Scholarlink

Abstract <p>Total organic carbon (TOC) is a key parameter for evaluating unconventional resources and CO<sub>2</sub> geological storage potential. The accurate determination of TOC relies on expensive and time-consuming experimental methods. Therefore, developing an accurate and reliable machine learning (ML) model is crucial. This study takes the 623 shale data of the Shuangyang Formation in the Yitong Basin of Northeast China as an example. With conventional logging (gamma ray log, GR, deep lateral resistivity log, RLLD, acoustic log, AC, and density log, DEN) as input and TOC data as output, we developed an accurate and interpretable Random Forest (RF) model. The RF model has an accuracy of 92% in the testing set, and in addition, the validation accuracy of the independent well (XT-1) is 88%. In addition, we use Shapley Additive explanations (SHAP) to provide global and local interpretations of the model, clarifying the influence between each input variable and TOC, as well as the mutual influence between variables. A force plot is also used to perform local interpretation analysis on a single set of data samples. Finally, we sort the input features and find that acoustic log (AC) has the highest contribution. The visualization analysis of predictive models will assist in the effectiveness evaluation of unconventional oil and gas exploration, as well as CO₂ geological storage.</p> Graphical abstract <p></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

Application of machine learning model in shale TOC content prediction based on well log data: enhancing model interpretability by SHAP

  • 期刊论文

  • 发表日期：

    2025年6月2日

查看全文
  • Ruibin Chen,
  • Xinyu Liu,
  • Sandong Zhou,
  • Weixin Zhang,
  • Hang Liu,
  • Detian Yan,
  • Hua Wang

摘要

Abstract

Total organic carbon (TOC) is a key parameter for evaluating unconventional resources and CO2 geological storage potential. The accurate determination of TOC relies on expensive and time-consuming experimental methods. Therefore, developing an accurate and reliable machine learning (ML) model is crucial. This study takes the 623 shale data of the Shuangyang Formation in the Yitong Basin of Northeast China as an example. With conventional logging (gamma ray log, GR, deep lateral resistivity log, RLLD, acoustic log, AC, and density log, DEN) as input and TOC data as output, we developed an accurate and interpretable Random Forest (RF) model. The RF model has an accuracy of 92% in the testing set, and in addition, the validation accuracy of the independent well (XT-1) is 88%. In addition, we use Shapley Additive explanations (SHAP) to provide global and local interpretations of the model, clarifying the influence between each input variable and TOC, as well as the mutual influence between variables. A force plot is also used to perform local interpretation analysis on a single set of data samples. Finally, we sort the input features and find that acoustic log (AC) has the highest contribution. The visualization analysis of predictive models will assist in the effectiveness evaluation of unconventional oil and gas exploration, as well as CO₂ geological storage.

Graphical abstract