Applying Gaussian process emulators for coastal wave modelling
Applying Gaussian process emulators for coastal wave modelling.
Malde, S.and Tozer, N.P. and Oakley, J. and Gouldby, B.P. and Liu, Y. and Wyncoll, D.
Coastal Engineering Journal . (Submitted) (2018)
Full text not available from this repository.
|Abstract:||In an ever increasing need to minimise costs, most coastal engineering work requires reliable information on the environmental forces and governing physical processes that need to be taken into account when designing or assessing the performance of coastal structures. This often involves the application of complex physical process based numerical models (simulators) that can be computationally expensive to run. In general, the more complex the process representation, the more computationally expensive the model simulations will be. Coastal process simulations, flood risk analysis and flood forecasting can all require substantial computational resources. The situation is exacerbated when repeated simulation runs are required in a sensitivity analysis or uncertainty analysis, for example.
When the number of model simulations are excessive, it is possible to make a compromise on time and space resolution or processes represented or apply alternative methods typically with a reduction in accuracy. One such relatively simply approach is populate or train a look-up table (LUT) using only a subset of the full set of runs and then use the LUT to predict the output for the relevant input values typically using linear interpolation. A limitation with such approaches is the number of training simulations required to maintain accuracy by adequately representing the input parameter space increases significantly with increasing dimension. In addition, the relationship between the inputs and outputs can often be non-linear and hence linear interpolation is not necessarily an appropriate interpolation method to apply.
This paper focusses on the use of the Gaussian process emulator (GPE) meta-modelling approach as an alternative approach to traditional LUTs. Using the specific example of wave transformation with the Simulating Waves Nearshore (SWAN) wave model, a GPE has been compared with a traditional LUT approach. In addition, the method of selecting the design points used to train the GPE has been explored and a refined algorithm, that takes account prior knowledge of the boundary conditions, has been introduced.
It is shown that the GPE approach requires significantly fewer model runs to obtain similar or higher accuracy, enabling a substantial reduction in overall computation time when compared to a traditional LUT approach. The refined algorithm also shows significant computational efficiencies, meaning that potential compromises on model resolution or the physical processes can be limited.|
|Uncontrolled Keywords:||Gaussian process emulator (GPE); SWAN modelling; computer modelling; maximum dissimilarity algorithm (MDA); look-up table (LUT); design selection; weighted design selection|
|Subjects:||Coasts > General|
|Deposited On:||11 Oct 2017 07:41|
|Last Modified:||06 Feb 2018 13:26|
Repository Staff Only: item control page