Scholarlink

<p>This study integrates machine learning to predict elemental composition and classify stratigraphic members within a Permian–Triassic epeiric homoclinal ramp. Using logs from six vertical wells, we trained random forest (RF), gradient boosting (GB), XGBoost (XGB), decision tree (DT), and multilayer perceptron (MLP) models to predict Ca, S, Si, and Al and to classify members. Exploratory data analysis and principal component analysis were used to relate inputs to elemental responses. The best elemental test <i>R</i><sup>2</sup> reached 0.86 (S with RF), while the lowest values were 0.55–0.56 (DT for Ca/Si). In a blind-well test, <i>R</i><sup>2</sup> values were modest overall (0.02–0.24), with ensembles performing best; sulfur remained the most stable target due to its strong density/sonic expression. However, the produced synthetic elemental logs demonstrated preserved stratigraphic cyclicity used in correlation. For member classification, ensembles and MLP outperformed DT in standard metrics (accuracy/area under the receiver operating characteristic [ROC] curve [AUC]). Compared with traditional interpretation and chemostratigraphy, the machine learning workflow accelerates the generation of synthetic elemental logs and member picks while providing quantitative, reproducible metrics; it complements core and expert judgment rather than replacing them. Potential applications include rapid well-to-well correlation and while-drilling decision support, with current generalizability primarily to this Permian–Triassic ramp and close regional equivalents.</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

Well Log-Based Prediction of Elemental Composition: A Machine Learning Approach for Classifying Stratigraphic Members in the Epeiric Homoclinal Ramps

  • 期刊论文

  • 发表日期：

    2026年2月5日

查看全文
  • Abdallah Abdelkarim,
  • Ammar Abdlmutalib,
  • John Humphrey,
  • Khalid Al-Ramadan

摘要

This study integrates machine learning to predict elemental composition and classify stratigraphic members within a Permian–Triassic epeiric homoclinal ramp. Using logs from six vertical wells, we trained random forest (RF), gradient boosting (GB), XGBoost (XGB), decision tree (DT), and multilayer perceptron (MLP) models to predict Ca, S, Si, and Al and to classify members. Exploratory data analysis and principal component analysis were used to relate inputs to elemental responses. The best elemental test R2 reached 0.86 (S with RF), while the lowest values were 0.55–0.56 (DT for Ca/Si). In a blind-well test, R2 values were modest overall (0.02–0.24), with ensembles performing best; sulfur remained the most stable target due to its strong density/sonic expression. However, the produced synthetic elemental logs demonstrated preserved stratigraphic cyclicity used in correlation. For member classification, ensembles and MLP outperformed DT in standard metrics (accuracy/area under the receiver operating characteristic [ROC] curve [AUC]). Compared with traditional interpretation and chemostratigraphy, the machine learning workflow accelerates the generation of synthetic elemental logs and member picks while providing quantitative, reproducible metrics; it complements core and expert judgment rather than replacing them. Potential applications include rapid well-to-well correlation and while-drilling decision support, with current generalizability primarily to this Permian–Triassic ramp and close regional equivalents.