Field data-based prediction of local scour depth around bridge piers using interpretable machine learning

Abstract

Local pier scour is one of the leading causes of bridge failure worldwide. It occurs when flowing water generates shear stresses at the water–sediment interface, leading to the erosion of soil particles or mass around the pier foundation. In this study, an efficient and accurate machine learning approach is developed for predicting local scour depth around bridge piers. Initially, the field data from the US geological survey database were preprocessed and divided into training, validation, and test sets. The hyperparameters of the models were then adjusted using Bayesian optimization and 5-fold cross-validation. Among the three machine learning models considered in this study, the eXtreme gradient boosting (XGB) model achieved the highest accuracy, which was significantly higher than those realized by four local scour estimation equations utilized in the study. To improve the interpretability of machine learning as a black-box model, SHapley Additive exPlanations (SHAP) was used to interpret the predictions of the XGB model. Interpretable ML analysis indicated that $y/b_n$ was the most influential factor, aligning with the focus on assessing the scour magnitude. In addition, the machine learning interpretation also indicates that the patterns captured by the XGB model are consistent with the theoretical understanding of factors affecting the local scour, thereby validating that the proposed model achieves reasonable predictions. Finally, the gap between laboratory and field data is explained, and a method to address such a gap is proposed considering accuracy and conservatism levels in the assessed scour atudes.