<p>In cold regions, persistent snow and ice accumulation on asphalt pavements poses a significant threat to traffic safety. Chloride-based deicing salts are widely used due to their effectiveness and low cost; however, infiltrated chloride ions interact with moisture and undergo repeated erosion–freeze–thaw cycles (EFTC), leading to accelerated pavement deterioration and reduced service life. Existing reports on chloride-induced EFTC damage primarily focus on salt concentration and freeze–thaw cycles, while the influence of asphalt type and material properties remains insufficiently investigated. In this study, the deterioration behavior of different asphalt concretes is examined under varying chloride concentrations, and an XGBoost model optimized using the snow geese algorithm (SGA) is developed to predict pavement damage variables (<InlineEquation ID="IEq1"> <EquationSource Format="MATHML"><math> <mi>D</mi> <mi>V</mi> </math></EquationSource> <EquationSource Format="TEX">$DV$</EquationSource> </InlineEquation>) after chloride exposure. SHAP values and partial dependence plots (PDP/ICE) are employed to interpret feature importance and interaction effects. The results indicate that <InlineEquation ID="IEq2"> <EquationSource Format="MATHML"><math> <mi>D</mi> <mi>V</mi> </math></EquationSource> <EquationSource Format="TEX">$DV$</EquationSource> </InlineEquation> increases with cycle number and chloride concentration due to ice expansion, salt crystallization, and chemical corrosion. 70# and 90# asphalt concretes exhibit pronounced deterioration, whereas SBS-modified asphalt concrete demonstrates superior resistance to EFTC. The optimal model configuration achieves a favorable balance between prediction accuracy and computational efficiency. Interpretability analyses identify exposure time and chloride concentration as the dominant predictors of <InlineEquation ID="IEq3"> <EquationSource Format="MATHML"><math> <mi>D</mi> <mi>V</mi> </math></EquationSource> <EquationSource Format="TEX">$DV$</EquationSource> </InlineEquation>, while softening point exhibits a minor negative influence and penetration and ductility show marginal contributions. A graphical user interface (GUI) is also developed to support engineering applications. Overall, the proposed framework provides a reliable and interpretable tool for pavement maintenance and rehabilitation decision-making in cold regions.</p>

Interpretable XGBoost model-based prediction of pavement damage after deicing salt exposure considering asphalt properties and residual chloride concentration

期刊论文
发表日期：

2026年4月2日

查看全文

Shiyuan Li,
Shengkang Zhang,
Huining Xu,
Xinxing Bian,
Ying Qu

摘要

In cold regions, persistent snow and ice accumulation on asphalt pavements poses a significant threat to traffic safety. Chloride-based deicing salts are widely used due to their effectiveness and low cost; however, infiltrated chloride ions interact with moisture and undergo repeated erosion–freeze–thaw cycles (EFTC), leading to accelerated pavement deterioration and reduced service life. Existing reports on chloride-induced EFTC damage primarily focus on salt concentration and freeze–thaw cycles, while the influence of asphalt type and material properties remains insufficiently investigated. In this study, the deterioration behavior of different asphalt concretes is examined under varying chloride concentrations, and an XGBoost model optimized using the snow geese algorithm (SGA) is developed to predict pavement damage variables ( $D V$ $DV$ ) after chloride exposure. SHAP values and partial dependence plots (PDP/ICE) are employed to interpret feature importance and interaction effects. The results indicate that $D V$ $DV$ increases with cycle number and chloride concentration due to ice expansion, salt crystallization, and chemical corrosion. 70# and 90# asphalt concretes exhibit pronounced deterioration, whereas SBS-modified asphalt concrete demonstrates superior resistance to EFTC. The optimal model configuration achieves a favorable balance between prediction accuracy and computational efficiency. Interpretability analyses identify exposure time and chloride concentration as the dominant predictors of $D V$ $DV$ , while softening point exhibits a minor negative influence and penetration and ductility show marginal contributions. A graphical user interface (GUI) is also developed to support engineering applications. Overall, the proposed framework provides a reliable and interpretable tool for pavement maintenance and rehabilitation decision-making in cold regions.