Predicting coastal wave overtopping is a significant challenge, exacerbated by climate change, increasing the frequency of severe flooding and rising sea levels. Digital twin technologies, which utilise artificial intelligence to mimic coastal processes and dynamics, may offer new opportunities to predict coastal wave overtopping and flooding reliably and computationally efficiently. This study investigates the effectiveness of training various artificial intelligence models using wave buoy, meteorological, and recorded coastal wave overtopping observations to predict the occurrence and frequency of overtopping at 10-minute intervals. These models have the potential for future large-scale global applications in estimating wave overtopping and flood forecasting, particularly in response to climate warming. The model types selected include machine-learning random forests, extreme gradient boosting, support vector machines, and deep-learning neural networks. These models were trained and tested using recorded observational overtopping events, to estimate wave overtopping and flood forecasting in Dawlish and Penzance (Southwest England). The random forests performed exceptionally well by accurately and precisely estimating coastal wave overtopping and non-overtopping 97% of the time within both locations, outperforming the other models. Moreover, the random forest model outperforms existing models, such as OWWL, for estimating overtopping. This research has profound implications for increasing preparedness and resilience to future coastal wave overtopping and flooding events by using these random forest models to predict overtopping and flood forecasting on wider global and climate scales. These trained random forests are significantly less computationally demanding than existing process-based models and can incorporate the important effect of wind on overtopping, which was neglected in existing empirical approaches.

File Attachments The attached files indicate the supplementary material for this study, including the code for constructing the AI models, the datasets for training the models, and some additional tables computing the performance metrics and model specifications. For information about the model hyperparameter tuning metrics, see "Appendix A." For code for constructing the neural network, random forest, and XGBoost, see "Neural Network Binary Script," "Random Forest Binary Script", and "XGBoost Binary Script," respectively. The training datasets for Dawlish and Penzance can be found under "Dawlish Complete Dataset" and "Penzance Complete Dataset," respectively. For information regarding the random forest model weaknesses regarding false negatives and positives, see "Penzance False Negatives," "Penzance False Positives," "Dawlish False Positives," and "Dawlish False Negatives." For information about how the figures were coded, see "Code for each graph."