Predictive Analysis of CO2 and CH4 Sorption Mechanisms in Unconventional Reservoirs

This study presents a robust machine learning framework for accurately predicting the sorption capacities of two greenhouse gases, CH₄ and CO₂, in unconventional reservoirs under diverse thermodynamic conditions. A comprehensive set of ensemble and hybrid modelsincluding ExtraTrees, XGBoost, LightGBM, Gradient Boosting, CatBoost, and four newly developed hybrid architectures (HM1–HM4) was trained and evaluated using an extensive experimental dataset. The hybrid models demonstrated consistently superior performance, with the best CH₄ model (HM3) achieving an R² of 0.9905 and the best CO₂ model (HM3) reaching an R² of 0.9949 on unseen test data. Unlike previous studies, where only selected models performed well for either CH₄ or CO₂, the proposed models exhibited high accuracy and generalizability across both gas types. Model interpretability was enhanced through SHAP and Partial Dependence Plot (PDP) analyses. For CH₄ sorption, total organic carbon (TOC) and pressure were identified as the most influential features, while for CO₂ sorption, Gas (%) representing CO₂ concentration was the dominant factor, followed by temperature and pressure. PDP analysis further revealed a strong linear relationship between CO₂ sorption and Gas (%), with TOC contributing significantly at lower levels before plateauing. Moisture content was found to have a mild negative effect on sorption in both cases. These results confirm the physical consistency of the models and their capacity to capture complex sorption behavior, offering valuable tools for gas-in-place estimation, CO₂ storage evaluation, and production forecasting in unconventional reservoirs.