Development of Advanced Machine Learning Models for Predicting CO2 Solubility in Brine
This study explores the application of advanced machine learning (ML) models to predict CO2 solubility in NaCl brine, a critical parameter for effective carbon capture, utilization, and storage (CCUS). Using a comprehensive database of 1404 experimental data points spanning temperature (−10 to 450 °C), pressure (0.098 to 140 MPa), and salinity (0.017 to 6.5 mol/kg), the research evaluates the predictive capabilities of five ML algorithms: Decision Tree, Random Forest, XGBoost, Multilayer Perceptron, and Support Vector Regression with a radial basis function kernel. Among these, XGBoost demonstrated the highest overall accuracy, achieving an R2 value of 0.9926, with low root mean square error (RMSE) and mean absolute error (MAE) of 0.0655 and 0.0191, respectively. A feature importance analysis revealed that pressure has the most impactful effect and positively correlates with CO2 solubility, while temperature generally exhibits a negative effect. A higher accuracy was found when the developed model was compared with one well-established empirical model and one ML-based model from the literature. The results underscore the potential of ML models to significantly enhance prediction accuracy over a wide data range, reduce computational costs, and improve the efficiency of CCUS operations. This work demonstrates the robustness and adaptability of ML approaches for modeling complex subsurface conditions, paving the way for optimized carbon sequestration strategies.