Machine learning methods for predicting shoreline change from submerged breakwater simulations

The design of submerged breakwaters for coastal protection requires many iterations of process-based simulations to determine combinations of input variables’ influence on the consequent shoreline response. By using machine learning, the complex non-linear multidimensional relation between input and modelled shoreline output can be modelled to predict shoreline change on unseen combinations of inputs with much faster simulations. To prove this, a neural network and Gaussian process regressor were trained on a dataset of 243 combinations of structural and wave forcing input parameters paired with their process-based modelled shoreline change output on a straight shoreline after one year. The models were capable of fitting the dataset of input–output relations comprising outputs of accretion and erosion, allowing for predictions on incremental input changes. The proposed hybrid modelling approach takes advantage of the computational advantages of machine learning to leverage physics-driven output data generated from process-based models to optimise the design options of submerged breakwaters relative to shoreline outcomes. Given the multidimensional multiparameter processes, the solution space is large, however, by focusing on key parameters of interest, resource-intensive processes can be optimised to collect high-quality data for important site parameters and using this machine learning method to develop a shoreline change model. • Machine learning applied to shoreline model data to predict input–output relation. • Neural network and Gaussian process regressor regressed a multidimensional dataset. • The Gaussian process regressor was found to be more effective on a sparse dataset. • Application of Monte Carlo-based optimisation using machine learning predictions.