Computer models are essential tools in the earth system sciences. They underpin our search for understanding of earth system functioning and support decision-and policy-making across spatial and temporal scales. To understand the implications of uncertainty and environmental variability on the identification of such earth system models and their predictions, we can rely on increasingly powerful Global Sensitivity Analysis (GSA) methods. Previous reviews have characterised the variability of GSA methods available and their usability for different tasks. In our paper we rather focus on reviewing what has been learned so far by applying GSA to models across the earth system sciences, independently of the specific algorithm that was applied. We identify and discuss 10 key findings with general applicability and relevance for the earth sciences. We further provide an A-B-C-D of best practise in applying GSA methods, which we have derived from analysing why some GSA applications provided more insight than others.
IntroductionComputer models are essential tools in the earth system sciences. They underpin our search for understanding of earth system functioning and influence decision-and policy-making at various spatial and temporal scales. For example, computer models of the atmospheric system are used to produce short-term weather forecasts, which inform operational decisions at regional or local scale, or to make long-term projections of the global climate, which forms the basis of the international debate around climate change. Global hydrologic models can now provide a coherent picture of hydrological dynamics across our planet under past, current and potential future conditions (Schewe et al., 2014); while integrated assessment models integrate our climate system with the socio-economic behaviour of society to assess the consequences of future policy scenarios (Stanton et al., 2009). Many other examples of the value of computer models can be made for a variety of earth science areas, from atmospheric circulation (Cotton et al., 1995) to biogeochemical processes in the sea (Soetaert et al., 2000), from mantle dynamics (Yoshida and Santosh, 2011) to tsunamis impacts (Gelfenbaum et al., 2011).A key issue in the development of computer models is that they can quickly exhibit complicated behaviours because of the potentially high level of interactions between their variables, and subsequently their parameters, even