Scholarlink

<p>Accurate pore-pressure estimation is critical for safe and efficient drilling, particularly in structurally complex fold-thrust belt settings where pressure behavior is difficult to constrain. The models were trained to predict an Eaton-derived proxy of pore pressure rather than direct measurements, reflecting the current limitation and highlighting the need for future validation with measured pressure values. This study develops a physics-guided machine-learning workflow for predicting Eaton-derived pore-pressure estimates within the Lumshiwal Formation of the Kohat Plateau, Pakistan, using wireline log data from the Makori-01 well. A dataset of 925 depth-indexed samples from a 141&#xa0;m interval was used to train and evaluate five supervised learning algorithms: Gradient Boosting Machine (GBM), Random Forest (RF), Support Vector Machine (SVM), AdaBoost, and Multi-Layer Perceptron (MLP). The dataset was randomly split into 80% training and 20% test subsets, and model performance was further evaluated using cross-validation to mitigate potential overfitting. The target variable was generated from an enhanced Eaton-based formulation constrained by sonic-derived velocity and additional petrophysical corrections. Among the tested models, GBM achieved the best predictive performance, with a test R<sup>2</sup> of 0.7713, RMSE of 23.34 psi, and MAE of 14.77 psi, followed closely by RF (R<sup>2</sup> = 0.7677), indicating that ensemble tree-based methods are most effective for reproducing the nonlinear pressure relationships embedded in the log responses. Although GBM slightly outperformed other models in terms of R² and error metrics, the differences were modest, and selection was based on overall predictive consistency and stability rather than formal statistical significance testing. Feature-importance analysis identified TNPH, GR, and RHOZ as the dominant predictors, highlighting the primary roles of porosity, lithology, and compaction state. Depth-dependent evaluation showed that model accuracy decreases and uncertainty increases with depth, with confidence intervals widening from about ± 3.5 psi in the shallow zone to ± 5.5 psi in the deep zone. Prediction uncertainty was quantified using 95% confidence intervals derived from residual errors of the test dataset, providing a practical assessment of reliability across depth intervals. These results demonstrate that GBM provides a robust surrogate framework for enhanced log-based pore-pressure estimation in the study area. However, because the workflow is trained against an Eaton-derived proxy from a single well, further validation using multi-well datasets and direct pressure measurements is required before broader operational application.</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

Machine Learning-Based Pore Pressure Prediction in a Complex Cretaceous Reservoir Formation, Kohat Plateau, Pakistan

  • 期刊论文

  • 发表日期：

    2026年6月18日

查看全文
  • Akbar Ali,
  • Rongyi Qian,
  • Majid Khan

摘要

Accurate pore-pressure estimation is critical for safe and efficient drilling, particularly in structurally complex fold-thrust belt settings where pressure behavior is difficult to constrain. The models were trained to predict an Eaton-derived proxy of pore pressure rather than direct measurements, reflecting the current limitation and highlighting the need for future validation with measured pressure values. This study develops a physics-guided machine-learning workflow for predicting Eaton-derived pore-pressure estimates within the Lumshiwal Formation of the Kohat Plateau, Pakistan, using wireline log data from the Makori-01 well. A dataset of 925 depth-indexed samples from a 141 m interval was used to train and evaluate five supervised learning algorithms: Gradient Boosting Machine (GBM), Random Forest (RF), Support Vector Machine (SVM), AdaBoost, and Multi-Layer Perceptron (MLP). The dataset was randomly split into 80% training and 20% test subsets, and model performance was further evaluated using cross-validation to mitigate potential overfitting. The target variable was generated from an enhanced Eaton-based formulation constrained by sonic-derived velocity and additional petrophysical corrections. Among the tested models, GBM achieved the best predictive performance, with a test R2 of 0.7713, RMSE of 23.34 psi, and MAE of 14.77 psi, followed closely by RF (R2 = 0.7677), indicating that ensemble tree-based methods are most effective for reproducing the nonlinear pressure relationships embedded in the log responses. Although GBM slightly outperformed other models in terms of R² and error metrics, the differences were modest, and selection was based on overall predictive consistency and stability rather than formal statistical significance testing. Feature-importance analysis identified TNPH, GR, and RHOZ as the dominant predictors, highlighting the primary roles of porosity, lithology, and compaction state. Depth-dependent evaluation showed that model accuracy decreases and uncertainty increases with depth, with confidence intervals widening from about ± 3.5 psi in the shallow zone to ± 5.5 psi in the deep zone. Prediction uncertainty was quantified using 95% confidence intervals derived from residual errors of the test dataset, providing a practical assessment of reliability across depth intervals. These results demonstrate that GBM provides a robust surrogate framework for enhanced log-based pore-pressure estimation in the study area. However, because the workflow is trained against an Eaton-derived proxy from a single well, further validation using multi-well datasets and direct pressure measurements is required before broader operational application.