Scholarlink

<p>Despite substantial research into shale pore structure, quantitative characterization of diverse pore types in shale remains a considerable challenge. Accurate delineation of shale pore types is crucial for improving further understanding of reservoir properties and predicting reservoir productivity. The present study investigated shales from the Upper Ordovician Wufeng Formation and Lower Silurian Longmaxi Formation to quantitatively characterize various pore types using machine learning (ML) coupled with scanning electron microscopy (SEM). Five ML models were evaluated using stratified 10-fold cross-validation: naive Bayes, artificial neural network-multilayer perceptron, random decision tree, logistic model tree, and fast random forest. Following segmentation, Fiji ImageJ was employed to extract pore area and Ferret diameter from organic matter, quartz, feldspar, carbonate minerals, and clay minerals to characterize the pore structure within each phase. The results demonstrated that the fast random forest model offered superior segmentation accuracy and efficiency, achieving an average intersection over union of 0.95 and an average F<sub>1</sub>-score of 0.96, while requiring only 42.6&#xa0;s on average to identify pores in each SEM image. Pore type development within the Wufeng–Longmaxi Formation shales exhibited significant variability, with each type displaying relatively limited and specific pore size distributions. Organic matter displayed the highest porosity, predominantly characterized by pores ranging 10–200&#xa0;nm, followed by clay minerals and quartz. Carbonate minerals exhibited the lowest porosity. The methodologies presented herein demonstrate a significant potential for the accurate and efficient characterization of diverse pore types in shale reservoirs.</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

Quantitative Characterization of Different Pore Types in Marine Shale Based on Machine Learning and SEM Images

  • 期刊论文

  • 发表日期：

    2025年5月31日

查看全文
  • Xinlei Wang,
  • Zhaodong Xi,
  • Songhang Zhang,
  • Shuheng Tang,
  • Zhifeng Yan

摘要

Despite substantial research into shale pore structure, quantitative characterization of diverse pore types in shale remains a considerable challenge. Accurate delineation of shale pore types is crucial for improving further understanding of reservoir properties and predicting reservoir productivity. The present study investigated shales from the Upper Ordovician Wufeng Formation and Lower Silurian Longmaxi Formation to quantitatively characterize various pore types using machine learning (ML) coupled with scanning electron microscopy (SEM). Five ML models were evaluated using stratified 10-fold cross-validation: naive Bayes, artificial neural network-multilayer perceptron, random decision tree, logistic model tree, and fast random forest. Following segmentation, Fiji ImageJ was employed to extract pore area and Ferret diameter from organic matter, quartz, feldspar, carbonate minerals, and clay minerals to characterize the pore structure within each phase. The results demonstrated that the fast random forest model offered superior segmentation accuracy and efficiency, achieving an average intersection over union of 0.95 and an average F1-score of 0.96, while requiring only 42.6 s on average to identify pores in each SEM image. Pore type development within the Wufeng–Longmaxi Formation shales exhibited significant variability, with each type displaying relatively limited and specific pore size distributions. Organic matter displayed the highest porosity, predominantly characterized by pores ranging 10–200 nm, followed by clay minerals and quartz. Carbonate minerals exhibited the lowest porosity. The methodologies presented herein demonstrate a significant potential for the accurate and efficient characterization of diverse pore types in shale reservoirs.