The armor layer of a breakwater dissipates wave energy through friction and percolation,thereby reducing wave overtopping.Most existing overtopping calculation methods employ a constant roughness coefficient related to armor block type,which fails to account for the influence of different structural parameters of the armor layer.To evaluate the rationality of using constant roughness coefficients as input in machine learning models for wave overtopping prediction,this study utilizes the overtopping database from the EU CLASH project.Data for single-slope breakwaters (including simple slopes,slopes with crown walls,slopes with berms,and composite slopes) encompassing overtopping,wave,and structural parameters are selected to develop a neural network model for fitting armor layer roughness coefficients.By comparing the model performance across different structural types,the applicability of the recommended roughness coefficients in the database for neural network-based overtopping prediction is assessed.The results reveal significant variations in model accuracy depending on the breakwater slope structure,indicating substantial limitations and uncertainties in the current practice of using constant roughness coefficients.To further enhance the precision of neural network models for overtopping prediction,input parameters should incorporate armor block characteristics that directly influence hydrodynamic performance.