Yesterday on WUWT, a post from Luboš Motl told us how climate science has been proposed as a vehicle for wealth redistribution by Hans Joachim Schellnhuber who is the director of the Potsdam Institute for Climate Impact Research and the main German government’s climate protection adviser. Interestingly it has been discovered that he co-authored a paper critical of Global Climate Models (GCM’s) in 2001. The paper and list of co-authors is below.
Global Climate Models Violate Scaling of the Observed Atmospheric Variability (link to PDF here)
R. B. Govindan,1,2 Dmitry Vyushin,1,2 Armin Bunde,2,* Stephen Brenner,3
Shlomo Havlin,1 and Hans-Joachim Schellnhuber4
1Minerva Center and Department of Physics, Bar-Ilan University, Ramat-Gan 52900, Israel
2Institut für Theoretische Physik III, Justus-Liebig-Universität Giessen, Heinrich-Buff-Ring 16, 35392 Giessen, Germany
3Department of Geography, Bar-Ilan University, Ramat-Gan 52900, Israel
4Potsdam Institute for Climate Impact Research, D-14412 Potsdam, Germany
(Received 1 November 2001; revised manuscript received 22 April 2002; published 21 June 2002)
Abstract:
We test the scaling performance of seven leading global climate models by using detrended fluctuation
analysis. We analyze temperature records of six representative sites around the globe simulated by the
models, for two different scenarios: (i) with greenhouse gas forcing only and (ii) with greenhouse gas
plus aerosol forcing. We find that the simulated records for both scenarios fail to reproduce the universal
scaling behavior of the observed records and display wide performance differences. The deviations
from the scaling behavior are more pronounced in the first scenario, where also the trends are clearly
overestimated.
DOI: 10.1103/PhysRevLett.89.028501 PACS numbers: 92.60.Wc, 02.70.Hm, 64.60.Ak, 92.70.Gt
In the conclusion the authors write:
To summarize, we have presented evidence that
AOGCMs fail to reproduce the universal scaling behavior
observed in the real temperature records. Moreover, the
models display wide differences in scaling for different
sites. When comparing the two scenarios, our results
suggest that the second scenario (CO2 plus aerosols)
exhibits better performance regarding the values of the
scaling exponents as well as the trends. The effect of
aerosols not only decreases the trends but also modifies
the fluctuations, to more closely resemble the real data.
This confirms in a way independent of the evaluations
made so far [5] that the incorporation of aerosols is
necessary to approach reality.
It is possible that the lack of long-term persistence is due
to the fact that certain forcings such as volcanic eruptions
or solar fluctuations have not been incorporated in the models.
However, we cannot rule out that systematic model
deficiencies (such as the use of equivalent CO2 forcing to
account for all other greenhouse gases or inaccurate spatial
and temporal distributions of sulphate emissions) prevent
the AOGCMs from correctly simulating the natural variability
of the atmosphere. Since the models underestimate
the long-range persistence of the atmosphere and overestimate
the trends, our analysis suggests that the anticipated
global warming is also overestimated by the models.
Oddly, though by his own peer reviewed admission, GCM’s don’t fully represent reality, and “global warming is also overestimated by the models”, that doesn’t stop Schellnhuber from using the conclusions of GCM’s to create his own alternate world reality where the industrialized nations have to pay carbon reparations to poorer ones.
h/t to Steve Mosher
Joel,
1. Do you think it makes sense to do an ensemble average of GCMs that
include volcanic forcing with those that don’t include volcanic forcing?
Joel answers Yes.
I find this answer odd. Joel points to a study that finds a better fit
to observations if volcanic forcing is INCLUDED. And yet would
Average those GCM that include volcanic forcing with those that don’t.
We are not talking forecasts here, we are talking hindcasts. Effectively
what this does is BROADEN the range or envelope of hindcasts. In
short, you make the envelope of hindcasts so wide that they are
“consistent with” observations. They is like an insurance policy
against falsification.
2. Would you ever base policy on a study that averaged runs from GCMs
that used volcanic forcing with runs from GCMs that did not?
Joel answers yes. here is what you have. You have two sets of models.
In one set the models do not model volcanic forcing. They have a certain
fit to the past, lets say they are warmer than they should be. tweaks
(parameterizations) are applied to moderate this misfit. In another
set of models where volcanic forcing is applied the fit to observations
is better. Less tweaking or different tweaking. Now, you do a FORECAST.
In all forecasts there is no volcanic forcing applied because you can’t predict
future volcanic forcing. Why would uses models that are physically WRONG
( no volcanic forcing) to do a forecast? Especially given that their parameterizations to achieve better hindcasts are tainted by the lack of historical volcanic forcing?
3. Would you reject a projection based on the average of models where only
some of the models including volcanic forcing?
Joel answer No. See above.
