Machine learning- based shale wettability prediction: Implications for H2, CH4 and CO2 geo-storage

Shale wettability determines shale gas productivities and gas (H2, CH4 and CO2) geo-storage efficiencies. However, shale wettability is a complex parameter which depends on multiple influencing factors, thus very time-consuming and costly to measure experimentally. Herein, we combined the eXtreme gradient boosting (XGBoost) and Shapley additive explanation (SHAP) machine learning methods to accurately predict brine advancing (theta A) and receding (theta R) contact angles and estimate shale wettability. The XGBoost model demon-strated much higher prediction accuracies than the commonly-used multiple linear regression and partial least squares regression models, e.g., R2 was 0.946-0.999, 0.794-0.821, and 0.635-0.674, respectively for these three models. The SHAP sensitivity analyses showed that total organic carbon content and gas molecular weight (MG) were the two most significant factors influencing shale wettability. In addition, shale hydrophobicity positively correlated with MG, calcite content, pressure and brine ionic strength, while negatively correlated with tem-perature and quartz content. This work provides an efficient approach for shale wettability estimation, thus aiding in the implementation of improved gas recovery and gas geo-storage processes, to further guarantee energy security and mitigate climate change.