In Impacts – II – GHG Emissions Projections: SRES and RCP we looked at projections of emissions under various scenarios with the resulting CO2 (and other GHG) concentrations and resulting radiative forcing.
Why do we need these scenarios? Because even if climate models were perfect and could accurately calculate the temperature 100 years from now, we wouldn’t know how much “anthropogenic CO2” (and other GHGs) would have been emitted by that time. The scenarios allow climate modelers to produce temperature (and other climate variable) projections on the basis of each of these scenarios.
The IPCC AR5 (fifth assessment report) from 2013 says (chapter 12, p. 1031):
Global mean temperatures will continue to rise over the 21st century if greenhouse gas (GHG) emissions continue unabated.
Under the assumptions of the concentration-driven RCPs, global mean surface temperatures for 2081–2100, relative to 1986–2005 will likely be in the 5 to 95% range of the CMIP5 models:
- 0.3°C to 1.7°C (RCP2.6)
- 1.1°C to 2.6°C (RCP4.5)
- 1.4°C to 3.1°C (RCP6.0)
- 2.6°C to 4.8°C (RCP8.5)
Global temperatures averaged over the period 2081– 2100 are projected to likely exceed 1.5°C above 1850-1900 for RCP4.5, RCP6.0 and RCP8.5 (high confidence), are likely to exceed 2°C above 1850-1900 for RCP6.0 and RCP8.5 (high confidence) and are more likely than not to exceed 2°C for RCP4.5 (medium confidence). Temperature change above 2°C under RCP2.6 is unlikely (medium confidence). Warming above 4°C by 2081–2100 is unlikely in all RCPs (high confidence) except for RCP8.5, where it is about as likely as not (medium confidence).
I commented in Part II that RCP8.5 seemed to be a scenario that didn’t match up with the last 40-50 years of development. Of course, the various scenario developers give their caveats, for example, Riahi et al 2007:
Given the large number of variables and their interdependencies, we are of the opinion that it is impossible to assign objective likelihoods or probabilities to emissions scenarios. We have also not attempted to assign any subjective likelihoods to the scenarios either. The purpose of the scenarios presented in this Special Issue is, rather, to span the range of uncertainty without an assessment of likely, preferable, or desirable future developments..
Readers should exercise their own judgment on the plausibility of above scenario ‘storylines’..
To me RCP6.0 seems a more likely future (compared with RCP8.5) in a world that doesn’t have any significant attempt to tackle CO2 emissions. That is, no major change in climate policy to today’s world, but similar economic and population development (note 1).
Here is the graph of projected temperature anomalies for the different scenarios. :
That graph is hard to make out for 2100, here is the table of corresponding data. I highlighted RCP6.0 in 2100 – you can click to enlarge the table:
Figure 2 – Click to expand
Probabilities and Lists
The table above has a “1 std deviation” and a 5%-95% distribution. The graph (which has the same source data) has shading to indicate 5%-95% of models for each RCP scenario.
These have no relation to real probability distributions. That is, the range of 5-95% for RCP6.0 doesn’t equate to: “the probability is 90% likely that the average temperature 2080-2100 will be 1.4-3.1ºC higher than the 1986-2005 average”.
A number of climate models are used to produce simulations and the results from these “ensembles” are sometimes pressed into “probability service”. For some concept background on ensembles read Ensemble Forecasting.
Here is IPCC AR5 chapter 12:
Ensembles like CMIP5 do not represent a systematically sampled family of models but rely on self-selection by the modelling groups.
This opportunistic nature of MMEs [multi-model ensembles] has been discussed, for example, in Tebaldi and Knutti (2007) and Knutti et al. (2010a). These ensembles are therefore not designed to explore uncertainty in a coordinated manner, and the range of their results cannot be straightforwardly interpreted as an exhaustive range of plausible outcomes, even if some studies have shown how they appear to behave as well calibrated probabilistic forecasts for some large-scale quantities. Other studies have argued instead that the tail of distributions is by construction undersampled.
In general, the difficulty in producing quantitative estimates of uncertainty based on multiple model output originates in their peculiarities as a statistical sample, neither random nor systematic, with possible dependencies among the members and of spurious nature, that is, often counting among their members models with different degrees of complexities (different number of processes explicitly represented or parameterized) even within the category of general circulation models..
