Abstract: [Objective] During CO₂ fluid injection into oil reservoirs or saline aquifers, CO₂-water-rock interactions can alter porous media properties, thereby influencing CO₂ microscale sweep efficiency, primarily due to capillary effects. Laboratory experiments and micro/nanoscale numerical simulations often struggle to isolate the specific contributions of individual pore properties, limiting targeted injection optimization for maximizing geological storage potential. [Methods] To investigate the dynamic evolution of pore structures and properties during multi-mineral competitive dissolution-precipitation reactions under CO₂ injection, by developing a lattice Boltzmann method (LBM) to simulate CO₂-water-rock interactions in shale oil reservoirs and to analyze pore properties (e.g., average wettability, roughness, porosity) and CO₂ microscale sweep efficiency. The LBM simulations generated a dataset covering various pore property scenarios to support causal machine learning. Using a double machine learning framework with a random forest algorithm, a causal inference prediction model was built for CO₂ microscale sweep efficiency, treating reaction time as a continuous treatment variable. [Results] This model quantified the relative importance of key pore parameters—porosity, wettability, and mean pore diameter—on sweep efficiency within the pore network. Results indicate that reservoirs with higher proportions of carbonate minerals (calcite) exhibit greater CO₂ microscopic sweep efficiency. The CO₂-water-rock reaction triggers calcite dissolution, forming preferential flow paths, while the secondary precipitation of oil-wet calcite induces localized wettability alteration. This dynamic "dissolution–secondary precipitation" process modifies capillary forces by altering pore-throat structure and physical properties, thereby influencing the microscopic sweep range of CO₂ fluids. However, under identical mineral proportions, CO₂ sweep efficiency varies among samples, with higher calcite proportions correlating with broader variation in sweep performance. [Conclusions] These findings underscore the crucial role of pore physical properties, beyond mineral composition alone, in governing sweep efficiency. Causal learning identified key pore-throat parameters that control CO₂ microscale sweep behavior, with wettability emerging as the most influential factor. Neutrally water-wet pore-throats exhibited the highest CO₂ sweep efficiency. [Significance] By constructing a Lattice Boltzmann model for CO₂-water-rock interactions and quantifying the impact of key physical parameters, this study provides a reference and guidance for the targeted adjustment of CO₂ injection strategies and the enhancement of geological CO₂ storage effectiveness.