Machine Learning-Based Prediction of Shale Adsorption Capacity after Supercritical CO2 Soaking

Abstract Supercritical CO2 (scCO2) soaking can alter the pore structure, mineral composition, and surface properties of shale, thereby affecting its adsorption capacity. Therefore, accurate prediction of shale adsorption capacity after scCO2 soaking is of great significance for evaluating CO2 geological sequestration potential and shale gas recovery performance. In this study, a database of the amount of CO2 adsorbed by shale was collected, which contained 315 data points from 26 shale samples before and after scCO2 soaking. Eight input variables were selected, including adsorption pressure, adsorption temperature, soaking pressure, soaking temperature, soaking time, quartz content, clay content, and total organic carbon (TOC) content, with CO2 adsorption capacity as the output variable. Four machine learning models, namely multilayer perceptron (MLP), support vector regression (SVR), extreme gradient boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM), were constructed and optimized to predict the shale adsorption capacity. The results showed that XGBoost exhibited the best predictive performance among the four models, achieving an R2 of 0.9772 on the test set. To address the insufficient predictive performance of XGBoost under high soaking pressure conditions, an XGBoost–Langmuir model was developed by inversely determining the Langmuir adsorption parameters from XGBoost predictions. Compared with the XGBoost and Langmuir models, the XGBoost-Langmuir model significantly improved prediction accuracy and reduced errors under high soaking pressures. Analyses of SHAP (SHapley Additive exPlanations) values and SHAP interaction values indicated that clay content exhibited asymmetric contributions to model predictions, adsorption temperature generally contributed negatively, and quartz content contributed positively, while the contribution of TOC to model predictions was associated with its interactions with pressure conditions. These results demonstrate that the proposed XGBoost-Langmuir model enhanced the reliability of predicting shale adsorption capacity after scCO2 soaking and revealed the changes in adsorption-controlling factors induced by scCO2 treatment.