Optimization of wave breaking criterion for Boussinesq-type model on coral reef terrain based ON BP neural network

Wave breaking on coral reefs, characterized by steep fore-reef slopes, poses a significant challenge for Boussinesq-type models, as conventional breaking criteria often fail. While machine learning is increasingly applied to predict specific outputs like wave height, this study introduces a more fundamental methodology by optimizing the model's internal breaking criterion (${\gamma }_b$). This approach enhances the model's core physics, enabling a more robust and physically realistic simulation of the entire wave transformation process. This study introduces a novel approach to dynamically predict the breaking criterion for the FUNWAVE-TVD model by employing a back-propagation (BP) neural network. The network was trained on a dataset generated from 66 high-resolution numerical simulations using the FUNWAVE-TVD model. Five key parameters—offshore water depth (h), incident wave height (H₀), wave period (T), reef flat water depth (h_r), and fore-reef slope ($\tan \alpha$)—were used as inputs to predict the optimal ${\gamma }_b$ value.Validation across four distinct experimental scenarios demonstrates that this method significantly improves simulation accuracy, with the goodness-of-fit (R²) increasing from an average of 0.63–0.87. This refined approach not only enhances the simulation of wave breaking over complex topographies but also provides a novel technical pathway to support safety assessments in island reef engineering.