Scholarlink

<p>A database containing 4025 rock mechanics records was compiled to develop computational models for the multi-output prediction of uniaxial compressive strength, Young’s modulus and Brazilian tensile strength. After duplicate removal, 3905 observations were retained and processed through median imputation, anomaly detection, feature standardization and physics-informed feature engineering. Three interaction features were generated to represent coupled geomechanical relationships among density, wave velocity, hardness and point-load characteristics. Random Forest, Extra Trees, Extreme Gradient Boosting, CatBoost and a stacking ensemble framework were developed and evaluated using coefficient of determination, root mean square error and mean absolute error. The stacking ensemble achieved the highest predictive performance for uniaxial compressive strength (r²-score = 0.78) and Young’s modulus (r²-score = 0.60), whereas Extreme Gradient Boosting produced the best prediction of Brazilian tensile strength (r²-score = 0.75). Repeated K-fold cross-validation yielded a mean r²-score of 0.69 with a standard deviation of 0.03, while bootstrap analysis produced a 95% confidence interval of 0.73–0.82. Sensitivity analysis indicated that the proposed interaction features exerted the strongest influence on model predictions. Statistical testing confirmed significant differences among the evaluated models. The developed framework provided accurate, robust and interpretable prediction of rock mechanical properties.</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

Statistical and Sensitivity Evaluation of Advanced Computational Models for Predicting Rock Mechanical Properties

  • 期刊论文

  • 发表日期：

    2026年7月7日

查看全文
  • Aman Jangir,
  • Biswajit Acharya

摘要

A database containing 4025 rock mechanics records was compiled to develop computational models for the multi-output prediction of uniaxial compressive strength, Young’s modulus and Brazilian tensile strength. After duplicate removal, 3905 observations were retained and processed through median imputation, anomaly detection, feature standardization and physics-informed feature engineering. Three interaction features were generated to represent coupled geomechanical relationships among density, wave velocity, hardness and point-load characteristics. Random Forest, Extra Trees, Extreme Gradient Boosting, CatBoost and a stacking ensemble framework were developed and evaluated using coefficient of determination, root mean square error and mean absolute error. The stacking ensemble achieved the highest predictive performance for uniaxial compressive strength (r²-score = 0.78) and Young’s modulus (r²-score = 0.60), whereas Extreme Gradient Boosting produced the best prediction of Brazilian tensile strength (r²-score = 0.75). Repeated K-fold cross-validation yielded a mean r²-score of 0.69 with a standard deviation of 0.03, while bootstrap analysis produced a 95% confidence interval of 0.73–0.82. Sensitivity analysis indicated that the proposed interaction features exerted the strongest influence on model predictions. Statistical testing confirmed significant differences among the evaluated models. The developed framework provided accurate, robust and interpretable prediction of rock mechanical properties.