Scholarlink

<p>This study presents a hybrid experimental–machine learning–optimization framework for predicting the unconfined compressive strength (UCS) of lime-stabilized subgrade soils. Laboratory experiments were conducted on low-plasticity clay with varying lime content (0–10%) and curing periods, generating a dataset that was further expanded and statistically validated. Multiple machine learning models, including MLR, SVR, ANN, DTR, RFR, XGBoost, and CatBoost, were developed and evaluated using performance metrics such as R², RMSE, MAE, IOA, and VAF. Among these, CatBoost achieved the highest accuracy (R² = 0.989, RMSE = 39.85&#xa0;kPa), demonstrating superior capability in capturing nonlinear soil behavior. Explainable AI techniques, including SHAP and partial dependence plots, were applied to interpret predictions, revealing lime content and maximum dry density as dominant positive factors, while plasticity index showed a negative influence. A multi-objective optimization framework using NSGA-II was implemented to maximize UCS while minimizing lime content and plasticity. Pareto-optimal solutions were ranked using TOPSIS, identifying optimal mix designs for practical applications. The proposed framework provides a reliable, interpretable, and efficient approach for sustainable soil stabilization and engineering decision-making by integrating experimental insights with advanced data-driven modeling and optimization strategies to enhance performance, reduce material usage, and support cost-effective infrastructure development in geotechnical engineering practice contexts.</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

Data-driven prediction and optimization of unconfined compressive strength in lime-stabilized soils using explainable machine learning

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Prashant Tiwari,
  • Smita Tung

摘要

This study presents a hybrid experimental–machine learning–optimization framework for predicting the unconfined compressive strength (UCS) of lime-stabilized subgrade soils. Laboratory experiments were conducted on low-plasticity clay with varying lime content (0–10%) and curing periods, generating a dataset that was further expanded and statistically validated. Multiple machine learning models, including MLR, SVR, ANN, DTR, RFR, XGBoost, and CatBoost, were developed and evaluated using performance metrics such as R², RMSE, MAE, IOA, and VAF. Among these, CatBoost achieved the highest accuracy (R² = 0.989, RMSE = 39.85 kPa), demonstrating superior capability in capturing nonlinear soil behavior. Explainable AI techniques, including SHAP and partial dependence plots, were applied to interpret predictions, revealing lime content and maximum dry density as dominant positive factors, while plasticity index showed a negative influence. A multi-objective optimization framework using NSGA-II was implemented to maximize UCS while minimizing lime content and plasticity. Pareto-optimal solutions were ranked using TOPSIS, identifying optimal mix designs for practical applications. The proposed framework provides a reliable, interpretable, and efficient approach for sustainable soil stabilization and engineering decision-making by integrating experimental insights with advanced data-driven modeling and optimization strategies to enhance performance, reduce material usage, and support cost-effective infrastructure development in geotechnical engineering practice contexts.