3.3 Interpretation of UKCP09 probabilistic climate projections

UKCP09 provides a state-of-the-art basis for assessing the risk of different outcomes consistent with current climate modelling capability and understanding. However it is not yet possible to provide probabilistic projections for all variables of interest. As knowledge improves in future, the projections are liable to change.

In this chapter we have described our methodology for probabilistic projection in UKCP09, based on perturbed physics ensemble of climate model simulations specifically designed to sample uncertainties in key physical and biogeochemical processes. This is done by perturbing poorly constrained parameters in a number of configurations of one particular climate model (HadCM3), to which is added a strategy for the sampling of structural modelling uncertainties (discrepancy, explained in Section 3.2.8) by using results from one of our perturbed physics ensembles to predict the results of an alternative ensemble of climate change simulations from models developed at different climate research institutes.

Our ensemble projections are converted into probabilistic projections using a Bayesian statistical framework developed to support inference of future information about real systems from complex but imperfect models (Goldstein and Rougier, 2004; Rougier, 2007). This process allows our projections to be constrained by a set of observations of past climate (Section 3.2.9), and also involves the use of expert judgements, for example in specifying prior distributions for uncertain model parameters. The probabilities which emerge from this approach represent the relative credibility of a family of different possible outcomes, taking into account our understanding of physics, chemistry, biology, observational evidence, and expert judgment. Climate change probabilities cannot be verified in the same way as (say) probabilistic weather forecasts, because we do not have the opportunity to test our projections over many historical forecast cycles. Rather, they should be interpreted as an attempt to quantify the relative risk of different future outcomes, consistent with climate modeling technology, physical understanding and observational evidence currently available.

The credibility of the UKCP09 projections should be judged, therefore, on whether the underlying experimental design captures the leading known drivers of uncertainty, and on the extent to which the projections are robust to reasonable variations in the experimental choices and assumptions. These have been highlighted throughout the chapter, and Annex 2 contains a number of tests of key assumptions, including our expert prior distributions for model parameters, our method of estimating discrepancy, and our method of selecting the appropriate level of detail in the observational information used to constrain our projections (specifically the number of eigenvectors retained in our analysis, as explained in Section 3.2.9). This Annex also tests our results by comparing them against an approach based on a different philosophy, in which probabilities of future change are sought using a method designed to maximize the role of the constraining observations, and to be as independent as possible from the set of climate models used (e.g. Allen et al. 2000; Stott et al. 2006a).

Some of our experimental choices are not yet testable, and arise from unavoidable limitations imposed by limited scientific understanding or modelling capability. For example, while we believe that our experimental design caters for the leading known drivers of uncertainty in 21st century climate change (in particular physical atmospheric feedback processes, and carbon cycle feedbacks), there are other possible forcing agents (e.g. non-sulphate aerosol species), or feedbacks (e.g. through methane cycle processes) which are not included in UKCP09. We have no positive evidence that such factors would, if included, provide sources of uncertainty comparable with those included in UKCP09 (at least for projection time scales of a century or less), but this remains an issue for future research. 

Further assumptions are imposed by limitations in computational resource. In particular, we sample uncertainties in surface and atmospheric physical processes more comprehensively than uncertainties in other earth system modules (ocean, sulphur cycle, carbon cycle), because it was not feasible to run the large ensembles of time-dependent climate change simulations which would be required. Thus we characterise uncertainties in these modules using simpler methods, applying the greater sophistication of our Bayesian calculations only to the treatment of surface and atmospheric uncertainties. In the case of the carbon cycle, however, we do make a simple attempt to account for variations in historical simulation skill between different ensemble members, and to account for structural modeling uncertainties by including results from a multi-model ensemble of projections (Friedlingstein et al. 2006), in addition to those from our perturbed physics ensemble.

We also assume that non-linear interactions between uncertainties in different components of the Earth system are important at the global scale, but not at the regional scale, because our finite computing resources were not able to support ensembles of climate projections with a comprehensive Earth System Model (ESM) in which uncertain processes in different components were simultaneously co-varied. Such an experiment is now in progress with HadCM3C, but UKCP09 relies on the assumption that regional interactions between earth system components are likely to be small compared with uncertainties arising when each component is sampled in isolation.

It is important that such caveats are clearly recognized. However, we believe that the UKCP09 methodology represents the most systematic and comprehensive attempt yet to provide climate projections which combine the effects of key sources of uncertainty, are constrained by a set of observational metrics representative of widely-accepted tests of climate model performance, and provide a state-of-the-art basis for the assessment of risk, within limits of feasibility imposed by current modelling capability and computing facilities.

