Abstract

Bridge pier scour is one of the primary causes of bridge foundation failure, making accurate prediction of final scour depth essential for the safe and economical design of hydraulic structures. This study develops interpretable ensemble machine learning models for predicting final scour depth around collar-protected circular bridge piers using a laboratory dataset comprising 48 experimental observations. Four dimensionless hydraulic and geometric parameters, namely (b/bc, b: bridge pier diameter, bc: collar diameter), (z/d50, z: collar elevation, d50: median diameter of bed particles), (z/bc), and (U/Uc, U: time average velocity, Uc: critical velocity of bed particles), were employed as model inputs, while the normalized final scour depth (dsf/bc) was considered as the target variable. Three ensemble learning algorithms, namely XGBoost (XGB), Random Forest (RF), and Extra Trees (ET), were developed and evaluated using training, testing, and 5-fold cross-validation procedures. The predictive performance of the models was assessed using the coefficient of determination (R2), Kling–Gupta Efficiency (KGE), root mean square error (RMSE), and mean absolute error (MAE). Among the investigated models, XGBoost demonstrated the highest predictive accuracy, achieving an R2 of 0.987, a KGE of 0.982, and an RMSE of 0.0117 on the training dataset. Furthermore, SHapley Additive exPlanations (SHAP) were employed to interpret the influence of individual input variables, revealing that hydraulic intensity and collar-related parameters exert the greatest influence on equilibrium scour prediction, consistent with established scour mechanics. The proposed framework combines high predictive accuracy with model interpretability, providing a reliable decision-support tool for bridge scour assessment and demonstrating the potential of explainable machine learning to support the design and management of scour protection measures under controlled hydraulic conditions.

IPC Classification

Keywords