Guest Post by Clyde Spencer
Because of recent WUWT guest posts, and their comments, I decided to do what I have been putting off for too long – reading what the IPCC has to say about climate modeling. The following are my remarks regarding what I thought were some of the most important statements found in FAQ 9.1; it asks and answers the question “Are Climate Models Getting Better, and How Would We Know?”
IPCC AR5 FAQ 9.1 (p. 824) claims:
“The complexity of climate models…has increased substantially since the IPCC First Assessment Report in 1990, so in that sense, current Earth System Models are vastly ‘better’ than the models of that era.”
They are explicitly defining “better” as more complex. However, what policy makers need to know is if the predictions are more precise, and more reliable! That is, are they useable?
FAQ 9.1 further states:
“An important consideration is that model performance can be evaluated only relative to past observations, taking into account natural internal variability.”
This is only true if model behavior is determined by tuning to match past weather, particularly temperature and precipitation. However, this pseudo model-performance is little better than curve fitting with a high-order polynomial. What should be done is to minimize the reliance on historical data, using first-principles more than what is currently done, doing a projection and waiting 5 or 10 years to see how well the projection forecasts the actual future temperatures. The way things are done currently – although using first-principles – may not be any better than using a ‘black box’ neural network approach to make predictions, because of the reliance on what engineers call “fudge factors” to tune with history.
FAQ 9.1 goes on to say:
“To have confidence in the future projections of such models, historical climate—and its variability and change—must be well simulated.”
It is obvious that if models didn’t simulate historical climate well, there would be no confidence in their ability to predict. However, historical fit alone isn’t sufficient to guarantee that projections will be correct. Polynomial-fits to data can have high correlation coefficients, yet are notorious for flying off into the Wild Blue Yonder when extrapolated beyond the data sequence. That is why I say above that the true test of skill is to let the models actually forecast the future. Another approach that might be used is to not tune the models to all historical data, but only tune them to the segment that is pre-industrial, or pre-World War II, and then let them demonstrate how well they model the last half-century.
One of the problems with tuning to historical data is that if the extant models don’t include all the factors that influence weather (and they almost certainly don’t), then the influence of the missing parameter(s) is/are proxied inappropriately by other factors. That is to say, if there was a past ‘disturbance in the force(ing)’, of unknown nature and magnitude, to correct for it, it would be necessary to adjust the variables that are in the models, by trial-and-error. Can we be certain that we have identified all exogenous inputs to climate? Can we be certain that all feedback loops are mathematically correct?
Inconveniently, it is remarked in Box 9.1 (p. 750):
“It has been shown for at least one model that the tuning process does not necessarily lead to a single, unique set of parameters for a given model, but that different combinations of parameters can yield equally plausible models (Mauritsen et al., 2012).”
These models are so complex that it is impossible to predict how an infinitude of combinations of parameters might influence the various outputs.
The kind of meteorological detail available in modern data is unavailable for historical data, particularly prior to the 20th Century! Thus, it would seem to be a foregone conclusion that missing forcing-information is assigned to other factors that are in the models. To make it clearer, in historical time, we know when most volcanoes erupted. However, what the density of ash in the atmosphere was can only be estimated, at best, whereas the ash and aerosol density of modern eruptions can be measured. Historical eruptions in sparsely populated regions may only be speculation based on a sudden decline in global temperatures that last for a couple of years. We only have qualitative estimates of exceptional events such as the Carrington Coronal Mass Ejection of 1859. We can only wonder what such a massive injection of energy into the atmosphere is capable of doing.
Recently, concern has been expressed about how ozone depletion may affect climate. In fact, some have been so bold as to claim that the Montreal Protocol has forestalled some undesirable climate change. We can’t be certain that some volcanos, such as Mount Katmai (Valley of Ten Thousand Smokes, AK), which are known to have had anomalous hydrochloric and hydrofluoric acid emissions (see page 4), haven’t had a significant effect on ozone levels before we were even aware of variations in ozone. For further insight on this possibility, see the following:
Continuing, FAQ 9.1 remarks:
“Inevitably, some models perform better than others for certain climate variables, but no individual model clearly emerges as ‘the best’ overall.”
This is undoubtedly because modelers make different assumptions regarding parameterizations and the models are tuned to their variable of interest. This suggests that tuning is over-riding the first-principles, and it dominates the results!
My supposition is supported by their subsequent FAQ 9.1 remark:
“…, climate models are based, to a large extent [my emphasis], on verifiable physical principles and are able to reproduce many important aspects of past response to external forcing.”
It would seem that tuning is a major weakness of current modeling efforts, along with the necessity for parameterizing energy exchange processes (convection and clouds), which occur at a spatial scale too small to model directly. Tuning is ‘the elephant in the room’ that is rarely acknowledged.
The authors of Chapter 9 acknowledge in Box 9.1 (p. 750):
“…the need for model tuning may increase model uncertainty.”
Exacerbating the situation is the remark in this same section (Box 9.1, p. 749):
“With very few exceptions … modelling centres do not routinely describe in detail how they tune their models. Therefore the complete list of observational constraints toward which a particular model is tuned is generally not available.”
Lastly, the authors clearly question how tuning impacts the purpose of modeling (Box 9.1, p. 750):
“The requirement for model tuning raises the question of whether climate models are reliable for future climate projections.”
I think that it is important to note that buried in Chapter 12 of AR5 (p. 1040) is the following statement:
“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 ….”
This is important because it implies that the quantitative correlations presented below are nominal values with no anchor to inherent uncertainty. That is, if the uncertainties are very large, then the correlations themselves have large uncertainties and should be accepted with reservation.
Further speaking to the issue of reliability is this quote and following illustration from FAQ 9.1:
“An example of changes in model performance over time is shown in FAQ 9.1, Figure 1, and illustrates the ongoing, albeit modest, [my emphasis] improvement.”
Generally, one should expect a high, non-linear correlation between temperatures and precipitation. It doesn’t rain or snow a lot in deserts, or at the poles (effectively cold deserts). Warm regions, i.e. the tropics, allow for abundant evaporation from the oceans and transpiration from vegetation, and provide abundant precipitatable water vapor. Therefore, I’m a little surprised that the following charts show a higher correlation between temperature and spatial patterns than is shown for precipitation and spatial patterns. To the extent that some areas have model temperatures that are higher than measured temperatures, then there have to be areas with lower than what is measured, in order to meet the tuning constraints of the global average. Therefore, I’m not totally convinced by the claims of high correlations between temperatures and spatial patterns. Might it be that the “surface temperatures” include the ocean temperatures, and because the oceans cover more than 70% of the Earth and don’t have the extreme temperatures of land, the temperature patterns are weighted heavily by sea surface temperatures? That is, would the correlation coefficients be nearly as high if only land temperatures were used?
The reader should note that the claimed correlation coefficients for both the CMIP3 and CMIP5 imply that only about 65% of the precipitation can be predicted by the location or spatial pattern. If precipitation patterns are so poorly explained compared to average surface temperatures, it doesn’t give me confidence that regional temperature patterns will have correlation coefficients as high as the global average.
To read any or all of the IPCC AR5, go to the following hyperlink: http://www.ipcc.ch/report/ar5/wg1/
*Intergovernmental Panel on Climate Change, Fifth Assessment Report: Working Group 1; Climate Change 2013: The Physical Science Basis: Chapter 9 – Evaluation of Climate Models