<p>Bridge pier scour was investigated using advanced machine learning approaches to enhance the prediction of normalized scour depth. Gene expression programming (GEP), three variants of physics-informed neural networks (PINNs), and a baseline artificial neural network were developed and evaluated using the USGS pier-scour database comprising 569 laboratory-scale data. GEP achieved the highest predictive accuracy, with a coefficient of determination R² of 0.883 and the lowest prediction errors, demonstrating strong generalization capability. Among the PINN variants, the GreyBox model achieved a balanced trade-off between physical consistency and predictive accuracy, while PINN-Lite reduced computational cost with moderate precision, and PINN-Melville highlighted the limitations of embedding simplified empirical relationships. Explainable artificial intelligence (XAI) techniques, including SHapley additive explanations, local interpretable model-agnostic explanations, individual conditional expectation plots, and partial dependence plots, were applied to interpret model predictions. Analysis identified velocity ratio and sediment ratio as dominant predictors, with coupled hydraulic-sediment interactions shaping scour responses. The results showed that changes in scour depth mainly depend on how strong the water flow is and how the sediment behaves. The integration of GEP, PINNs, and XAI provides a powerful framework combining predictive accuracy with physical interpretability, offering valuable insights for bridge design, safety assessment, and scour risk management.</p>