Artificial intelligence for hurricane storm surge hazard assessment. (1st February 2022)

• Record Type: Journal Article
• Authors: Ayyad, Mahmoud; Hajj, Muhammad R.; Marsooli, Reza
• Is Part Of: Ocean engineering. Volume 245(2022)
• Publication Date: 2022-02-01
• Subjects: Artificial Neural Network -- Storm surge -- Tropical cyclones -- Return period; Ocean engineering -- Periodicals; Ocean engineering; Periodicals; 620.4162
• DOI: 10.1016/j.oceaneng.2021.110435
• Languages: English
• ISSNs: 0029-8018

Abstract

Abstract: Risk-informed coastal management requires assessment of extreme flood hazards from a large number of storm scenarios. To account for impact of climate change based on potential variations in greenhouse gas concentration and climate models, the number of storm scenarios would be even larger. Although physics-based hydrodynamic numerical models could predict flood levels and their impact from storm scenarios, the high computational cost of the solutions hinders the ability to perform the required number of simulations. Towards alleviating that cost, we show that physics-based simulations can be combined with Artificial Neural Network models to support more faster and effective prediction of low-probability events that account for uncertainties associated with climate change. We show this capability by predicting 10, 100, and 1, 000 years return periods for peak storm surge height at a specific location on an idealized coastline. A large data set of synthetic tropical cyclones is generated from physics-based simulations and used for training, validating and testing the constructed neural network model. The ANN predicted values are validated against values from the physics-based simulations. The advantage of the combined approach is that, once the training was complete, it was performed in a fraction of the time required for the physics-based simulations. Highlights: We assess coastal flood hazards from a large number of storm scenarios. We perform large number of physics-based simulations to generate a large data set. We used the generated data set to train the artificial neural network model. We expand ANN capability by combining physics-based simulations with deep ANN model. We show that ANN is a cost-effective approach to predict low-probability flood hazards.