Emulation Engines: Choice and Quantification of Uncertainty for Complex Hydrological Models

Complex, mechanistic hydrological models can be computationally expensive, have large numbers of input parameters, and generate multivariate output. Model emulators can be constructed to approximate these complex models with substantial computational savings, making activities such as sensitivity analysis, calibration and uncertainty analysis feasible. Success in the use of an emulator relies on it making accurate and precise predictions of the model output. However, it is often unclear what type of emulation approach will be suitable. We present a comparison of reduced-rank, multivariate emulators built upon different ‘emulation engines’ and apply them to the Australian Water Resource Assessment System model. We examine first-order and second-order approaches which focus on specifying the mean and covariance, respectively. We also introduce a nonparametric approach for quantifying the uncertainty associated with the emulated prediction where this has bounded support. Our results demonstrate that emulation engines based on second-order approaches, such as Gaussian processes, can be computationally burdensome and may be comparable in performance to computationally efficient, first-order methods such as random forests.Supplementary materials accompanying this paper appear online.