Online climate change projections report 3.1 Introduction
It is clear from Chapter 1 and Chapter 2 that future climate over the UK (and elsewhere) will be influenced by an array of factors. Some of these affect external forcing of climate through changes to the Earth’s radiation balance resulting from natural changes (e.g. volcanic eruptions or variations in solar output) or man-made changes (emissions of greenhouse gases, aerosols and their precursors), while others affect physical and biogeochemical feedback processes which enhance or reduce the response to this forcing. In addition, internal climate variability exerts a significant influence on climate, in addition to the effects of forced changes. All of these factors introduce uncertainty into projections of future climate because none of them can be predicted perfectly. This is due, in general, to imperfect knowledge of either the detailed behaviour or the current observed states of the relevant systems.
We currently have no agreed method of quantifying the relative likelihood of alternative pathways for future man-made emissions (Section 2.4). For UKCP09, we therefore focus on the task of estimating distributions of future changes in climate for each of three specific emissions scenarios ( A1FI, A1B and B1, explained in Section 2.4 and Annex 1, and referred to elsewhere in UKCP09 as High, Medium and Low). These scenarios assume no future changes in natural external forcing, apart from a prescribed repetition of the 11 year cycle of solar insolation based on past observations. Regional climate changes in response to these emissions will be determined by complex interactions between a number of Earth System processes, plausible projections of which require the use of detailed three-dimensional global climate models (GCMs). As discussed in Section 2.3, ensemble approaches provide an obvious method of exploring the uncertainties associated with GCM projections. Multi-model ensembles (MMEs, e.g. Meehl et al, 2005), constructed by pooling projections from alternative GCMs developed at different modelling centres, provide a valuable indication of the range of possible future changes. However, stakeholders faced with climate-sensitive policy and adaptation decisions will typically require more than a simple specification of a possible range (Pittock et al. 2001). This is widely recognised in the climate science community, and consequently methods have been suggested to derive probability distributions for regional changes from MME results (e.g. Tebaldi et al. 2005; Greene et al. 2006; Furrer et al. 2007; Watterson, 2008), giving estimates of the relative probability of different future outcomes within the envelope of possible changes. Motivations for such approaches stem from results showing that combining projections from different models can increase the skill of historical climate simulations (e.g. Reichler and Kim, 2008) or seasonal forecasts (e.g. Hagedorn et al. 2005), because the errors in different models are partially independent. Furthermore, the models are assembled from a large pool of alternative components, thus sampling to some extent the effects of variations in basic structural assumptions such as choice of model grid, numerical integration scheme or the fundamental physical assumptions employed in the parameterisation of sub-grid scale processes such as convection, boundary layer transports, cloud and precipitation formation, etc (see Box 2.1). However, multi-model ensembles are rather small in size, consisting typically of 10–20 models, some of which might be run several times from different initial states. Also, the set of models is assembled on an opportunity basis, not being designed to sample systematically some underlying space of possible model formulations (Allen and Stainforth, 2002). This creates the need for substantial assumptions in converting their results into estimated probabilities for climate change, essentially because it is not clear how to identify a distribution of possible outcomes of which the MME is a sample. Different studies address this issue in different ways, and therefore generate significantly different results (see Tebaldi and Knutti, 2007).
Another issue is that probabilistic projections are conditional on the set of uncertainties sampled in the ensemble simulations. In order to provide a credible basis for decision making, a critical prerequisite is that these are designed to sample all sources of uncertainty known to be likely to exert a significant influence on climate over the time frame of interest (here, the 21st century). For a given scenario of future emissions, these would include internal climate variability and uncertainties in atmospheric and oceanic processes, which give rise to different realisations of 21st century climate in the latest MME produced for the IPCC AR4 (Figure 2.5). However additional sources of uncertainty, notably carbon cycle feedbacks (Box 2.1) and the uncertainty in downscaling GCM simulations to local scales, also need to be considered. In order to produce probabilistic projections for UKCP09, we have therefore developed a new approach aimed at sampling the key uncertainties systematically, using a purpose-built set of ensemble simulations involving several different configurations of the HadCM3 climate model.
The method is based on the notion of the perturbed physics ensemble (PPE), in which alternative variants of a single GCM are created by altering the values of uncertain model parameters (Murphy et al. 2004; Stainforth et al. 2005). These parameters control important small scale processes in the model (such as the formation and precipitation of cloud droplets, the reflectivity of sea ice or the transfer of heat, moisture or momentum between the surface and the atmosphere), and are uncertain because we lack sufficiently detailed observations or sufficiently precise theoretical understanding to constrain their values accurately. A major advantage is that PPEs can be designed to ensure that all the key process uncertainties are sampled in a manner consistent with current scientific understanding. This is achieved by asking experts to identify which model parameters control the key processes, and then to specify distributions for the chosen parameters, consistent with the present state of knowledge concerning the identified processes. We can then construct a set of ensemble runs which select alternative values of the parameters drawn from these distributions, ensuring that the relevant uncertainties are well sampled.
The PPE approach therefore facilitates the construction of probabilistic projections consistent with current understanding of model uncertainties (Section 3.3), and it is also possible to test the sensitivity of the results to reasonable variations in the definition of the space of possible model variants implied by the specified distributions for model parameters (see Annex 2). However, the model on which the PPE is based (in our case HadCM3) will inevitably contain some structural errors in its physical representation of the real climate system, which cannot be resolved by varying the model parameters (Murphy et al. 2004). These structural errors determine how informative the model simulations are about the real system, so it is critical to account for the additional uncertainty implied by their presence (Goldstein and Rougier, 2004). We address this by using our PPE results to predict the results of members of a multimodel ensemble developed at other modelling centres, and containing structural assumptions partially independent of HadCM3. This allows us to estimate the effects of structural errors (subject to assumptions discussed in Section 3.2.8), and to present probabilistic projections which combine information from both perturbed physics and multi-model ensemble results.
The methodology is described in Section 3.2, this being a somewhat abridged (though also updated) version of that given by Murphy et al. (2007). Section 3.3 provides a brief summary of key strengths and limitations of our approach, and a discussion of how the probabilistic climate change estimates it provides for UKCP09 should be interpreted by users. The robustness of these estimates to plausible variations in key assumptions is discussed in Annex 2.