Research Outputs

Abstract

In recent years, coastal management has been facing new challenges: socio-economic growth and consequent climate change impose new boundary conditions pushing coastal systems towards unseen states. For adaptation and mitigation strategies as well as risk management, the resilience of systems to these projected changes must be tested and quantified using predictive tools, given the scarcity of observations. Process-based models, which limit the number of assumptions, are the preferred tools. However, these models are computationally expensive and therefore unattractive for global sensitivity and uncertainty analyses. Input and model reduction techniques, as well as behavioural empirical models, have been widely used to overcome these computational difficulties. In this paper, we propose a process-based hybrid workflow—that combines statistical and machine learning with a process-based numerical model—to provide sensitivity analyses on complex systems. As an example we explore salt intrusion in estuaries. The novelty of the method presented is the implementation of an adaptive sampling technique of numerical experiments with a process-based hydrodynamic model, and the training of a neural network to augment the set of numerical runs executed. The first uses predictive uncertainty to automatically explore the response of the complex system to varying environmental boundary conditions and geomorphological configurations. The second is trained to provide system responses around the sampled points. This exploration is closed by simulating the extremes in the output space as found by a genetic algorithm. This scheme is shown to be highly efficient in non-linear, heteroscedastic, and highly non-stationary systems.