Scholarlink

<p>Nuclear Magnetic Resonance (NMR) logs provide critical insights into pore-scale fluid distribution, including clay-bound water (CBW), free fluid index (FFI), and bulk volume irreducible (BVI), but their high acquisition cost limits widespread availability. This study presents an interpretable machine learning (ML) workflow for predicting these NMR-derived parameters from conventional well logs, including gamma ray (GR), bulk density (RHOB), neutron porosity (APLC), photoelectric factor (PEF), and resistivity (RLA5). Data from three wells (WEB-1, PAPYRUS-1, and BAMBOO-1) in the West El-Burullus gas field, offshore the Nile Delta, Egypt, are used to develop and evaluate two supervised learning models: Extreme Gradient Boosting (XGB) and Multilayer Perceptron (MLP). Model development incorporated rigorous data preprocessing, systematic hyperparameter optimization, and SHAP-based interpretability analysis to ensure physically meaningful predictions. Results demonstrate that the XGB model consistently outperformed the MLP model, achieving testing R² values of 0.885, 0.895, and 0.846 for CBW, FFI, and BVI, respectively. Blind-well validation further confirmed model generalization, with XGB achieving R² values of 0.811 for CBW, 0.825 for FFI, and 0.774 for BVI in the PAPYRUS-1 well. In the BAMBOO-1 well, where NMR measurements were unavailable, the predicted profiles reproduced geologically consistent stratigraphic trends. SHAP analysis revealed that model predictions were controlled by physically plausible relationships between conventional logs and reservoir properties, enhancing confidence in the workflow. The proposed workflow demonstrates that ensemble-based approaches such as XGB can provide reliable, interpretable, and computationally efficient predictions of advanced petrophysical properties from conventional logs. Although the results are encouraging, the limited dataset size constrains broader generalization. Future work should focus on expanding the database, improving BVI prediction through advanced feature engineering, and validating the workflow across a wider range of geological settings.</p> Graphical Abstract <p></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

Robust prediction of advanced reservoir petrophysical parameters using conventional logs: Insights from XGB and neural network models

  • 期刊论文

  • 发表日期：

    2026年7月13日

查看全文
  • Ahmed Bakry,
  • Ashraf Ghoneimi,
  • Muhammad Nabih,
  • Qamar Yasin,
  • Mohamed Abd Elaziz,
  • Ahmed Elgazar

摘要

Nuclear Magnetic Resonance (NMR) logs provide critical insights into pore-scale fluid distribution, including clay-bound water (CBW), free fluid index (FFI), and bulk volume irreducible (BVI), but their high acquisition cost limits widespread availability. This study presents an interpretable machine learning (ML) workflow for predicting these NMR-derived parameters from conventional well logs, including gamma ray (GR), bulk density (RHOB), neutron porosity (APLC), photoelectric factor (PEF), and resistivity (RLA5). Data from three wells (WEB-1, PAPYRUS-1, and BAMBOO-1) in the West El-Burullus gas field, offshore the Nile Delta, Egypt, are used to develop and evaluate two supervised learning models: Extreme Gradient Boosting (XGB) and Multilayer Perceptron (MLP). Model development incorporated rigorous data preprocessing, systematic hyperparameter optimization, and SHAP-based interpretability analysis to ensure physically meaningful predictions. Results demonstrate that the XGB model consistently outperformed the MLP model, achieving testing R² values of 0.885, 0.895, and 0.846 for CBW, FFI, and BVI, respectively. Blind-well validation further confirmed model generalization, with XGB achieving R² values of 0.811 for CBW, 0.825 for FFI, and 0.774 for BVI in the PAPYRUS-1 well. In the BAMBOO-1 well, where NMR measurements were unavailable, the predicted profiles reproduced geologically consistent stratigraphic trends. SHAP analysis revealed that model predictions were controlled by physically plausible relationships between conventional logs and reservoir properties, enhancing confidence in the workflow. The proposed workflow demonstrates that ensemble-based approaches such as XGB can provide reliable, interpretable, and computationally efficient predictions of advanced petrophysical properties from conventional logs. Although the results are encouraging, the limited dataset size constrains broader generalization. Future work should focus on expanding the database, improving BVI prediction through advanced feature engineering, and validating the workflow across a wider range of geological settings.

Graphical Abstract