Machine learning predictions for local scour around complex reinforced concrete piers

Accurately predicting flood-induced local scour around complex reinforced concrete (RC) piers remains challenging, as many existing studies rely on simplified single-column assumptions. Although recent models have sought to address complex pier geometries, they frequently fail to incorporate sufficiently influential parameters, thereby limiting their predictive accuracy and applicability. In this study, three advanced machine learning (ML) techniques – Gaussian process regression, ensemble learning and an artificial neural network (ANN) – integrated with Bayesian optimisation for hyperparameter tuning were employed. A dataset of 336 clear-water experiments, encompassing a wide range of hydraulic and structural conditions, was compiled for model training and validation. Model performance was evaluated using ten-fold cross-validation and a test set. Among the models, the ANN demonstrated the highest performance on the test set, achieving a root mean square error of 0.134, mean absolute error of 0.093, mean absolute percentage error of 17.3% and a coefficient of determination of 0.871, significantly outperforming existing empirical formulas and previously established ML approaches. Additionally, Shapley additive explanations were used to interpret feature contributions, revealing that flow-related parameters had the greatest influence, followed by sediment characteristics and structural geometry. These findings provide both accurate predictive tools and physical insight into the scour process around complex RC piers.