Prediction of Pure Mineral-H2-Brine Wettability Using Data-Driven Machine Learning Modeling: Implications for H2 Geo-Storage

Abstract Greenhouse gases, particularly carbon dioxide (CO2), have the effect of raising the Earth's temperature. To combat this issue and reduce carbon emissions, it is advisable to shift towards the widespread utilization of cleaner fuels, such as hydrogen. The establishment of a global-scale hydrogen economy, coupled with hydrogen geological storage, presents a viable solution to meet the world's energy demands while accommodating peak usage periods. In geological hydrogen (H2) storage, the rock formation wetting characteristics are essential to regulate fluid dynamics, injection rates, the spread of gas within the rock matrix, and safety considerations. The wetting characteristics of minerals within the rock are significantly influenced by geological factors. To assess the wetting behavior of a mineral/H2/brine system under geo-storage conditions, innovative approaches have emerged. This research utilized a combination of advanced machine learning models, such as fully connected neural networks, adaptive gradient boosting, random forests, decision trees, and extreme gradient boosting to forecast the wettability characteristics of a ternary system comprising hydrogen (H2), brine, and specific rock minerals (namely quartz and mica). The predictions were made under various conditions, including different pressures ranging from 0 to 25 MPa, temperatures spanning from 308 to 343 K, and salinities of 10 wt.% NaCl solution. The machine learning models demonstrated remarkable accuracy in predicting mineral/H2/brine system's wettability (contact angles, advancing and receding). Incorporation of various experimental values have established correlations based on ML techniques. The performance and reliability of these models were rigorously assessed using statistical methods and graphical analyses. The deployed ML models consistently provided accurate predictions of wettability across diverse operational scenarios. Notably, the suggested model exhibited a root mean square error (RMSE) of 0.214 during training and 0.810 during testing. Furthermore, sensitivity analysis revealed that pressure exerted the most significant influence on mineral/H2/brine system's wettability. These ML model outcomes can be effectively utilized to anticipate hydrogen geological storage capacities and ensure the security of restraint in large-scale developments.