Abstract

Predicting CO₂ solubility in aqueous systems is vital for advancing carbon capture, utilization, and storage (CCUS) technologies and for enhanced oil recovery and geologic modeling. This study addresses the limitations of classical thermodynamic models in handling complex electrolyte systems by introducing a machine learning (ML) - based framework for accurate and efficient CO₂ solubility prediction.  This study demonstrates that classical EOS models are often computationally intensive and require detailed compositional input and interaction parameters as they are built on thermodynamic principles involving activity coefficients, chemical potential, fugacity, and empirically derived interaction parameters. Classical EOS models were implemented in this study with CO₂ predictions.  Two thousand seven hundred thirty experimental data points were compiled from literature covering various temperatures, pressures, and salinities ranges. Ionic strength was calculated using the Debye-Hückel method. The dataset includes multiple salts such as NaCl, CaCl₂, KCl, NaHCO₃, K₂CO₃, and MgSO₄.

Three ML models were developed: Random Forest (RF), Gradient Boosting (GB), and an Ensemble model combining both. The data underwent preprocessing steps like standardization, polynomial feature expansion, and outlier removal. The Ensemble model delivered the highest predictive accuracy, achieving R² values of 0.9916 (training), 0.9832 (validation), and 0.9934 (testing) while maintaining robustness and avoiding overfitting. Feature sensitivity analysis revealed pressure as the most influential variable (~45%), followed by ionic strength and temperature. The model captured key solubility trends, such as increased solubility with pressure and a salting-out effect with higher ionic strength. Notably, K₂CO₃ showed the highest CO₂ uptake at 353 K. The study also developed a graphical user interface for real-time CO₂ solubility estimation in CCUS applications. Overall, this work demonstrates the potential of ML to provide a reliable, scalable, user-friendly, and computationally efficient alternative to traditional solubility models, especially in complex aqueous systems. Future work could focus on expanding the dataset under extreme conditions and integrating more advanced ion interaction models for enhanced predictive power.