Scholarlink

<p>Unlocking the potential of subsurface carbon storage requires cutting-edge machine learning (ML) to predict key petrophysical properties such as wettability and revolutionize CO<sub>2</sub> sequestration. This study evaluated five ML models including extreme gradient boosting (XGBoost), support vector regression (SVR), multilayer perceptron (MLP), adaptive boosting (AdaBoost), and random forest (RF) to predict CO<sub>2</sub> wettability under varied geological conditions. The XGBoost model, enhanced with particle swarm optimization (PSO), outperformed the other ML models in terms of prediction accuracy. The training results yielded a coefficient of correlation (R<sup>2</sup>) of 0.9968, a mean absolute error (MAE) of 0.0084, and a root mean square error (RMSE) of 0.0270. The testing results were R<sup>2</sup> = 0.9312, MAE = 0.0284, and RMSE = 0.0412. The ML models’ predicted performance follows this sequence: PSO-XGBoost &gt; RF &gt; AdaBoost &gt; MLP &gt; SVR. A feature importance analysis using Shapley additive explanations (SHAP) identified pressure, mineral composition and salinity as the major controlling factors of CO<sub>2</sub> wettability. These findings provide insights into optimizing CO<sub>2</sub> sequestration strategies and ML methodology for geological applications, particularly in subsurface engineering, including generalizability across diverse geological conditions and computational efficiency for rapid scenario testing.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Improving CO2 Sequestration Through Machine-Learning-Driven Prediction of Wettability in Tight Reservoirs

  • 期刊论文

  • 发表日期：

    2026年3月5日

查看全文
  • Elieneza Nicodemus Abelly,
  • Feng Yang,
  • Edwin Ernest Nyakilla,
  • Fravian Mwinzarubi Marobo,
  • Christopher Nyangi Mkono,
  • Omari Athumani Mbuji,
  • Erasto E. Kasala,
  • Allou Koffi Franck Kouassi

摘要

Unlocking the potential of subsurface carbon storage requires cutting-edge machine learning (ML) to predict key petrophysical properties such as wettability and revolutionize CO2 sequestration. This study evaluated five ML models including extreme gradient boosting (XGBoost), support vector regression (SVR), multilayer perceptron (MLP), adaptive boosting (AdaBoost), and random forest (RF) to predict CO2 wettability under varied geological conditions. The XGBoost model, enhanced with particle swarm optimization (PSO), outperformed the other ML models in terms of prediction accuracy. The training results yielded a coefficient of correlation (R2) of 0.9968, a mean absolute error (MAE) of 0.0084, and a root mean square error (RMSE) of 0.0270. The testing results were R2 = 0.9312, MAE = 0.0284, and RMSE = 0.0412. The ML models’ predicted performance follows this sequence: PSO-XGBoost > RF > AdaBoost > MLP > SVR. A feature importance analysis using Shapley additive explanations (SHAP) identified pressure, mineral composition and salinity as the major controlling factors of CO2 wettability. These findings provide insights into optimizing CO2 sequestration strategies and ML methodology for geological applications, particularly in subsurface engineering, including generalizability across diverse geological conditions and computational efficiency for rapid scenario testing.