Scholarlink

<p>Petrographic characteristics have long been known to control rock strength, including uniaxial compressive (UCS) and tensile strength (TS). However, the influence of these characteristics on granite strength has not been fully understood. Machine learning (ML) approach, as a data-driven research paradigm, requires a large, complete, and consistent dataset to accurately predict the non-linear relationship in granite. Currently, such comprehensive datasets are lacking, and the complexity of ML models makes it challenging to interpret the relationships between input properties and predicted strength. To address these challenges, 578 granite samples from 105 publications are analyzed in this study, encompassing physico-mechanical and petrographic parameters. Various data imputation methods are tested, with the Multivariate Imputation by Chained Equations (MICE) technique performing best. Several ML approaches are used to predict the UCS and TS of granite, with the Extreme Gradient Boosting (XGBoost) model demonstrated superior performance, achieving the lowest prediction errors and highest correlation coefficients. To interpret the predicted results, SHapley Additive exPlanations (SHAP), Partial Dependence Plots (PDP) combined with Individual Conditional Expectation (ICE) plots, and feature importance rankings are employed. The analysis revealed that porosity, grain size, and biotite content are the most critical factors, each showing a significant negative correlation with granite strength. Conversely, plagioclase, K-feldspar, and quartz content have no significant impact on UCS and TS of granite.</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

The Impact of Petrographic Characteristics on Granite Strength: Insights from an Interpretable Machine Learning Model

  • 期刊论文

  • 发表日期：

    2025年6月26日

查看全文
  • Kaiwei Tian,
  • Zeqi Zhu,
  • Qian Sheng,
  • Ning Tian,
  • Peng Wu,
  • Ningxi Zhou

摘要

Petrographic characteristics have long been known to control rock strength, including uniaxial compressive (UCS) and tensile strength (TS). However, the influence of these characteristics on granite strength has not been fully understood. Machine learning (ML) approach, as a data-driven research paradigm, requires a large, complete, and consistent dataset to accurately predict the non-linear relationship in granite. Currently, such comprehensive datasets are lacking, and the complexity of ML models makes it challenging to interpret the relationships between input properties and predicted strength. To address these challenges, 578 granite samples from 105 publications are analyzed in this study, encompassing physico-mechanical and petrographic parameters. Various data imputation methods are tested, with the Multivariate Imputation by Chained Equations (MICE) technique performing best. Several ML approaches are used to predict the UCS and TS of granite, with the Extreme Gradient Boosting (XGBoost) model demonstrated superior performance, achieving the lowest prediction errors and highest correlation coefficients. To interpret the predicted results, SHapley Additive exPlanations (SHAP), Partial Dependence Plots (PDP) combined with Individual Conditional Expectation (ICE) plots, and feature importance rankings are employed. The analysis revealed that porosity, grain size, and biotite content are the most critical factors, each showing a significant negative correlation with granite strength. Conversely, plagioclase, K-feldspar, and quartz content have no significant impact on UCS and TS of granite.