Predicting the interfacial tension of CO₂ and NaCl aqueous solution with machine learning

Achieving carbon neutrality requires effective strategies to reduce CO₂ emissions, and geological sequestration of CO₂ is considered among the most promising and economically viable options. The interfacial tension (IFT) between the CO₂ and the surrounding liquid (underground salt water or brine, NaCl) is a key parameter that affects the storage capacity of CO₂ in saline aquifers; however, the experimental measurement of IFT is often time-consuming, labor-intensive, and reliant on expensive equipment, and empirical correlations demonstrate a low level of accuracy. Machine learning (ML) techniques have been suggested as an alternative approach, and the current literature related to interfacial phenomena utilizes a wide array of basic and advanced ML models for predicting IFT, though often without a comparative analysis, raising the question of which model is most appropriate for this specific application. In this work, multiple machine learning models, including linear regression (LR), support vector machine (SVM), decision tree regressor (DTR), random forest regressor (RFR), and multilayer perceptron (MLP), are used to predict the IFT of the CO₂ and aqueous solution of NaCl. Models are trained using an experimental dataset that covers a wide range of temperature, pressure, and salinity (NaCl) conditions for CO₂-brine IFT. Hyperparameter tuning algorithms are utilized to optimize each model, and the performance is evaluated using metrics such as mean absolute error (MAE) and mean absolute percentage error (MAPE). The best-performing algorithms are found to be SVM and MLP, with a MAPE of 0.97% and 0.99% and a MAE of 0.39 mN/m and 0.40 mN/m, respectively. The linear regression model demonstrated the worst performance with a MAPE of 4.25% and an MAE of 1.7 mN/m. The feature importance analysis reveals that pressure is the main parameter affecting the IFT. Our findings indicate a notable enhancement in prediction accuracy over previous ML studies in this area. Moreover, the results from this study suggest that even the basic ML models that were investigated, when properly tuned and optimized, are sufficient for accurate IFT predictions. This demonstrates that ML models offer a cost-effective and efficient alternative to experimental methods, potentially optimizing designs for CO₂ sequestration.