Exploring the Behaviour of Mega-Nourishment Interventions and Predicting Morphological Changes Using Artificial Intelligence and with Interactive Apps

Kumar, Pavitra ORCID: https://orcid.org/0000-0002-4683-724X and Leonardi, Nicoletta (2023) Exploring the Behaviour of Mega-Nourishment Interventions and Predicting Morphological Changes Using Artificial Intelligence and with Interactive Apps. In: AGU Fall Meeting 2023, 11 December 2023 - 15 December 2023, San Francisco, USA. (Unpublished)

Accepted Version File not available for download. Available under License In Copyright. Download (149kB)

Abstract

There is a growing trend towards implementing nature-based solutions in various engineering domains. Coastal protection, for example, has seen the implementation of nature-based solutions such as mega-nourishments like sand engines. However, there are lot of uncertainties in their long-term behaviour and response to diverse coastal conditions which are yet to be explored. The behaviour of mega-nourishment interventions can be explored with the help of Artificial Intelligence (AI), specifically deep learning. In this study, application of Artificial Neural Network (ANN) is tested in combination with fully coupled hydrodynamics and morphological model (Delft3D) for predicting morphological changes and understanding the behaviour of Sand Engine. Two frameworks (Sand Engine App and Sand Engine Surface) are proposed which support the choice of coastal protection schemes through the synthesis of numerical modelling outputs into ANN models. For the Sand Engine app, 12 ANN ensemble models were trained on the simulated data of sand engines, placed at different locations along Morecambe Bay (UK), to predict the impact of different sand engines on water depth, wave height, and sediment transport in its surrounding area when subjected different wave forcings. These ensemble models provided good performance with majority of the models having testing regression greater than 0.9. For the Sand Engine Surface app, Long Short-Term Memory (LSTM) models were trained on simulated data of a simplified case, with sand engine placed at general coastline, to predict volume of sand remaining and bathymetry evolution of sand engines over time (for more than a year). The LSTM network was customized to include 172 parallel LSTM cells for prediction of complete time series based on feature inputs (sand engine configurations and coastal conditions). Models provided good accuracy with testing regression for bathymetry evolution greater than 0.7 and 0.9 for volume of sand remaining. Both the Sand Engine App and Sand Engine Surface app, designed in MATLAB, include their respective models and simulation results. These applications provide prediction outcomes along with the simulation results, allowing easy comparison and analysis.