Online climate change projections report 3.2.10 Doubled CO2
As explained in sections 3.2.7–3.2.9, probabilistic projections of equilibrium climate changes in response to doubled CO2 provide the cornerstone of the UKCP09 methodology. This process produces projections of changes in the UKCP09 variables at five global climate model (HadSM3) grid boxes covering the UK landmass (and also a further nine points covering surrounding marine regions), for every month of the year. Here we provide a few illustrations of how this part of the method works in practice, and what criteria are considered in assessing the credibility of the results.
Figure 3.4 shows an example, for changes in the 20-year average of surface air temperature (Tmean) over Wales, in March. The green histogram shows our perturbed physics ensemble of 280 HadSM3 simulations, while the multi-model ensemble (MME) results are shown as black ticks along the x-axis. The MME results provide a means of estimating the impact of structural errors in HadSM3, via the discrepancy term described in Section 3.2.8. We estimate discrepancy by taking each MME member in turn, and use a search algorithm to find four locations within the HadSM3 parameter space which match the results of the MME member most closely, based on multivariate global patterns of both historical climate and changes in response to doubled CO2 (see Section 3.2.9). Once the four HadSM3 analogues have been found, discrepancy values can be calculated for any variable of interest (e.g. temperature change over Wales in March). This is done by applying our emulator to estimate projected changes from the four HadSM3 variants, and comparing those with the simulated projection of the corresponding variable from the target MME member. Repeating this procedure for each of the 12 MME members gives 48 discrepancy estimates in total, from which a mean and variance can be calculated (we assume the discrepancy distribution to be ).
The coloured curves in Figure 3.4 show how we build up our probabilistic projection from the model simulations. We use our emulator trained on the perturbed physics ensemble results (see Section 3.2.3) to estimate results for a much larger ensemble of model variants sampling the full parameter space of HadSM3. This gives us the red curve, which also contains the impact of the variance of discrepancy (but not the mean value of discrepancy, as we wish to illustrate the impact of this separately). In Figure 3.4 the sampling of the full parameter space, combined with the addition of discrepancy variance, leads to a slight broadening of the distribution of possible changes (red curve cf green histogram). The median value is also shifted slightly towards a smaller warming, this being an effect of the improved sampling of parameter space inherent in the red curve. We also weight points in parameter space according to emulated estimates of the set of historical climate variables described in Section 3.2.9. This weighting process constrains the emulated projections according to the fit to observations, and will in general alter the characteristics of the of projected changes. In Figure 3.4 the probabilities of small or large temperature increases are reduced by the weighting (blue curve cf red curve), while the probabilities of intermediate changes increase somewhat. The mean discrepancy is then added to the projected changes at each location in the HadSM3 parameter space, to produce the final (posterior) probabilistic projection (black curve cf blue curve).
We cannot make a blanket assumption that this procedure will lead to the production of a credible result. For example, a basic assumption of our approach is that robust probabilities would be difficult to infer from small multi-model ensembles in isolation (see Section 3.1), and that perturbed physics ensembles are therefore needed to supply a more systematic means of sampling key process uncertainties to first order. If this is the case, then we would expect the spread of changes simulated by the perturbed physics ensemble to encompass that described by the multi-model ensemble, as it does in Figure 3.4.
We checked all the UKCP09 variables according to this criterion, and generally found that the spread of MME responses did lie within that of the HadSM3 ensemble. For surface latent heat flux, however, two MME members were often found to give projections at or beyond an extreme of the range given by our HadSM3 ensemble (Figure 3.5 shows a typical example). This signals that for latent heat flux the simulated changes are strongly dependent on detailed choices made in the physics of different climate models, and cannot be assumed to be approximately independent of how our experimental design was constructed (for example our decision to base the perturbed physics ensemble on HadCM3/HadSM3, rather than on some other climate model). In Figure 3.5 the outlying MME responses lead to a large discprepancy variance, which substantially inflates the spread in the red, blue and black curves, leading in particular to the projection of a significant probability for negative change in latent heat flux. This is not supported by any of the underlying model simulations. We therefore conclude that the method cannot be used to provide robust probabilistic projections for latent heat flux.
Another issue concerns the magnitude of the shift in the final projections resulting from the mean of the discrepancy term (black cf blue curve in Figure 3.4). If the perturbed physics ensemble is an effective means of sampling key uncertainties to first order, we would expect the mean value of discrepancy to exert a limited (albeit non-trivial) influence on the final results. This is indeed the case in Figure 3.4. Here, it is important to understand that the mean discrepancy can in theory be large, even when the multi-model and perturbed physics ensemble results cover similar ranges. This is because the procedure used to match MME members to their nearest perturbed physics ensemble analogues is conducted using information based on a wide range of historical and future climate information derived from global multivariate patterns. This is done to ensure that it will only be possible to find a perfect match (across all variables and regions) if the perturbed physics analogues truly replicate all aspects of the representations of physical processes simulated in their target MME members. Any remaining disparities (for some particular local variable like temperature change over Wales in March) will then be a consequence of true structural differences between HadSM3 and the MME members. Note that if we had attempted to calculate the discrepancy by conducting the matching exercise using a more limited choice of variables (say using only temperature changes over the UK), we would have risked finding misleadingly good matches over the chosen variables (through a convenient local compensation of errors effectively achieved via statistical overfitting), accompanied by unrealistically poor matches over other variables or regions not included in the matching process.
Figure 3.6 shows a histogram of the shifts in Tmean arising from the mean of the discrepancy, considering the 60 Tmean projections obtained by pooling monthly changes at all five UK land points in HadSM3. In most cases the mean discrepancy is within the range plus or minus 0.5ºC (as in Figure 3.4), and therefore provides a significant but not dominant contribution to the final projection, compared to the spread of responses simulated by the HadSM3 ensemble, or emulated across the full HadSM3 parameter space. In such cases, we typically find that the median of the posterior distribution lies somewhere between the medians of the HadSM3 and MME ensembles.