As Joel explains:
“My longer answer would be that it is certainly worth investigating whether the GCMs that include volcanic forcing give significantly different projections than those that don’t. ”
Well, the paper you site says that including forcing improves the skill WRT
hindcasts. Also, note the wiggle room in the phrase “significantly different”
So here is a thought experiment. Let’s suppose you have two sets of models. Those with and those without volcanic forcing. Both are run in forecast without volcanic forcing. Would you expect to find a difference?
“However, I suspect that the answer to this question would be “no”. And, since we can’t predict future volcanic eruptions, including them has to be done in some guesstimation way anyway.”
If you found that the answer was “yes” would you question the inclusion of models that didn’t include volcanic forcing? or would you include the non physical models so that you get a bigger envelop of projections and thus insulate yourself from future claims of falsification? Finally, WRT guesstimates on volcanic forcing.
A. Hansen did this so there is precedent.
B. What do you think the SRES are? they are a wide range of guesses
about future emissions. I would hazard that we can guess the future
average forcing from volcanoes better than we can guess the future
concentrations of GHGs, after all we have a good handle on the average
volcanic forcing going back 100s of thousands of years.
I suppose one can also ask the question more generally of joel. You have a
collection of GCMs. Some show good skill in hindcast. Others suck. Would
you use those that suck to do forecasts?
Video posted by John B (23:18:47): “We’re going green” lyric sung while collecting hands full of dollars. So true!
Remmitt:
I made a similar comment while putting ones in a stripper’s g-string last night. Really I did. “You’re going green”.
jeez:
That’s another example of unwise spending, but at least pleasant and by free choice…
Are you kidding?
That’s about the most efficient kind of economic stimulus there is.
Dancers immediately pour that money right back into the economy on drinks, makeup, shoes, Ed Hardy purses, DUI lawyers, and iPhones (the dancers’ choice), which they lose and replace bimonthly.
Joel Shore,
Would just like to say that I appreciate you posting here.
steven mosher: Here are a few points that I would make –
(1) The paper that I referenced does not says “that including forcing improves the skill WRT hindcasts” at least in the sense that I think you mean hindcasts, i.e., looking at how well the global temperature in the models follows the measured global temperature. Rather, it looks at a very technical aspect of the temperature record (its scaling behavior, i.e., how the temperature is correlated in time) and compares the models and data. This is a very different issue.
(2) If you are so concerned about how the inclusion of models that don’t include volcanic forcing affects the reported results, you can probably go and look and see if it really significantly broadens the future predictions. I personally doubt that it does.
(3) The fact is that the 20th century global temperature record does not that strongly constrain the predicted future warming because there is a lot of uncertainty about the magnitude of the aerosol forcing, so there is a fairly large range of possible climate sensitivities that the global temperature record is compatible with. Better constraints on the climate sensitivity are derived from looking at other empirical data, such as the last glacial maximum and the climate response to the Mt. Pinatubo eruption. The best constraints are contained by combining all three.
(4) The model parameters are not really tweaked to hindcast the global temperature record. What parameters exist are mainly set by other climatology issues or by mimicking the physical processes themselves. (The possible exception to this may be a sort of unconscious tuning that has been done whereby the models with higher sensitivity to aerosols also tend to have higher sensitivity to GHGs as a paper by Kiehl et al. claimed.)
(5) At some point when the constraints on the climate sensitivity, the aerosol forcing and the like become tighter then it may become a little more important to include volcanic forcing at least in the hindcast. But, I don’t think that we are there yet.
TonyB: What you have identified are the kinds of issues that one deals with in science all of the time. Data is seldom available without some (often severe) data quality issues. I don’t know what to tell you except that the way one deals with it is the way that climate scientists have been dealing with it: by having different independent analyses done on the data sets, by making estimates of the errors introduced due to various known data issues, by looking at different measures of a similar thing (e.g., in addition to direct air thermometer measurements, there are ocean temperature measurements, measurements of the advance and retreat of glacials, borehole temperature measurements, and various temperature proxy measurements).
Work at the forefront of science is seldom as easy and straightforward as presented in science textbooks. This is probably one of the main reasons why scientists and “lay people” tend to reach rather different conclusions regarding the strength of the evidence in a field such as climate science where lay people are motivated to investigate the science. If other fields of science were subjected to the similar sort of study by lay people that climate science has been, I think that these people would find similar deficiencies…and they would probably be left believing hardly any of the theories of modern science. And yet, I think these theories have been very important and successful in advancing our knowledge, and the reason I think that is the case is that, while each individual piece of data or each experiment may have its problems, the whole set of data and experiments taken together generally have a high degree of redundancy that means that the overall picture that emerges is more likely to be correct than one would expect by looking at each experiment in isolation.
To be honest, I have spent almost my entire career as a computational physicist being continually surprised at how well the models that I use agree with the experimental data in spite of the fact that I can almost always identify many concerns that I have with either the data or the model (e.g., many things that the model is ignoring).
