Physics-guided explainable machine learning for predicting temporal scour evolution around hybrid spur dikes under unsteady flood hydrographs

Accurate prediction of local scour around river-training structures during flood events remains a major challenge because scour evolution is governed by highly nonlinear interactions among hydraulic conditions, structural permeability, sediment transport, and time-dependent bed morphology. Although machine learning has recently improved scour prediction, existing studies have primarily focused on equilibrium scour under steady flow conditions and have largely treated predictive models as black boxes, limiting physical interpretation. This study develops an explainable physics-guided machine learning framework for predicting the temporal evolution of normalized scour depth around hybrid spur dikes (Bandal-like) under unsteady flood hydrographs. A laboratory dataset comprising 1,308 observations collected under nine hydrograph scenarios with different peak discharges, hydrograph durations, and permeability ratios was used to train and evaluate four regression algorithms, including Support Vector Regression, Random Forest, Gradient Boosting Regression, and CatBoost. Hyperparameters were optimized using GridSearchCV with five-fold cross-validation, and a physics-guided cumulative-maximum constraint was incorporated to maintain temporal consistency of the predicted cumulative scour envelope by preventing nonphysical decreases. Among the investigated models, CatBoost achieved the highest predictive performance on the independent testing dataset, with \({R}²=0.9110\) , RMSE = 0.0504, and MAE = 0.0343, and successfully reproduced the principal characteristics of temporal scour evolution across the investigated flood hydrographs. Explainable artificial intelligence analyses using CatBoost Feature Importance, Permutation Feature Importance, and SHAP consistently identified relative time ( \(t/{T}b\) ) as the dominant predictor, followed by structural permeability ( \({R}ₐ\) ), while SHAP interaction analysis revealed nonlinear and condition-dependent relationships among hydrograph position, relative discharge, flow depth, and structural permeability. In particular, the predictive contribution of relative discharge became more pronounced near peak-flow conditions, while the influence of permeability varied with hydraulic loading and hydrograph stage. The proposed framework combines predictive performance with interpretable model-based insights and physics-guided temporal consistency, providing a complementary approach to conventional scour assessment within the investigated experimental domain.