Using GEM-SA to explore wind speed dependencies in HF radar backscatter power spectra

A simulator is used to model and predict the behaviour of a complex physical system. Sensitivity analysis (SA) is sometimes carried out as a way of understanding the physical system. SA is the study of how the output of a simulator is sensitive to changes in the values of an input. Usually SA is carried out using Monte-Carlo methods, but this often requires 1000s of runs (ie 1000s of simulator inputs and corresponding outputs). If one run of the simulator is computationally expensive, we substitute the simulator for an emulator. An emulator is a statistical representation of in the input-output relationship of the simulator. To do SA using an emulator requires usually less than 100 runs of the simulator's computer code, which is a lot less than Monte-Carlo. In this dissertation, we build a Gaussian process emulator using GEM-SA (an emulator building program) from a simulator with a 5-dimensional input and 512-dimensional output. After conducting various diagnostic checks to ensure that the emulator is accurately representing the simulator, we then carry out sensitivity analysis to answer the question at the centre of our investigation: How is HF radar backscatter power spectra (the simulator's output) sensitive to changes in wind speed (one of the simulator's inputs)?