..In summary, there does not exist at present a single agreed on and robust formal methodology to deliver uncertainty quantification estimates of future changes in all climate variables. As a consequence, in this chapter, statements using the calibrated uncertainty language are a result of the expert judgement of the authors, combining assessed literature results with an evaluation of models demonstrated ability (or lack thereof) in simulating the relevant processes (see Chapter 9) and model consensus (or lack thereof) over future projections. In some cases when a significant relation is detected between model performance and reliability of its future projections, some models (or a particular parametric configuration) may be excluded but in general it remains an open research question to find significant connections of this kind that justify some form of weighting across the ensemble of models and produce aggregated future projections that are significantly different from straightforward one model–one vote ensemble results. Therefore, most of the analyses performed for this chapter make use of all available models in the ensembles, with equal weight given to each of them unless otherwise stated.
And from one of the papers cited in that section of chapter 12, Jackson et al 2008:
In global climate models (GCMs), unresolved physical processes are included through simplified representations referred to as parameterizations.
Parameterizations typically contain one or more adjustable phenomenological parameters. Parameter values can be estimated directly from theory or observations or by “tuning” the models by comparing model simulations to the climate record. Because of the large number of parameters in comprehensive GCMs, a thorough tuning effort that includes interactions between multiple parameters can be very computationally expensive. Models may have compensating errors, where errors in one parameterization compensate for errors in other parameterizations to produce a realistic climate simulation (Wang 2007; Golaz et al. 2007; Min et al. 2007; Murphy et al. 2007).
The risk is that, when moving to a new climate regime (e.g., increased greenhouse gases), the errors may no longer compensate. This leads to uncertainty in climate change predictions. The known range of uncertainty of many parameters allows a wide variance of the resulting simulated climate (Murphy et al. 2004; Stainforth et al. 2005; M. Collins et al. 2006). The persistent scatter in the sensitivities of models from different modeling groups, despite the effort represented by the approximately four generations of modeling improvements, suggests that uncertainty in climate prediction may depend on underconstrained details and that we should not expect convergence anytime soon.
Stainforth et al 2005 (referenced in the quote above) tried much larger ensembles of coarser resolution climate models, and was discussed in the comments of Models, On – and Off – the Catwalk – Part Four – Tuning & the Magic Behind the Scenes. Rowlands et al 2012 is similar in approach and was discussed in Natural Variability and Chaos – Five – Why Should Observations match Models?
The way I read the IPCC reports and various papers is that clearly the projections are not a probability distribution. Then the data gets inevitably gets used as a de facto probability distribution.
“All models are wrong but some are useful” as George Box said, actually in a quite unrelated field (i.e., not climate). But it’s a good saying.
Many people who describe themselves as “lukewarmers” believe that climate sensitivity as characterized by the IPCC is too high and the real climate has a lower sensitivity. I have no idea.
Models may be wrong, but I don’t have an alternative model to provide. And therefore, given that they represent climate better than any current alternative, their results are useful.
We can’t currently create a real probability distribution from a set of temperature prediction results (assuming a given emissions scenario).
How useful is it to know that under a scenario like RCP6.0 the average global temperature increase in 2100 has been simulated as variously 1ºC, 2ºC, 3ºC, 4ºC? (note, I haven’t checked the CMIP5 simulations to get each value). And the tropics will vary less, land more? As we dig into more details we will attempt to look at how reliable regional and seasonal temperature anomalies might be compared with the overall number. Likewise rainfall and other important climate values.
I do find it useful to keep the idea of a set of possible numbers with no probability assigned. Then at some stage we can say something like, “if this RCP scenario turns out to be correct and the global average surface temperature actually increases by 3ºC by 2100, we know the following are reasonable assumptions … but we currently can’t make any predictions about these other values..”
Long-term Climate Change: Projections, Commitments and Irreversibility, M Collins et al (2013) – In: Climate Change 2013: The Physical Science Basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change
Scenarios of long-term socio-economic and environmental development under climate stabilization, Keywan Riahi et al, Technological Forecasting & Social Change (2007) – free paper
Error Reduction and Convergence in Climate Prediction, Charles S Jackson et al, Journal of Climate (2008) – free paper
Note 1: As explored a little in the last article, RCP6.0 does include some changes to climate policy but it seems they are not major. I believe a very useful scenario for exploring impact assessments would be the population and development path of RCP6.0 (let’s call it RCP6.0A) without any climate policies.
For reasons of”scenario parsimony” this interesting pathway avoids attention.