Scholarlink

<p>Predicting the Drilling Rate Index (DRI) with both high accuracy and physical interpretability remains a major challenge in rock excavation engineering, due to the nonlinear coupling of mechanical, petrographic, and abrasivity parameters. This study presents an Explainable Hybrid Gradient-Boosting Framework that integrates ensemble learning, Bayesian optimization, and explainable artificial intelligence (XAI) to deliver geomechanically interpretable DRI predictions for granitic rocks. A comprehensive experimental database of 41 granitic rock samples, encompassing 27 geotechnical variables, including uniaxial compressive strength (35–182 MPa), Brazilian tensile strength (5–20 MPa), brittleness index (S20) (25–60%), Schmidt rebound hardness (26–70), and Cerchar abrasivity index (1.3–3.9), formed the modeling foundation. Four gradient-boosting algorithms (XGBoost, LightGBM, CatBoost, and AdaBoost) were optimized via Bayesian hyperparameter tuning and validated under stratified fivefold cross-validation. A Shallow Neural Network (SNN) was additionally implemented as a benchmark for direct comparison on the same dataset. All optimized ensemble models achieved <i>R</i><sup>2</sup>test &gt; 0.994 and RMSE &lt; 0.9 DRI units, with CatBoost delivering the highest accuracy (<i>R</i><sup>2</sup>test = 0.9987, RMSEtest = 0.43). Bayesian posterior inference confirmed narrow uncertainty bounds (MAE<sub>95</sub>% CI ≤  ± 0.05), and Taylor diagram analyses verified near-perfect correlation and variance fidelity across all models. SHAP and Partial Dependence Plot analyses identified brittleness index (S<sub>20</sub>), Sievers’ J value (<i>S</i><sub><i>J</i></sub>), Brazilian tensile strength (BTS), and Cerchar abrasivity index (CAI) as the dominant DRI controls, collectively accounting for ≈ 65% of predictive variance, with physically consistent nonlinear threshold behavior confirmed across the full drillability spectrum. The optimal CatBoost model was operationalized in a graphical user interface (GUI) for real-time DRI estimation, supporting practical excavation decision-making.</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

Explainable Gradient-Boosting Framework for Robust Prediction and Geomechanical Interpretation of Drilling Rate Index (DRI) in Granitic Rocks

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Ebrahim Sharifi Teshnizi,
  • Suhaib Rasool Wani,
  • Harish Panghal,
  • Mohammad Ghafoori,
  • Gholam Reza Lashkaripour

摘要

Predicting the Drilling Rate Index (DRI) with both high accuracy and physical interpretability remains a major challenge in rock excavation engineering, due to the nonlinear coupling of mechanical, petrographic, and abrasivity parameters. This study presents an Explainable Hybrid Gradient-Boosting Framework that integrates ensemble learning, Bayesian optimization, and explainable artificial intelligence (XAI) to deliver geomechanically interpretable DRI predictions for granitic rocks. A comprehensive experimental database of 41 granitic rock samples, encompassing 27 geotechnical variables, including uniaxial compressive strength (35–182 MPa), Brazilian tensile strength (5–20 MPa), brittleness index (S20) (25–60%), Schmidt rebound hardness (26–70), and Cerchar abrasivity index (1.3–3.9), formed the modeling foundation. Four gradient-boosting algorithms (XGBoost, LightGBM, CatBoost, and AdaBoost) were optimized via Bayesian hyperparameter tuning and validated under stratified fivefold cross-validation. A Shallow Neural Network (SNN) was additionally implemented as a benchmark for direct comparison on the same dataset. All optimized ensemble models achieved R2test > 0.994 and RMSE < 0.9 DRI units, with CatBoost delivering the highest accuracy (R2test = 0.9987, RMSEtest = 0.43). Bayesian posterior inference confirmed narrow uncertainty bounds (MAE95% CI ≤  ± 0.05), and Taylor diagram analyses verified near-perfect correlation and variance fidelity across all models. SHAP and Partial Dependence Plot analyses identified brittleness index (S20), Sievers’ J value (SJ), Brazilian tensile strength (BTS), and Cerchar abrasivity index (CAI) as the dominant DRI controls, collectively accounting for ≈ 65% of predictive variance, with physically consistent nonlinear threshold behavior confirmed across the full drillability spectrum. The optimal CatBoost model was operationalized in a graphical user interface (GUI) for real-time DRI estimation, supporting practical excavation decision-making.