Scholarlink

<p>Porosity is a fundamental petrophysical parameter that governs fluid storage capacity and reservoir productivity. However, accurately predicting porosity in heterogeneous reservoirs remains a significant challenge, as conventional methods are often limited by high costs, low efficiency, and reduced accuracy. To overcome these limitations, this study proposes an innovative lithology-guided machine and deep learning (ML/DL) framework that integrates multiple predictive models, including Recurrent Neural Networks (RNN), Artificial Neural Networks (ANN), and Extreme Gradient Boosting (XGBoost), along with advanced uncertainty quantification. The RNN model is specifically employed to capture depth-dependent continuity within well-log data, enabling the framework to model geological variations with higher precision. Meanwhile, ANN and XGBoost provide complementary modeling for nonlinear relationships, enhancing the overall robustness of the predictions. Furthermore, Monte Carlo dropout and bootstrap resampling are incorporated to quantify prediction uncertainty, providing a more comprehensive understanding of model reliability. The methodology is applied to well-log data from the Longmaxi Shale Formation in the Jiaoye Wells of the Sichuan Basin, where lithology-sensitive modeling is crucial to differentiate the distinct porosity characteristics of shale and sandstone intervals. Among the evaluated models, the RNN consistently outperforms the others, achieving the lowest RMSE (0.024) and the highest R² (0.981) on the test dataset. In contrast, XGBoost (RMSE = 0.051, R² = 0.912) and ANN (RMSE = 0.053, R² = 0.905) exhibit relatively lower predictive accuracy. These results affirm that the RNN effectively captures nonlinear geological relationships and depth-dependent variations, establishing a more reliable and precise framework for porosity prediction. The findings highlight the substantial potential of deep learning techniques to enhance reservoir characterization and support informed decision-making in shale reservoirs.</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

Lithology-driven machine and deep learning modelling for porosity prediction in shale reservoir systems using well logging data

  • 期刊论文

  • 发表日期：

    2026年5月30日

查看全文
  • Shaukat Khan,
  • Zhishui Liu,
  • Zhiqiang Lu,
  • Wakeel Hussain,
  • Muhammad Sajid,
  • Manaf Muhammad,
  • Abdul Wahid,
  • Muhammad Ubaid Umar

摘要

Porosity is a fundamental petrophysical parameter that governs fluid storage capacity and reservoir productivity. However, accurately predicting porosity in heterogeneous reservoirs remains a significant challenge, as conventional methods are often limited by high costs, low efficiency, and reduced accuracy. To overcome these limitations, this study proposes an innovative lithology-guided machine and deep learning (ML/DL) framework that integrates multiple predictive models, including Recurrent Neural Networks (RNN), Artificial Neural Networks (ANN), and Extreme Gradient Boosting (XGBoost), along with advanced uncertainty quantification. The RNN model is specifically employed to capture depth-dependent continuity within well-log data, enabling the framework to model geological variations with higher precision. Meanwhile, ANN and XGBoost provide complementary modeling for nonlinear relationships, enhancing the overall robustness of the predictions. Furthermore, Monte Carlo dropout and bootstrap resampling are incorporated to quantify prediction uncertainty, providing a more comprehensive understanding of model reliability. The methodology is applied to well-log data from the Longmaxi Shale Formation in the Jiaoye Wells of the Sichuan Basin, where lithology-sensitive modeling is crucial to differentiate the distinct porosity characteristics of shale and sandstone intervals. Among the evaluated models, the RNN consistently outperforms the others, achieving the lowest RMSE (0.024) and the highest R² (0.981) on the test dataset. In contrast, XGBoost (RMSE = 0.051, R² = 0.912) and ANN (RMSE = 0.053, R² = 0.905) exhibit relatively lower predictive accuracy. These results affirm that the RNN effectively captures nonlinear geological relationships and depth-dependent variations, establishing a more reliable and precise framework for porosity prediction. The findings highlight the substantial potential of deep learning techniques to enhance reservoir characterization and support informed decision-making in shale reservoirs.