Maximising the value of computer experiments using multi-method global sensitivity analysis

Abstract:

Global Sensitivity Analysis (GSA) is increasingly recognised as an essential technique for a structured and quantitative approach to the calibration and diagnostic evaluation of environmental models. However, the implementation and interpretation of GSA is complicated by a number of choices that users need to make and for which multiple, equally sensible, options are often available. These choices include in the first place the choice of the GSA method, as well as many implementation details like the definition of the sampling space and strategy. The issue is exacerbated by computational complexity, in terms of both computing time and storage space needed to run the model, which might strongly constrain the number of experiments that can be afforded. While several algorithmic improvements can be adopted to reduce the computing burden of specific GSA methods, in this talk we discuss how a multi-method approach can be established to maximise the information gathered from an individual sample of model evaluations. Using as an example the GSA of a land surface model, we show how different analytical and approximation techniques can be applied sequentially to the same sample of model inputs and outputs, providing complimentary information about the model behaviour from different angles, and allowing for testing the impact of the choices made to generate the sample. We further expand our analysis to show how GSA is interconnected with model calibration and uncertainty analysis, so that a careful design of the simulation experiment can be used to address different questions simultaneously.