Scholarlink

<p>Numerous recent studies in the machine learning field show nearly perfect accuracy in permeability forecasting from well logs. Usually, the coefficient of determination is above 0.90. Nevertheless, many of these findings have an underlying methodological problem. Models that employ identical log measurements will undergo training aimed at predicting the output of empirical equations. In these cases, the algorithm learns a predetermined mathematical function. The circularity problem is quantitatively tested with 112 laboratory-measured core samples from the Tensleep Sandstone at Teapot Dome, Wyoming, along with over two million well-log measurements from 347 wells. When a Random Forest model was trained using empirical transforms, the resulting R² score was 1.00. However, those same transforms yielded an R² score of − 1.44 when correlated to Klinkenberg permeability from core plugs, providing a systematic bias of − 2.26 log₁₀ units, which is equivalent to quantifying underestimation of permeability by a factor of ~ 182. Models that are calibrated against core measurements and validated through leave-one-out cross-validation obtain a best R² of 0.70 (95% bootstrap confidence interval: 0.58–0.79), with a 65% reduction of root mean square error compared to the transforms. A feature ablation study shows that depth-related variables contribute 46% of model importance, and their removal still results in a RMSE 54% lower than empirical equations under leave-one-out validation, indicating that log-derived features provide predictive signal beyond stratigraphic position alone. For future permeability prediction studies, a five-requirement validation protocol is recommended as a minimum standard.</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

Are machine learning permeability models predicting rock properties or reproducing equations? Core validation evidence from the teapot dome field

  • 期刊论文

  • 发表日期：

    2026年6月24日

查看全文
  • Shaho Mohammed Ali,
  • Roja Dashty,
  • Mohammed Ameen

摘要

Numerous recent studies in the machine learning field show nearly perfect accuracy in permeability forecasting from well logs. Usually, the coefficient of determination is above 0.90. Nevertheless, many of these findings have an underlying methodological problem. Models that employ identical log measurements will undergo training aimed at predicting the output of empirical equations. In these cases, the algorithm learns a predetermined mathematical function. The circularity problem is quantitatively tested with 112 laboratory-measured core samples from the Tensleep Sandstone at Teapot Dome, Wyoming, along with over two million well-log measurements from 347 wells. When a Random Forest model was trained using empirical transforms, the resulting R² score was 1.00. However, those same transforms yielded an R² score of − 1.44 when correlated to Klinkenberg permeability from core plugs, providing a systematic bias of − 2.26 log₁₀ units, which is equivalent to quantifying underestimation of permeability by a factor of ~ 182. Models that are calibrated against core measurements and validated through leave-one-out cross-validation obtain a best R² of 0.70 (95% bootstrap confidence interval: 0.58–0.79), with a 65% reduction of root mean square error compared to the transforms. A feature ablation study shows that depth-related variables contribute 46% of model importance, and their removal still results in a RMSE 54% lower than empirical equations under leave-one-out validation, indicating that log-derived features provide predictive signal beyond stratigraphic position alone. For future permeability prediction studies, a five-requirement validation protocol is recommended as a minimum standard.