Adhesive failure in the interfacial transition zone caused by aggregate particles is a major contributor to various forms of distress in asphalt mixtures.Accurate prediction of such failure holds significant research ...Adhesive failure in the interfacial transition zone caused by aggregate particles is a major contributor to various forms of distress in asphalt mixtures.Accurate prediction of such failure holds significant research value.However,conventional predictive models often struggle with the complex,highly nonlinear relationships between input variables and the damage indicator,and they generally lack interpretability—hindering understanding of the prediction process.This study introduces the Extremely Randomized Trees(Extra Trees)algorithm,which demonstrates superior performance in modeling non-linear relationships and offers strong interpretability.By applying Extra Trees,highly accurate and stable damage predictions were achieved(coefficient of determination,R2=0.984;mean absolute percentage error=8.14%).Model interpretability was further explored through correlation analysis,feature importance evaluation,and partial dependence plots,with results validated via ablation experiments.Additionally,the model's sensitivity to dataset size was investigated.Experimental findings confirm that Extra Trees significantly outperforms other algorithms in terms of both accuracy and stability.Its treebased structure also facilitates a deeper understanding of the roles of input features in the damage prediction process.展开更多
基金supported by the National Natural Science Foundation of China(Grant No.52478449)the Basic Research Priorities Program of Jiangsu(Grant No.BK20232036)the Natural Science Foundation of Jiangsu(Grant No.BK20221468)。
文摘Adhesive failure in the interfacial transition zone caused by aggregate particles is a major contributor to various forms of distress in asphalt mixtures.Accurate prediction of such failure holds significant research value.However,conventional predictive models often struggle with the complex,highly nonlinear relationships between input variables and the damage indicator,and they generally lack interpretability—hindering understanding of the prediction process.This study introduces the Extremely Randomized Trees(Extra Trees)algorithm,which demonstrates superior performance in modeling non-linear relationships and offers strong interpretability.By applying Extra Trees,highly accurate and stable damage predictions were achieved(coefficient of determination,R2=0.984;mean absolute percentage error=8.14%).Model interpretability was further explored through correlation analysis,feature importance evaluation,and partial dependence plots,with results validated via ablation experiments.Additionally,the model's sensitivity to dataset size was investigated.Experimental findings confirm that Extra Trees significantly outperforms other algorithms in terms of both accuracy and stability.Its treebased structure also facilitates a deeper understanding of the roles of input features in the damage prediction process.
