Scholarlink

<p>Marine gas hydrates are both an untapped energy source and a drilling hazard, making reliable subsurface assessment essential. We analysed wells in the deepwater Mahanadi Basin, offshore India, to estimate gas hydrate saturation and pore pressure by combining established petrophysical methods with machine learning (ML). Conventional empirical methods, though widely used, often fail to capture complex lithological and geomechanical relationships, especially where limited well information introduces substantial uncertainty. In this study, we present an integrated, data-driven workflow for estimation of gas hydrate saturation and pore pressure using ensemble machine learning (ML) techniques. Archie’s law was first applied to resistivity logs to derive gas hydrate saturations, while Eaton’s and Bower’s methods provided pore pressure profiles from sonic, resistivity, and density data. To overcome empirical limitations, Random Forest, Gradient Boosting, Extremely Randomised Trees, and Bootstrap Aggregating regressors were trained on well NGHP-01-08 and validated on well NGHP-01-19 from the Mahanadi Basin, offshore India. The ML models successfully identified nonlinear patterns in well log responses (resistivity, P-wave velocity, gamma ray, bulk density, and stratigraphic depth bins) and accurately reproduced both smooth background trends and sharp local anomalies, including hydrate-rich intervals near the Bottom Simulating Reflector. Among the tested models, Random Forest achieved the highest predictive accuracy in terms of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^{2}\)</EquationSource> </InlineEquation> across both training and testing phases, followed by Gradient Boosting, Extremely Randomized Trees, and Bootstrap Aggregating. For pore pressure estimation, Random Forest and Gradient Boosting regressors were trained on well NGHP-01-08 and validated on well NGHP-01-19, with both models delivering excellent predictive accuracy across Eaton’s and Bower’s methods, and Gradient Boosting showing improved generalization in heterogeneous intervals. The integrated approach reduces prediction uncertainty, enhances cross-well applicability, and provides a transferable methodology for hydrate and pressure characterization in other offshore basins with 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

Comparative assessment of machine learning and conventional methods for gas hydrate saturation and pore pressure estimation in the Mahanadi Basin, India

  • 期刊论文

  • 发表日期：

    2026年7月1日

查看全文
  • Aafreen,
  • Komal Tiwari,
  • Pradeep Kumar Yadav,
  • Uma Shankar

摘要

Marine gas hydrates are both an untapped energy source and a drilling hazard, making reliable subsurface assessment essential. We analysed wells in the deepwater Mahanadi Basin, offshore India, to estimate gas hydrate saturation and pore pressure by combining established petrophysical methods with machine learning (ML). Conventional empirical methods, though widely used, often fail to capture complex lithological and geomechanical relationships, especially where limited well information introduces substantial uncertainty. In this study, we present an integrated, data-driven workflow for estimation of gas hydrate saturation and pore pressure using ensemble machine learning (ML) techniques. Archie’s law was first applied to resistivity logs to derive gas hydrate saturations, while Eaton’s and Bower’s methods provided pore pressure profiles from sonic, resistivity, and density data. To overcome empirical limitations, Random Forest, Gradient Boosting, Extremely Randomised Trees, and Bootstrap Aggregating regressors were trained on well NGHP-01-08 and validated on well NGHP-01-19 from the Mahanadi Basin, offshore India. The ML models successfully identified nonlinear patterns in well log responses (resistivity, P-wave velocity, gamma ray, bulk density, and stratigraphic depth bins) and accurately reproduced both smooth background trends and sharp local anomalies, including hydrate-rich intervals near the Bottom Simulating Reflector. Among the tested models, Random Forest achieved the highest predictive accuracy in terms of \(R^{2}\) across both training and testing phases, followed by Gradient Boosting, Extremely Randomized Trees, and Bootstrap Aggregating. For pore pressure estimation, Random Forest and Gradient Boosting regressors were trained on well NGHP-01-08 and validated on well NGHP-01-19, with both models delivering excellent predictive accuracy across Eaton’s and Bower’s methods, and Gradient Boosting showing improved generalization in heterogeneous intervals. The integrated approach reduces prediction uncertainty, enhances cross-well applicability, and provides a transferable methodology for hydrate and pressure characterization in other offshore basins with similar geological settings.