<p>Reliable prediction of the long-term mechanical durability of lime-stabilised subgrade soils under repeated wetting–drying (W–D) cycles is a critical bottleneck for performance-based geotechnical design of transportation and earthwork infrastructure: empirical durability assessment per ASTM D559 is slow, costly and limited in covering the joint variation of soil index properties, lime dosage, curing time and number of cycles. To address this gap, a hyperparameter-optimised Extreme Gradient Boosting (XGBoost) framework was developed and benchmarked for predicting the unconfined compressive strength (UCS) of lime-stabilised soils after W–D cycles, trained on 444 ASTM D559-compliant data points compiled from 20 peer-reviewed studies. Eight key hyperparameters were systematically tuned using six metaheuristic optimisation algorithms—Particle Swarm Optimisation (PSO), Ant Colony Optimisation for Continuous Domains (ACOR), Harris Hawk’s Optimisation (HHO), the Equilibrium Optimizer (EO), the Improved Multi-operator Differential Evolution Algorithm (IMODE) and LSHADE-cnEpSin—under identical search bounds, identical 50-particle/100-iteration budgets and a 10-fold cross-validated RMSE objective. Algorithm performance was compared by Wilcoxon signed-rank tests over 30 independent seeds, and the pipeline was benchmarked against multiple linear regression with and without second-order interactions, Random Forest and LightGBM on the same train/test split. The optimised XGBoost–LSHADE-cnEpSin model achieved a test-set <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^{2}\)</EquationSource> </InlineEquation> of 0.955 and a 90% split-conformal prediction interval with 93.3% empirical coverage and a mean half-width of 0.70 MPa, providing a distribution-free uncertainty bound suitable for reliability-based design. Shapley Additive Explanations (SHAP), permutation importance, partial dependence analysis and SHAP interaction values jointly identified Lime Content, Curing Time and Number of W–D cycles as the dominant predictors and revealed a damping interaction between Lime Content and cycle-induced damage. The framework provides an accurate, transparent and uncertainty-aware computational tool for assessing the long-term durability of lime-stabilised soils within the calibration envelope (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\textrm{CC}\in [20,60]\%\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\textrm{LC}\in [2,10]\%\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\textrm{CT}\in [7,56]\)</EquationSource> </InlineEquation> d, <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(N\in [0,15]\)</EquationSource> </InlineEquation> cycles), supporting more efficient and risk-informed preliminary geotechnical design.</p>

Predicting the unconfined compressive strength of lime-stabilized soils under wet-dry cycles using extreme gradient boosting optimized with metaheuristic algorithms

期刊论文
出版模式：

开放获取
发表日期：

2026年7月7日

查看全文

Fakhr Eddine M’harzi Alaoui,
Issam Aalil,
Youssef Taki,
Omar Dadah,
Somaya Ben Abbou

摘要

Reliable prediction of the long-term mechanical durability of lime-stabilised subgrade soils under repeated wetting–drying (W–D) cycles is a critical bottleneck for performance-based geotechnical design of transportation and earthwork infrastructure: empirical durability assessment per ASTM D559 is slow, costly and limited in covering the joint variation of soil index properties, lime dosage, curing time and number of cycles. To address this gap, a hyperparameter-optimised Extreme Gradient Boosting (XGBoost) framework was developed and benchmarked for predicting the unconfined compressive strength (UCS) of lime-stabilised soils after W–D cycles, trained on 444 ASTM D559-compliant data points compiled from 20 peer-reviewed studies. Eight key hyperparameters were systematically tuned using six metaheuristic optimisation algorithms—Particle Swarm Optimisation (PSO), Ant Colony Optimisation for Continuous Domains (ACOR), Harris Hawk’s Optimisation (HHO), the Equilibrium Optimizer (EO), the Improved Multi-operator Differential Evolution Algorithm (IMODE) and LSHADE-cnEpSin—under identical search bounds, identical 50-particle/100-iteration budgets and a 10-fold cross-validated RMSE objective. Algorithm performance was compared by Wilcoxon signed-rank tests over 30 independent seeds, and the pipeline was benchmarked against multiple linear regression with and without second-order interactions, Random Forest and LightGBM on the same train/test split. The optimised XGBoost–LSHADE-cnEpSin model achieved a test-set \(R^{2}\) of 0.955 and a 90% split-conformal prediction interval with 93.3% empirical coverage and a mean half-width of 0.70 MPa, providing a distribution-free uncertainty bound suitable for reliability-based design. Shapley Additive Explanations (SHAP), permutation importance, partial dependence analysis and SHAP interaction values jointly identified Lime Content, Curing Time and Number of W–D cycles as the dominant predictors and revealed a damping interaction between Lime Content and cycle-induced damage. The framework provides an accurate, transparent and uncertainty-aware computational tool for assessing the long-term durability of lime-stabilised soils within the calibration envelope ( \(\textrm{CC}\in [20,60]\%\) , \(\textrm{LC}\in [2,10]\%\) , \(\textrm{CT}\in [7,56]\) d, \(N\in [0,15]\) cycles), supporting more efficient and risk-informed preliminary geotechnical design.