Online climate change projections report Annex 2.4
In Chapter 2, we identify three basic sources of uncertainty in projected climate change, associated with emissions of greenhouse gases, aerosols and their precursors, internal climate variability arising from natural unforced variations in the atmospheric and oceanic circulation, and uncertainty in modelling the forced response to emissions. For a given emissions scenario (in this case SRES A1B, the UKCP09 medium scenario), we consider the relative contributions of internal variability and modelling uncertainty to the total uncertainty expressed in the UKCP09 projections. We consider first an example involving the same variables analysed in Figure A2.1 (i.e. changes to summer and winter temperature and precipitation over the global climate model grid box representing Wales), thus omitting uncertainty arising from the downscaling step of Section 3.2.11, which is considered later. We partition modelling uncertainty into a few components representing key elements of our methodology. These consist of:
- Parameter uncertainty, arising from uncertainties in the values of climate model input parameters that control key physical processes. UKCP09 is based on a comprehensive strategy for sampling parameter uncertainties in the atmospheric component of the HadCM3 climate model, by combining a large ensemble of model simulations with emulation of the outputs of possible model variants for which we do not possess an actual simulation (Section 3.2.3). In addition, we sample parameter uncertainties in ocean and sulphur cycle processes using a more limited strategy based on 17 member ensembles of alternative model variants. We define parameter uncertainty to include all of these sources of uncertainty (including uncertainty arising from emulator error in the case of atmospheric parameters), but note that atmospheric parameters provide the dominant contribution. Our method for the quantification of uncertainties in carbon cycle processes, which we consider under a separate heading below), also contains a substantial contribution from parameter uncertainties associated with terrestrial ecosystem processes in HadCM3C (the configuration of HadCM3 including an interactive carbon cycle).
- Structural uncertainty, which measures the additional uncertainty due to modelling errors which cannot be resolved by varying uncertain parameters in HadCM3 (Section 3.2.8). As a proxy for this, we use information from alternative contemporary climate models, assuming that errors in our ability to predict their historical and future simulations of climate form reasonable estimates of structural errors in the ability of HadCM3 to simulate the real climate system. Note that our strategy estimates the impacts of structural errors in atmospheric processes, but not in ocean transport or sulphur cycle processes.
- Timescaling uncertainty is the uncertainty that arises from the need to predict time-dependent climate responses from the simulations of the equilibrium response to doubled levels of carbon dioxide which form the basis of our strategy for sampling uncertain atmospheric model parameters (see Sections 3.2.4 and 3.2.6). The uncertainties associated with timescaling include the effects of internal variability. We remove these in the analysis below, in order to isolate uncertainties arising from methodological assumptions in our procedure, for example that time-dependent climate changes can be assumed to be linearly related to changes in globally averaged temperature.
- Carbon cycle uncertainty. This is assessed in a separate category because carbon cycle feedbacks (e.g. Friedlingstein et al. 2006) are recognised to give rise to a level of uncertainty in global temperature projections comparable to that due to atmospheric processes. These are sampled by combining 15 perturbed variants of HadCM3C with simulations from an alternative multi-model ensemble of nine coupled climate-carbon cycle models (see Sections 3.2.4 and 3.2.6).
Uncertainty due to internal variability is estimated from long control simulations of members of the PPE_A1B ensemble carried out with no changes to the applied external forcing. We quantify timescaling uncertainty by executing our methodology with parameter and carbon cycle uncertainties removed (by fixing values for all model parameters in all Earth System components to those used in the standard published variants of the relevant HadCM3 configuration), and with the future component of the structural uncertainty set to zero. The component of timescaling uncertainty due to internal variability is then subtracted, in order to isolate the aspects that could potentially be removed by improvements to the methodology in future.
The contributions from parameter, carbon cycle and structural uncertainty are calculated by repeating the probabilistic projections, each time removing one or more of these components (either by fixing relevant parameters to their standard values, or by setting future structural uncertainty to zero), and then comparing the spread of the projected changes for 2070–2099 relative to 1961–1990. For instance, to estimate the increase in spread due to carbon cycle uncertainty we run the projection twice, the first time sampling the carbon cycle parameters as described in Section 3.2.6, and the second time fixing the carbon cycle parameters to their standard values. A limitation of this approach is that the change in spread due to addition of carbon cycle uncertainty depends on which other sources of uncertainty have previously been sampled, as the uncertainties combine in nonlinear ways. For instance, carbon cycle feedbacks (and their associated uncertainties) are larger when temperature changes are high, and only when the other sources of uncertainty are sampled do the temperature changes become large enough for a large carbon cycle feedback. So we run all eight permutations of fixing/sampling parameter, carbon cycle and structural uncertainty (with internal variability and timescaling uncertainties always included). From this set of eight, we have four pairs of runs which can each be used to look at the increase in spread that arises from allowing each of the three types of uncertainty to be sampled rather than kept fixed. Then we take the root-mean-square change in spread, and plot the relative size of the contributions in a pie chart in Figure A2.5. Spread is measured as the distance between the 10 and 90% probability levels of relevant probability distributions.
For the four examples shown in Figure A2.5, parameter uncertainty provides the largest contribution (22–31%). This occurs despite the fact that formal observational constraints have been applied to limit the impact of parameter uncertainties (particularly the dominant contribution from atmospheric model parameters), whereas this is not the case for the other components of uncertainty in Figure A2.5. In fact each of the other components typically adds a significant contribution of its own (in the range 12-27%), and no single source of uncertainty dominates. For winter precipitation no contribution from (the methodological aspects of) timescaling is shown, as the total timescaling uncertainty (i.e. including internal variability) is found to be the same as our independent estimate of internal variability in isolation (derived from model control simulations as described above). While we focus here on contributions to the spread of our probabilistic projections, we stress that each of the elements of the methodology considered in Figure A2.5 (apart from internal variability) can also shift the distributions, thus affecting aspects such as the mean, median or mode. For example adding carbon cycle feedbacks increases the mean projected warming (as well as adding uncertainty), while the mean reduction in summer precipitation projected over much of the UK is ameliorated somewhat by the inclusion of the uncertainty associated with structural model errors, since our projections of the changes simulated by other climate models tend to be too dry.
Figure A2.6 repeats the analysis of Figure A2.5 for an earlier projection period, 2010–2039. This demonstrates the changing role of different contributions to uncertainty at different lead times. In particular, internal variability increases in significance, becoming the largest contribution in three of the four cases. The other components are generally smaller than at 2070–99, though parameter un-certainty still contributes at least 20% in all cases.