An explicit machine learning model for brine-gas interfacial tension prediction: Implications for H₂, CH₄, and CO₂ geo-storage

Brine-gas interfacial tension (γ) is a key parameter to determine pore-scale fluid distributions/dynamics and thus influence reservoir-scale gas geo-storage (GGS) efficiencies. Note that γ at GGS conditions is a complex function of gas mole fraction (x_H2, x_CH4, and x_CO2), ionic strength (I), temperature (T), and pressure (P) – which is time-consuming to measure experimentally and challenging to quantify theoretically. Therefore, innovatively this work integrates the physics-constrained generative adversarial network and the genetic algorithm-based symbolic regression (PCGAN-GA-SR) to establish an explicit correlation for predicting γ. The proposed PCGAN-GA-SR model outperformed models without synthetic data or physical constraints (R²=0.89 vs. <0.75). The proposed correlation also outperformed existing correlations by considering more parameters (6 vs. 3), reducing prediction errors, consisting of less terms (7 vs. > 10), and covering broader systems. Sensitivity analyses indicate that P is the most dominant factor affecting γ. The correlation predicts that γ increases with higher x_H2 and x_CH4, but decreases with higher x_CO2. These results are qualitatively consistent with experimental data and physical knowledge. This work provides an efficient and reliable tool that can be utilized for evaluations of GGS capacity and security. These insights contribute to the successful implementation of large-scale GGS projects, thus promoting the decarbonization and energy transition.