Another key point is that we cannot make a universal assumption that probabilistic predictions can be provided for all variables that users might be interested in. As discussed in Section 3.2.10, our method is based on the assumption that robust probabilities cannot be inferred from small multi-model ensembles in isolation (see Section 3.1), and that larger perturbed physics ensembles can be used as an alternative means of sampling key process uncertainties to first order. If this is the case, then we would expect that: (a) the spread of changes simulated by the 12 member multi-model ensemble used in UKCP09 should lie more or less within that simulated by our corresponding perturbed physics ensemble; (b) even if (a) is satisfied, the discrepancy term calculated from the multi-model ensemble results should supply a modest (albeit non-trivial) component to the total uncertainty reflected in our probability distributions. With the exception of the latent heat flux variable (see Section 3.2.10), we find that criteria (a) and (b) are satisfied for the UKCP09 projection variables.

However, there were two further variables for which probabilities could not be provided, for different reasons. In the case of soil moisture content, the issue was that different models define this variable in slightly different ways, so it was not possible to calculate a discrepancy term by comparing the perturbed physics results against simulations of a consistently defined quantity in the multi-model ensemble. Secondly, it was not possible to provide probabilistic projections of fractional changes in snowfall. This is because the logarithmic transformation applied prior to our statistical calculations (in order to avoid the possibility of projecting reductions below the absolute bound of –100% — see Section 3.2.3) sometimes resulted in distributions with a highly skewed upper tail. This suggested a non-negligible probability for substantial increases in snowfall, not supported by the climate model results. This arose because the logarithm of snowfall varies rapidly at small snowfall values, and small values are often simulated in the climate model runs. This in turn means that statistical uncertainties (variances resulting from emulation error, downscaling error and timescaling error) calculated in the transformed variable tend to have large values. However our method does not account for changes in this variance as a function of the value of the projection variable, so these large variances are then assumed to apply to all projected values, leading to an unrealistic inflation of the upper tail of the attempted probabilistic projection. Changes in wind speed and snowfall derived from our 11 member regional climate model ensemble projections are discussed in Chapter 4, noting that these simulations sample only a subset of the uncertainties considered in the fuller probabilistic analysis applied to other variables.

For users, an important question concerns how climate projections will change in future. Should planners make decisions now, based on estimates showing a wide range of possible changes, or should they delay in the hope that more precise information will be available in (say) 10 years time? On the one hand, modellers have striven successfully to improve their models over the past decade or so (e.g. Reichler and Kim, 2008), yet the range of future global projections in the IPCC AR4 (Meehl et al. 2007) was not significantly narrower than in the previous IPCC assessment, and the range of projected changes over the UK has certainly not narrowed. On the other hand, some of the errors in climate models tend to be systematic across different models, partly due to shared features such as limited resolution. Examples, including a tendency to underestimate the frequency of blocking anticyclones over Europe in winter, are given in Annex 3. The presence of common errors gives rise to the possibility that ensemble climate projection exercises of the future might give different results to those deriving from the current generation of models, at least for some aspects of climate.

In practice, therefore, the prospects for better projections will depend on which variables or which future periods users are most interested in. For example, uncertainties in the UKCP09 projections are substantial even for a couple of decades ahead (Sections 4.4.2 and 4.5), due to the significant influence of internal variability at regional scales, and then grow larger through the 21st century due to the additional influence of uncertain climate change feedbacks (Box 2.1). Prospects for reducing uncertainties in near-term changes are likely to rest mainly on constraining projections of internal variability by initializing climate models with ocean observations (Smith et al, 2007; Keenlyside et al, 2008), and through improvements in the ability of models to simulate regional modes of variability. For example, increased horizontal or vertical resolution might lead to better simulation of features such as the North Atlantic storm track, or the coupling between sea surface temperature anomalies and atmospheric circulation anomalies. At longer lead times progress would also depend on improvements in our ability to represent thermodynamic climate feedbacks and carbon cycle processes, and their complex interactions. An active dialogue between users and climate research scientists will therefore be crucial, in order to ensure that adaptation decisions are taken on the basis of up-to-date information concerning the potential for emerging research to update projections currently available, such as UKCP09.

As mentioned above, improvements in climate models are one potential route to improved projections in future. By improved, we mean both more comprehensive sampling of climate feedbacks (through the use of comprehensive ESMs), and smaller uncertainties through the development of models with higher resolution and better representations of sub-grid scale processes. Initialization of climate models with observations (also mentioned above) has potential to improve projections of near-term climate over the next decade or so, and possibly longer. Uncertainties could also be reduced by developments in experimental design, subject to available computing resources. For example, future exercises of this type could potentially be based entirely on simulations in which the atmosphere model is coupled to a full dynamical ocean component, rather than a simple mixed layer ocean (see Section 3.2.3). This would remove the need for scaling approaches to infer time-dependent climate changes from equilibrium changes, and hence narrow the probability distributions significantly, as our timescaling procedure is responsible for a significant component of the total uncertainty captured in our probabilities (see Annex 2). It would also allow a wider range of observational metrics to be used in constraining the projections.

In summary, the UKCP09 projections should be seen as a comprehensive summary of possible climate futures consistent with understanding, models and resources available at present, but users should be aware that the projections could change in future, as the basis for climate prediction evolves over time.

Previous  Next