CO2 Mineralization Reaction-Induced Pore-Structure Evolution and Its Effects on CO2 Sweep Behavior in Shale: Stage-Dependent Causal Discovery and Causality-Informed Machine Learning Assisted by LBM Simulations

Abstract CO2 injection in shale oil reservoirs can effectively enhance oil recovery while also facilitating CO2 sequestration. During CO2 mineralization, rock dissolution and the secondary precipitation of supersaturated mineral ions critically govern the evolution of reservoir pore structure and, consequently, CO2 sweep behavior. Therefore, to elucidate the stage-dependent factors controlling CO2 sweep and pore-throat evolution, we combined CO2 mineralization experiments, lattice Boltzmann simulations, causal discovery, and machine-learning prediction.X-ray diffraction (XRD) and confocal laser scanning microscopy (CLSM) were used to characterize shale mineral composition and pore morphology. The extracted pore geometry, mineral attributes, and proportions were incorporated into a previously developed and experimentally validated multi-mineral-informed lattice Boltzmann model to simulate the time-dependent evolution of pore structure and CO2 concentration. A database covering 200,000 LBM time steps was established. Based on characteristic changes in pore-scale parameters, the mineralization process was divided into an early dissolution stage, an intermediate equilibrium stage, and a late precipitation stage. Graphical Lasso and information-geometric causal inference were used to identify stage-specific directed relationships, which were incorporated into a causal-weight transformation for prediction. The cumulative CO2 sweep coefficient increased from 0.006 to 0.326, with increments of 0.098, 0.062, and 0.158 in the three stages, respectively, demonstrating pronounced stage-wise evolution. During the early dissolution stage, maximum pore diameter and surface roughness directly governed CO2 sweep, with normalized weights of 0.641 and 0.359. During the intermediate equilibrium stage, reservoir wettability and reactive mineral fraction became dominant, with weights of 0.502 and 0.498. During the late precipitation stage, wettability, porosity, surface roughness, and reactive mineral fraction jointly controlled sweep behavior, with weights of 0.324, 0.238, 0.220, and 0.219. Reaction-stage partitioning substantially improved prediction, while causal weighting further reduced errors for most models; the MAPE values of KNN-5, second-order polynomial Ridge, random forest, and gradient boosting decreased to 6.71%, 0.62%, 0.35%, and 4.35%, respectively. This work clarifies the stage-dependent causal mechanisms linking mineralization, pore-structure evolution, wettability, and CO2 sweep behavior, providing a physically interpretable framework for predicting CO2-accessible pore ranges during shale oil recovery and geological carbon sequestration.