrium. These reaction-induced pore networks dynamically alter rock permeability, yet their ephemeral nature makes direct characterization challenging.

We present an integrated methodology combining time-resolved synchrotron x-ray microtomographic imaging (4DSµCT) with generative artificial intelligence to quantify reaction-induced porosity evolution. Using 4DSµCT, we captured spatio-temporal pore network dynamics during KBr-KCl replacement, a well-established analogue for interface-coupled dissolution-precipitation processes. Advanced statistical microstructural descriptors and Minkowski functionals revealed intricate coupling between dissolution-precipitation mechanisms, transport regimes, and evolving connectivity governing transient permeability.

To extend insights beyond experimental limitations, particularly for high-temperature systems (>500°C) where direct imaging remains infeasible, we developed Pore-Edit GAN, a StyleGAN2-ADA framework trained on ~29,000 tomographic images. This model generates statistically realistic microstructures while enabling semantic editing of porosity and connectivity. We applied our approach to hydrothermally altered monzonite from the Oslo Rift, where feldspar replacement reactions at ~10 km depth created now-isolated pore networks. By navigating the GAN latent space along learned connectivity directions, we reconstructed plausible transient pore configurations, effectively reversing the porosity isolation that occurred as reactions ceased.

Voxel-based finite element simulations of incompressible Stokes flow through these AI-reconstructed networks yield permeabilities reaching 4.5×10⁻¹⁵ m², a two-order-of-magnitude enhancement upon pore reconnection, consistent with established transient crustal permeability-depth relations. This convergence of synchrotron capabilities, deep generative models, and computational fluid dynamics establishes a quantitative framework for predicting transport properties in reactive geological systems where direct observation remains challenging.