Interpretable machine learning for bridge-pier scour prediction and flood resilience

Abstract

Bridge-pier scour is a leading cause of flood-induced bridge failure, yet practice still lacks transparent, physics-informed tools that link data-driven prediction with design guidance. This study develops an interpretable, physics-aware machine-learning framework to predict equilibrium scour depth and translate those predictions into actionable strategies for flood-resilient infrastructure. Using the 2014 U.S. Geological Survey Pier-Scour Database (569 laboratory cases), five models: Gradient Boosting, AdaBoost (Tree), XGBoost, Gaussian Process (RBF kernel), and Kernel Ridge (polynomial), were trained and evaluated with K-fold cross-validation. Model performance was evaluated using R2, RMSE, and MAE. Gradient Boosting performed best, achieving training and testing R2 of 0.99 and 0.96, a near-ideal parity fit, and consistent accuracy across folds. Interpretability is provided by SHAP, whose attributions align with hydraulics; the pier width normal to flow accounts for 70.6% of the total importance in predicting scour depth. Predicted scour is mapped to four scenario envelopes that capture rare, peak, and sustained hydraulic extremes and yield clear design checks for flood resilience. A physics-based imputation scheme for sediment critical velocity and duration of flow is integrated in the framework so that missing inputs are handled in a hydraulically consistent way. The developed models are deployed in an interactive web app, allowing practitioners to obtain code-free scour predictions across all learners. Applied to the Knik River bridge and benchmarked against related work, the framework improves accuracy and provides actionable margins for design verification, maintenance prioritization, retrofit planning, emergency response, and transparent risk communication.