Scholarlink

<p>Traditional approaches for estimating Total Organic Carbon (TOC), such as the ΔLogR method, often show limited accuracy in the Longmaxi shale due to their inability to account for subsurface heterogeneity and nonlinear relationships between TOC and petrophysical logs. This study presents a sequential workflow comprising unsupervised lithofacies clustering followed by supervised TOC regression. K-means clustering is first applied to GR, DEN, CNL, and U logs to characterize lithological variability and depositional heterogeneity within the formation. Subsequently, supervised machine learning models are developed using uranium (U), density (DEN), gamma ray (GR), neutron porosity (CNL), photoelectric factor (PE), and deep resistivity (RD) logs for quantitative TOC prediction. The evaluated models include Light Gradient Boosting Machine (LightGBM), Multilayer Perceptron (MLP), Random Forest (RF), and a stacking model. The stacking model achieved the best performance with R² values of 0.9772 (train) and 0.9643 (test), outperforming LightGBM, MLP, and RF. In contrast, the conventional ΔLogR method yielded significantly lower accuracy (R² = 0.4646 train; 0.4614 test). SHAP analysis identifies bulk density as the most influential predictor, consistent with the inverse relationship between density and organic richness. The results demonstrate that lithofacies clustering enhances geological interpretation, while supervised ensemble learning substantially improves TOC prediction in the Longmaxi Formation, Sichuan Basin, China.</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

Enhanced estimation of total organic carbon in the longmaxi shale formation using supervised machine learning with lithofacies characterization via K-means clustering

  • 期刊论文

  • 发表日期：

    2026年4月2日

查看全文
  • Shaukat Khan,
  • Zhishui Liu,
  • Zhiqiang Lu,
  • Wakeel Hussain,
  • Sohail Ahmed,
  • Muhammad Sajid,
  • Hafiz Hussain Ahmad,
  • Obaidullah,
  • Imdadullah

摘要

Traditional approaches for estimating Total Organic Carbon (TOC), such as the ΔLogR method, often show limited accuracy in the Longmaxi shale due to their inability to account for subsurface heterogeneity and nonlinear relationships between TOC and petrophysical logs. This study presents a sequential workflow comprising unsupervised lithofacies clustering followed by supervised TOC regression. K-means clustering is first applied to GR, DEN, CNL, and U logs to characterize lithological variability and depositional heterogeneity within the formation. Subsequently, supervised machine learning models are developed using uranium (U), density (DEN), gamma ray (GR), neutron porosity (CNL), photoelectric factor (PE), and deep resistivity (RD) logs for quantitative TOC prediction. The evaluated models include Light Gradient Boosting Machine (LightGBM), Multilayer Perceptron (MLP), Random Forest (RF), and a stacking model. The stacking model achieved the best performance with R² values of 0.9772 (train) and 0.9643 (test), outperforming LightGBM, MLP, and RF. In contrast, the conventional ΔLogR method yielded significantly lower accuracy (R² = 0.4646 train; 0.4614 test). SHAP analysis identifies bulk density as the most influential predictor, consistent with the inverse relationship between density and organic richness. The results demonstrate that lithofacies clustering enhances geological interpretation, while supervised ensemble learning substantially improves TOC prediction in the Longmaxi Formation, Sichuan Basin, China.