Scholarlink

<p>Shear-wave velocity (Vs) is essential for seismic interpretation, reservoir characterization, and geomechanical analysis, yet its direct measurement using dipole sonic tools remains limited by acquisition cost, leaving Vs logs absent in many wells. This study presents a leakage-aware, physically interpretable, and reproducible machine-learning workflow to estimate Vs from conventional well logs in the Datta Formation, Kohat Plateau, Pakistan. A dataset of 749 clean samples from the Well-A comprising compressional velocity, bulk density, neutron porosity, gamma ray, deep resistivity, photoelectric factor, and caliper logs, was used as predictor variables, while Vs derived from shear slowness logs served as the reference target. Five widely used empirical Vp–Vs relationships were first evaluated as baseline models, all of which showed severely negative <i>R</i><sup>2</sup> values and mean biases ranging from + 0.90 to + 1.33&#xa0;km/s, confirming their inapplicability to this formation. Three machine-learning models Random Forest (RF), Support Vector Machine (SVM), and Multilayer Perceptron (MLP) were then developed using a systematic 60/20/20 training-validation-testing strategy. RF achieved the strongest test-set performance (<i>R</i><sup>2</sup> = 0.88, RMSE = 0.0175&#xa0;km/s), substantially outperforming both empirical baselines and the SVM and MLP models. Uncertainty quantification using RF ensemble variance, bootstrap resampling, and split-conformal prediction intervals showed that RF also provides the most reliable calibrated uncertainty estimates, achieving 90.7% conformal coverage at the nominal 90% level with a mean interval width of 0.061&#xa0;km/s. Feature importance analysis consistently identified compressional velocity as the dominant predictor, followed by gamma ray and bulk density, supporting the physical interpretability of the workflow. The proposed approach offers a reliable and reproducible solution for Vs estimation in wells lacking direct shear-wave measurements, with practical applications in seismic and geomechanical studies in similar geological 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

Interpretable and uncertainty-aware machine learning for shear-wave velocity estimation from conventional well logs

  • 期刊论文

  • 发表日期：

    2026年7月16日

查看全文
  • Akbar Ali,
  • Rongyi Qian

摘要

Shear-wave velocity (Vs) is essential for seismic interpretation, reservoir characterization, and geomechanical analysis, yet its direct measurement using dipole sonic tools remains limited by acquisition cost, leaving Vs logs absent in many wells. This study presents a leakage-aware, physically interpretable, and reproducible machine-learning workflow to estimate Vs from conventional well logs in the Datta Formation, Kohat Plateau, Pakistan. A dataset of 749 clean samples from the Well-A comprising compressional velocity, bulk density, neutron porosity, gamma ray, deep resistivity, photoelectric factor, and caliper logs, was used as predictor variables, while Vs derived from shear slowness logs served as the reference target. Five widely used empirical Vp–Vs relationships were first evaluated as baseline models, all of which showed severely negative R2 values and mean biases ranging from + 0.90 to + 1.33 km/s, confirming their inapplicability to this formation. Three machine-learning models Random Forest (RF), Support Vector Machine (SVM), and Multilayer Perceptron (MLP) were then developed using a systematic 60/20/20 training-validation-testing strategy. RF achieved the strongest test-set performance (R2 = 0.88, RMSE = 0.0175 km/s), substantially outperforming both empirical baselines and the SVM and MLP models. Uncertainty quantification using RF ensemble variance, bootstrap resampling, and split-conformal prediction intervals showed that RF also provides the most reliable calibrated uncertainty estimates, achieving 90.7% conformal coverage at the nominal 90% level with a mean interval width of 0.061 km/s. Feature importance analysis consistently identified compressional velocity as the dominant predictor, followed by gamma ray and bulk density, supporting the physical interpretability of the workflow. The proposed approach offers a reliable and reproducible solution for Vs estimation in wells lacking direct shear-wave measurements, with practical applications in seismic and geomechanical studies in similar geological settings.