Counterfactual Machine-Learning Analysis of Carbon Dioxide Rich Gas Brine Interfacial Tension Under Subsurface Conditions

Abstract Interfacial tension (IFT) between CO2-rich gas and brine influences capillary trapping, plume migration and containment during geological carbon storage. A harmonized literature database was filtered to mixtures containing at least 0.75 mole fraction CO2, yielding 5,215 observations from 379 experimental groups. A deterministic group-aware partition prevented rows from the same group entering both model development and testing. A single Extra Trees surrogate was fitted over eleven directly settable predictors - pressure, temperature, gas composition and mass-based brine descriptors - so that every counterfactual constraint can be stated in physical units. It achieved grouped-CV RMSE 3.61 ± 0.32 mN/m and held-out RMSE 2.87 mN/m (R2=0.932), against 8.04 mN/m for ridge regression on the same predictors. Counterfactuals were generated in physical units with gas closure, salinity coupling, observed-range bounds and an applicability-domain gate. 20 of 24 retained recommendations reached the configured target by ensemble mean, and 18 also met the tree-agreement threshold. These outputs are screening hypotheses rather than causal prescriptions.