Scholarlink

<p>Acoustic impedance (<i>Z</i>) is a fundamental parameter in geophysical subsurface characterization, governing seismic reflection coefficients and enabling reservoir property quantification through seismic inversion. Conventional derivation of <i>Z</i> relies on density (<i>ρ</i>) and P-wave velocity (<i>V</i><sub><i>p</i></sub>) logs, yet these datasets are frequently unavailable due to operational constraints, tool limitations, or borehole irregularities. Existing empirical methods, such as neutron porosity-based formulas, suffer from restrictive assumptions -including matrix/fluid constant dependencies, low shale tolerance (&lt; 25%), and negligible secondary porosity - that limit applicability in heterogeneous formations. To overcome these challenges, we present a robust machine learning workflow that predicts Z directly from commonly available well logs, circumventing the need for sonic or density data. A multi-well dataset comprising gamma-ray (<i>GR</i>), neutron porosity (<i>NPHI</i>), deep resistivity (<i>R</i><sub><i>D</i></sub>), and formation tops were analyzed. Pearson correlation identified <i>GR</i>, <i>NPHI</i>, and log-transformed resistivity (<i>R</i><sub><i>Dlog</i></sub>) as optimal predictors. Data preprocessing included Isolation Forest-based outlier removal and logarithmic resistivity transformation. The XGBoost regressor - selected for its scalability in handling nonlinear interactions - was trained on 80% of the data, with hyperparameters optimized via cross-validated grid search. Model performance was evaluated using mean absolute error (<i>MAE</i>), root MSE (<i>RMSE</i>), and coefficient of determination (<i>R²</i>). The optimized model achieved an R² of 0.916 (training) and 0.808 (testing), with <i>RMSE</i> values of 718.3 and 1070, respectively. Independent validation on a blind well demonstrated strong generalization (<i>R²</i> = 0.869, <i>RMSE</i> = 981.3), with predicted <i>Z</i> logs showing stratigraphic fidelity and suppression of high-amplitude artifacts inherent to sonic-derived impedance. Compared to empirical methods, the ML workflow eliminates reliance on matrix/fluid constants, accommodates shale volumes &gt; 25%, and mitigates errors from secondary porosity or gas effects. This provides a scalable, cost-effective solution to enhance seismic inversion accuracy in data-scarce or complex lithological settings.</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

XGBoost regression for robust acoustic impedance prediction in the absence of density and sonic logs

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月18日

查看全文
  • Khaled Saleh,
  • Muhammad A. El Hameedy,
  • Walid M. Mabrouk,
  • Ahmed Metwally

摘要

Acoustic impedance (Z) is a fundamental parameter in geophysical subsurface characterization, governing seismic reflection coefficients and enabling reservoir property quantification through seismic inversion. Conventional derivation of Z relies on density (ρ) and P-wave velocity (Vp) logs, yet these datasets are frequently unavailable due to operational constraints, tool limitations, or borehole irregularities. Existing empirical methods, such as neutron porosity-based formulas, suffer from restrictive assumptions -including matrix/fluid constant dependencies, low shale tolerance (< 25%), and negligible secondary porosity - that limit applicability in heterogeneous formations. To overcome these challenges, we present a robust machine learning workflow that predicts Z directly from commonly available well logs, circumventing the need for sonic or density data. A multi-well dataset comprising gamma-ray (GR), neutron porosity (NPHI), deep resistivity (RD), and formation tops were analyzed. Pearson correlation identified GR, NPHI, and log-transformed resistivity (RDlog) as optimal predictors. Data preprocessing included Isolation Forest-based outlier removal and logarithmic resistivity transformation. The XGBoost regressor - selected for its scalability in handling nonlinear interactions - was trained on 80% of the data, with hyperparameters optimized via cross-validated grid search. Model performance was evaluated using mean absolute error (MAE), root MSE (RMSE), and coefficient of determination (R²). The optimized model achieved an R² of 0.916 (training) and 0.808 (testing), with RMSE values of 718.3 and 1070, respectively. Independent validation on a blind well demonstrated strong generalization (R² = 0.869, RMSE = 981.3), with predicted Z logs showing stratigraphic fidelity and suppression of high-amplitude artifacts inherent to sonic-derived impedance. Compared to empirical methods, the ML workflow eliminates reliance on matrix/fluid constants, accommodates shale volumes > 25%, and mitigates errors from secondary porosity or gas effects. This provides a scalable, cost-effective solution to enhance seismic inversion accuracy in data-scarce or complex lithological settings.