Scholarlink

<p>This study aims to develop and compare hybrid ensemble learning models to estimate the resilient modulus (M<sub>r</sub>) of stabilized weak subgrade soils. The dataset comprised 580 M<sub>r</sub> measurements for waek subgrade soils stabilized with lime, class C fly ash, and cement kiln dust. To develop the models, the most significant variables of soil and additive parameters were first identified using the Neighborhood Component Analysis (NCA) method. Then, three hybrid ensemble learning methods, including Gradient Boosting Regression Tree-Equilibrium Optimization Algorithm (GBRT-EOA), Adaptive Boosting-EOA (AdaBoost-EOA), and Random Forest-EOA (RF-EOA), were developed to predict M<sub>r</sub> and compared with artificial neural network and support vector machine models as well as models reported in previous studies. Results showed that GBRT-EOA emerged as the superior algorithm with an R<sup>2</sup> of 0.986 and 0.97 for the training and testing sets, respectively. The Shapely Additive exPlanations (SHAP) analysis showed that additive percentage was the most significant variable for predicting M<sub>r</sub>. Parametric analysis from SHAP analysis also showed that increasing additive percentage and calcium oxide content raised M<sub>r</sub>, whereas increasing liquid limit and deviatoric stress led to a decrease in M<sub>r</sub>.</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 Hybrid Ensemble Learning Methods for Predicting the Resilient Modulus of Stabilized Weak Subgrade Soils

  • 期刊论文

  • 发表日期：

    2025年12月20日

查看全文
  • Mahdi Ghodratabadi,
  • Ali Reza Ghanizadeh

摘要

This study aims to develop and compare hybrid ensemble learning models to estimate the resilient modulus (Mr) of stabilized weak subgrade soils. The dataset comprised 580 Mr measurements for waek subgrade soils stabilized with lime, class C fly ash, and cement kiln dust. To develop the models, the most significant variables of soil and additive parameters were first identified using the Neighborhood Component Analysis (NCA) method. Then, three hybrid ensemble learning methods, including Gradient Boosting Regression Tree-Equilibrium Optimization Algorithm (GBRT-EOA), Adaptive Boosting-EOA (AdaBoost-EOA), and Random Forest-EOA (RF-EOA), were developed to predict Mr and compared with artificial neural network and support vector machine models as well as models reported in previous studies. Results showed that GBRT-EOA emerged as the superior algorithm with an R2 of 0.986 and 0.97 for the training and testing sets, respectively. The Shapely Additive exPlanations (SHAP) analysis showed that additive percentage was the most significant variable for predicting Mr. Parametric analysis from SHAP analysis also showed that increasing additive percentage and calcium oxide content raised Mr, whereas increasing liquid limit and deviatoric stress led to a decrease in Mr.