Scholarlink

<p>Unconfined compressive strength (UCS) is a fundamental design parameter for chemically stabilized and fiber-reinforced soils, yet its experimental determination is both time-intensive and resource-demanding, particularly given the 28-day minimum curing period required for meaningful pozzolanic strength development. This study presents a machine learning (ML) framework for predicting the UCS of lime-treated, expansive clay soils reinforced with polypropylene fibers (PPF). A multi-source database of 310 experimental records compiled from 13 independent published studies was supplemented with original laboratory data collected on a high-plasticity clay (MH) from Medea, Algeria, incorporating five input features: lime content, PPF content, fiber length, curing time, and the initial UCS of the untreated soil. Five ML algorithms were systematically developed and benchmarked — XGBoost, CatBoost, LightGBM, TabNet, and TabPFN — encompassing both gradient boosting ensemble and deep tabular learning paradigms. Hyperparameter optimization was performed via Bayesian search using Optuna with 100 trials per model. Results demonstrate that gradient boosting models consistently outperformed deep tabular approaches, with CatBoost and LightGBM achieving the highest predictive accuracy on the test set (R² = 0.96, RMSE = 125.48&#xa0;kPa and 126.06&#xa0;kPa, respectively), while TabNet yielded the lowest performance (R² = 0.92, RMSE = 187.95&#xa0;kPa). SHapley Additive exPlanations (SHAP) applied to the best-performing model revealed that lime content dominated UCS predictions (mean |SHAP| = 248.51&#xa0;kPa), followed by initial soil strength (208.84&#xa0;kPa) and curing time (137.14&#xa0;kPa), while PPF content and fiber length provided secondary yet physically meaningful contributions (59.98 and 33.72&#xa0;kPa, respectively). The proposed framework constitutes an interpretable, data-driven surrogate that substantially reduces reliance on extended laboratory testing and supports efficient mix-design optimization in geotechnical ground improvement practice.</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

Predicting the UCS of lime-treated clay soils reinforced with polypropylene fibers using ensemble and deep tabular learning models with SHAP Interpretability

  • 期刊论文

  • 发表日期：

    2026年7月17日

查看全文
  • Anfal Rihab Cheriet,
  • Ali Benzaamia,
  • Bilal Sari Ahmed,
  • Mohamed Ghrici,
  • Khoudja AliBenyahiya

摘要

Unconfined compressive strength (UCS) is a fundamental design parameter for chemically stabilized and fiber-reinforced soils, yet its experimental determination is both time-intensive and resource-demanding, particularly given the 28-day minimum curing period required for meaningful pozzolanic strength development. This study presents a machine learning (ML) framework for predicting the UCS of lime-treated, expansive clay soils reinforced with polypropylene fibers (PPF). A multi-source database of 310 experimental records compiled from 13 independent published studies was supplemented with original laboratory data collected on a high-plasticity clay (MH) from Medea, Algeria, incorporating five input features: lime content, PPF content, fiber length, curing time, and the initial UCS of the untreated soil. Five ML algorithms were systematically developed and benchmarked — XGBoost, CatBoost, LightGBM, TabNet, and TabPFN — encompassing both gradient boosting ensemble and deep tabular learning paradigms. Hyperparameter optimization was performed via Bayesian search using Optuna with 100 trials per model. Results demonstrate that gradient boosting models consistently outperformed deep tabular approaches, with CatBoost and LightGBM achieving the highest predictive accuracy on the test set (R² = 0.96, RMSE = 125.48 kPa and 126.06 kPa, respectively), while TabNet yielded the lowest performance (R² = 0.92, RMSE = 187.95 kPa). SHapley Additive exPlanations (SHAP) applied to the best-performing model revealed that lime content dominated UCS predictions (mean |SHAP| = 248.51 kPa), followed by initial soil strength (208.84 kPa) and curing time (137.14 kPa), while PPF content and fiber length provided secondary yet physically meaningful contributions (59.98 and 33.72 kPa, respectively). The proposed framework constitutes an interpretable, data-driven surrogate that substantially reduces reliance on extended laboratory testing and supports efficient mix-design optimization in geotechnical ground improvement practice.