Interpretability-enhanced temporal surrogate modeling for wave-induced scour vulnerability assessment of vertical-wall structures

Scour compromises the stability and safety of vertical-wall structures, necessitating a systematic method to assess scour vulnerability. However, two key vulnerability elements remain unresolved, namely suitable grading thresholds and an efficient temporal scour surrogate model. This study utilizes machine learning to address these vulnerability elements using 23,215 wave-induced temporal scour depths, obtained from full-scale numerical simulations in front of vertical-wall structures. First, three clustering algorithms are adopted to derive three grading thresholds based on the relative scour depth. Then, the temporal scour surrogate model is developed by comparing six supervised learning algorithms. Subsequently, the SHapley Additive exPlanations method is utilized to interpret and visualize the feature contributions to model predictions. Furthermore, time-dependent vulnerability assessment is conducted by integrating the surrogate model with Monte Carlo simulations. The derived grading thresholds obtained from Agglomerative Clustering are 0.0361, 0.0726, and 0.1432, quantifying the scour severity of vertical-wall structures. The temporal scour surrogate model trained using Residual Multi-Layer Perceptron exhibits the optimal generalization performance in substituting nonlinear scour processes. The model interpretability analysis indicates physical consistency with the scour development process, while the evolution of fragility curves illustrates the progressive escalation of vulnerability with increasing wave intensity and duration.