Scholarlink

<p>This study presented a hybrid machine learning (ML) approach for predicting the small-strain shear modulus (<i>G</i><sub>max</sub>) in granular soils that integrated AdaBoost, Decision Tree, and CatBoost models with the Gorilla Troops Optimization algorithm to improve predictive accuracy and model robustness. The approach addressed key limitations of conventional empirical models and standalone ML models in capturing complex parameter interactions across varying soil conditions. A database of 816 samples was compiled using four key soil parameters: void ratio, confining pressure, coefficient of uniformity, and particle shape descriptor. Among the developed models, CatBoost outperformed the other ML models and empirical correlations available in the literature for <i>G</i><sub>max</sub> estimation, achieving coefficient of determination (R²) values of 0.986 (training) and 0.994 (testing) with minimal associated errors. To enhance model interpretability and transparency, Shapley additive explanations, partial dependence plots, and individual conditional expectation analyses were applied. The results showed that confining pressure was the most influential predictor, while the particle shape descriptor had the least effect on <i>G</i><sub>max</sub>. The proposed approach provides a reliable, interpretable tool for engineers, supporting more accurate <i>G</i><sub>max</sub> estimation and reducing uncertainty in geotechnical design.</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

Enhancing Predictive Accuracy and Interpretability of Small Strain Shear Modulus for Granular Soils Using Machine Learning Models

  • 期刊论文

  • 发表日期：

    2026年2月18日

查看全文
  • Muhammad Suliman,
  • Majid Khan,
  • Zia Ul Islam

摘要

This study presented a hybrid machine learning (ML) approach for predicting the small-strain shear modulus (Gmax) in granular soils that integrated AdaBoost, Decision Tree, and CatBoost models with the Gorilla Troops Optimization algorithm to improve predictive accuracy and model robustness. The approach addressed key limitations of conventional empirical models and standalone ML models in capturing complex parameter interactions across varying soil conditions. A database of 816 samples was compiled using four key soil parameters: void ratio, confining pressure, coefficient of uniformity, and particle shape descriptor. Among the developed models, CatBoost outperformed the other ML models and empirical correlations available in the literature for Gmax estimation, achieving coefficient of determination (R²) values of 0.986 (training) and 0.994 (testing) with minimal associated errors. To enhance model interpretability and transparency, Shapley additive explanations, partial dependence plots, and individual conditional expectation analyses were applied. The results showed that confining pressure was the most influential predictor, while the particle shape descriptor had the least effect on Gmax. The proposed approach provides a reliable, interpretable tool for engineers, supporting more accurate Gmax estimation and reducing uncertainty in geotechnical design.