Machine Learning and Deep Learning Framework for Accurate Prediction of CH4 Solubility in Brine Systems Using Physicochemical Descriptors

Abstract

Methane, with a higher global warming potential than carbon dioxide, is a key driver of climate change, making accurate prediction of its solubility in water and brine essential due to its impact on reservoir behavior, gas migration, hydrate formation, corrosion, flow assurance, and environmental risks such as leakage. However, classical thermodynamic models perform poorly in complex electrolyte systems, while machine learning approaches often require multiple salt-specific inputs, limiting their flexibility and requiring model redesign or retraining when brine composition changes. This study proposes a data-driven framework using a large experimental dataset and physicochemical descriptors, including temperature, pressure, ionic strength (IS), effective free water fraction (EFW), and ion-specific descriptor (ISD), to capture realistic gas–brine interactions. Four machine learning and deep learning models (CatBoost, AdaBoost-DT, GrowNet, and TabNet) were developed and compared, along with benchmark thermodynamic models (CPA-MHV1 and SRK-MHV1). Among the investigated approaches, CatBoost demonstrated the best predictive capability and generalization performance, achieving excellent accuracy on the independent test dataset with MAE × 100 = 0.0089, RMSE = 0.0001, and R2 = 0.9972. SHAP analysis revealed that pressure is the most influential parameter affecting methane solubility, followed by temperature and electrolyte-related descriptors, where IS, ISD, and EFW successfully represent salting-out effects and ion–solvent interactions. Williams’ applicability domain analysis further confirmed the reliability of the developed model, showing that 97.34% of the data points were located within the model applicability domain, while 1.17% were identified as high-leverage points and 1.49% as suspected outliers. Overall, the proposed framework provides an accurate, fast, and interpretable tool for CH4 solubility prediction in brine systems.