Interpretable machine learning scheme for predicting bridge pier scour depth

Predicting scour depth around bridge piers is challenging due to the involvement of various uncertain factors and complex processes, making it difficult to obtain accurate results using traditional deterministic models. Recently, nonlinear analysis and prediction using machine learning (ML) techniques that glean statistical structures from input/output data have received substantial attention. With the use of ML, however, interpretation becomes difficult with increased complexity and diversity of the analysis parameters. The study herein proposes a local scour depth prediction model around piers using interpretable ML. A scour depth prediction model was constructed using the state-of-the-art eXtreme Gradient Boositng (XGB) model, and Shapley additive explanations (SHAP) based on the cooperative game theory. The ML model demonstrated excellent prediction accuracy (R2: 0.75, RMSE: 0.23) compared with existing empirical formulas used for assessment of pier scour. Model analysis results demonstrated that for each input variable, the distribution of SHAP values for scour depth is consistent with the general theoretical knowledge about the factors influencing scour depth, suggesting that the predictions of the XGB model are reasonable. Considering accuracy and conservatism, design recommendations using two statistical parameters: mean absolute percentage error (MAPE) and the percentage of predicted scour depth exceeding measured scour depth, referred to as the "level of conservatism" were proposed as a means of providing the flexibility of including a target level of accuracy and conservatism in the computed scour depth.