Abstract:
The focus of this study is to respond to climate change-induced risk fluctuations and probabilistically quantify the relevant impacts on the resilience of coastal and inland environments, using advanced Bayesian machine learning algorithms. For the coastal environments, we focus on the Mangrove Forests, which are known to provide excellent protection against erosion by dampening incident wave energy. In order to understand and quantify to what extent rehabilitation of coastal mangrove forests can enhance resilience against climate change impacts, we need to efficiently model the hydro-morphodynamics across mangrove forests. On the other hand, due to acceleration of the effects from climate change over the last 3 decades, the inland areas have become more prone to extreme flooding. An accurate predication of flood inundation depth and extent is crucial to plan for an effective flood management scheme to enhance the flood resilience of urban areas against the extreme climatic events.
Modelling and simulation of the hydro-morphodynamics of mangrove environments and spatio-temporal outputs of a 2D inland flood model are computationally very expensive, given the nonlinear and complex nature of these problems. In order to tackle these challenges, we developed advanced machine learning methods, including physics-informed neural networks and Gaussian process (GP) emulation, which have recently gained much attention in addressing the climate and geophysical modelling problems due to their ability in providing fast and accurate results, while preserving the binding physics laws and requiring small amounts of data from complex mathematical evaluations, usually represented by Navier-Stokes PDEs.

This study explores appropriateness and robustness of GP models to emulate the results from a hydraulic inundation model. The developed GPs produce real-time predictions based on the LISFLOOD-FP simulation output. Furthermore, we build a depth-averaged hydro-morphodynamics model using a finite element model, which can be used for simulating coastal processes. Navier Stokes, the underlying equation to solve for fluid dynamics, is used as the governing equation that constrains the neural network to respect the conservation of mass, energy, and momentum. Then, the latest development of applying physics-informed neural networks to quantify the ability of mangrove environments in attenuating waves and preventing erosion will be illustrated.
The enhanced prediction performance and computational power of the proposed methods will be demonstrated for two complex real-world case studies, including the modelling of hydro-morphodynamics at the Sundarbans, the largest mangrove forest in the world located between India and Bangladesh, which is regularly subjected to tropical cyclones and other extreme climatic events; and modelling the worst flood recorded in history for Tadcaster, a riverside town located between York and Leeds in the Northeast of England, occurred in 2015.
 
Keywords: Climate modelling; Flood; Erosion; Gaussian Process; Physics-informed neural networks.
 

Speaker: Professor Alireza Daneshkhah Coventry University

Krisitian Romano (University of Warwick) for RSS West Midlands Local Group