Machine learning prediction of methane, ethane, and propane solubility in pure water and electrolyte solutions: Implications for stray gas migration modeling

Abstract

Hydraulic fracturing is an effective technology for hydrocarbon extraction from unconventional shale and tight gas reservoirs. A potential risk of hydraulic fracturing is the upward migration of stray gas from the deep subsurface to shallow aquifers. The stray gas can dissolve in groundwater leading to chemical and biological reactions, which could negatively affect groundwater quality and contribute to atmospheric emissions. The knowledge of light hydrocarbon solubility in the aqueous environment is essential for the numerical modelling of flow and transport in the subsurface. Herein, we compiled a database containing 2129 experimental data of methane, ethane, and propane solubility in pure water and various electrolyte solutions over wide ranges of operating temperature and pressure. Two machine learning algorithms, namely regression tree (RT) and boosted regression tree (BRT) tuned with a Bayesian optimization algorithm (BO) were employed to determine the solubility of gases. The predictions were compared with the experimental data as well as four well-established thermodynamic models. Our analysis shows that the BRT-BO is sufficiently accurate, and the predicted values agree well with those obtained from the thermodynamic models. The coefficient of determination (R²) between experimental and predicted values is 0.99 and the mean squared error (MSE) is 9.97 × 10⁻⁸. The leverage statistical approach further confirmed the validity of the model developed.