Comparative Analysis of Parametric and Non-Parametric Machine Learning Models for CO₂ Solubility Estimation

This study investigates the application of parametric and non-parametric supervised machine learning techniques for CO₂ solubility estimation in brine. CO₂ solubility in brine is crucial in estimating the CO₂ storage capacity of geological formations, assessing CO₂ surface mixing, and evaluating the performance of CO₂-enhanced oil recovery projects. Hence this research aims to provide an alternative approach for accurate CO₂ solubility prediction, critical for various applications in CO₂ geo-storage and EOR projects. Independent variables including pressure (0.1-40 MPa), temperature (273-474 K), salinity (0-6 mol/kg), and salt type (NaCl, CaCl₂, and MgCl₂) were sourced from the credible scientific literature. To ensure data quality, a comprehensive data exploration process was conducted. Subsequently, the dataset was then split into training (70%) and testing (30%) sets for model development and evaluation. Hyperparameter tuning was employed to optimize model performance. Statistical metrics and visualizations were also used to evaluate model performance. The model reliability and the statistical validity of the dataset were assessed using William's plot. Sensitivity analysis and feature importance were also explored using correlation and model-intrinsic methods. The findings of this study showed that non-parametric models (random forest (RF) and decision tree (DT)) significantly outperformed the parametric model (multiple linear regression (MLR)). Specifically, RF and DT exhibited exceptional general behaviour and robustness, achieving an overall coefficient of determination>0.97, mean absolute error<0.08 mol/kg, and mean squared error<0.02 mol²/kg² while MLR yielded 0.76, 0.272 mol/kg, and 0.155 mol²/kg², respectively. This disparity is attributed to the inherent limitations of parametric models in capturing complex relationships between CO₂ solubility and its influencing factors. The RF model was adjudged the best-performing model and demonstrated considerable prediction accuracy compared to the Søreide-Whitson with improved binary interaction parameter (m-SW) and the activity-fugacity models during external validation. The paradigm also exhibited superior predictive performance over genetic algorithm-derived correlation and radial basis function neural network proposed by other scholars. The correlation and model intrinsic methods also revealed that pressure exerted the greatest positive influence on CO₂ solubility in aqueous systems, while temperature and salinity showed negative effects. Additionally, pressure and salt type were identified as the most and least influential variables, respectively. William's plot analysis indicated ~2% of the total dataset as vertical suspect and good high-leverage instances. This signifies the authenticity and reliability of the database and constructed models. The investigation uncovers insights into the impact of pressure, temperature, and salinity on CO₂ dissolution. This research represents a significant step forward in understanding and enhancing CO₂ storage optimization strategies while ensuring efficient resource utilization.