Joel
Thanks for your candour. However it is explained away, much of the theoretical information presented is contradicted by observations from the real world. It seems to me that climate science has standards not as rigorous as those applied to other sciences and the burden of proof falls far short of what should be expected.
This is probably the first science born in the computer age and many believe the models to be more accurate than we can in reality currently achieve.
According to the IPCC, Climate Change 2007: The Physical Science Basis “The set of available models may share fundamental inadequacies, the effects of which cannot be quantified.”
best wishes
Tonyb
Perhaps I should explain how these GCM models are constructed.
Except on a molecular and smaller scales, which is why we can develop from first principles statistics such as Max Boltz, larger scale actual natural events are not truly random.
On these larger scales random events are essentially a mechanical artifact about which we can deduce, in the case for example of the roulette wheel, that over time equal numbers of reds and blacks will happen.
We can also observe on this larger scale that the occurence of an individual event, for example being struck by lightning, also shows true randomness. Which is not to say that if you stand where lightning tends to strike you will not be more likely to be struck.
But we cannot predict when that might happen.
Beyond that we cannot go.
We only know that large scale natural systems are not truly random so even if we were to consider the balance of the outcomes over the lifetime of the Universe itself we would be no wiser. For instance we now know that black holes obey the second law of thermodynamics, so the fact that a largish black hole might have a theoretical lifetime longer than the universe itself is irrelevant since it must disappear when the universe itself comes to an end.
In short unlike our roulette wheel the balance of the outcome, equally red or black, does not apply to large scale natural systems even over the lifetime of the universe.
As you might expect given the irreversible nature of change in the universe normally defined as the Arrow of Time.
We now call such variable natural systems chaotic to distinguish them from truly random events and outcomes. But they have a degree of randomness which defies prediction in terms of a balanced outcome over time: even to the end of time itself.
Which is why it is absurd to say that weather, which we cannot predict with any certainty, depending on where you live, for more than few days
is not the same as climate which it is averred we can forecast fifty years ahead. Really?
Well let us see how this supposed GCM prediction is done.
Given that, unlike our roulette wheel, we have no simple mechanical cause, we have to begin by trying to define the variables we think affect the climate and build up a dynamic model from that.
Since we do not know the future we can only test our model against the past, and if it does not fit we can adjust the weighting of the variables until it does.
If of course it turns out it does not predict the future we can then take the new data which show it to be wrong and amend the variables until we now get a perfect hindcast.
Does this mean our model is better than before? And will now predict the future?
No: probably not. So if it fails again we adjust it again to fit the results. Is it now better than before? Probably not.
This fallacy, known to wise gamblers, wise bankers, statisticians and the like, occurs because it is imagined that every variable has not only been accounted for but that moreover as a test if the variables are set to produce a neutral outome then there must necessarily be a balanced outcome over time.
Except in the real world the outome is not necessarily balanced over time, nor is there any reason to suppose it should be.
Yet this fallacy that the outcome must balance over time around the trend line of the model is used to suggest that whilst the model cannot predict the immediate future it can still foretell the far future.
What balderdash.
Worse in large scale GCM models if you do include more than handful of variables the model collapses into a puddle before it has got to the prediction for one year: and simply swings over and settles at on of two extreme outcomes, hellfire or absolute zero, neither of which seem likely.
To prevent the model doing this is it necessary to impose a constraint which forces the model to remain within reasonable limits, although enormities can be committed in defining what reasonable is depending on what the modeller believes might be acceptable to the client.
Whether such a constraint exists in Nature or is merely a figment of the imagination of the modeller I leave you to ponder.
But you might are to note that unlike gamblers fallacies whih tend to fall apart in practice fairly quickly climate change is a long term affair so it has taken the last decade to convincingly falsify the models.
As I said in my post above it is now apparent these models failed to predict the current cooling trend, cannot foretell how long it will last or how deep it will be and cannot even tell us how much of the recent warming was due to natural causes.
Thus in the words of Dr. Stockdale at the IPCC conference the models are biased which is damaging our forecasts. Och aye.
What he did not say was that if you test the models against the actual outcome their predictive power turns out to be no better than random. A test you can do for yourself.
Whih is why I said you would do as well with a set of Tarot cards, or perhaps you might prefer a Mystic Medium.
As for the idea espoused by Dr. Pope that in many ways we know more about the climate in 2050 than we do about next decade or two, it’s enough to make a cat laugh.
Kindest Regards.
Joel Shore (10:12:07) :
a jones:
Joel Shore seems to imagine that real greenhouses chiefly lose heat by conduction: see above post.
I would have thought any competent physicist, well in fact any schoolboy with an elementary scientific education, would know that whether they use glass or plastic, real greenhouses retain heat by blocking convective action with their transparent barrier so the warm air does not escape upwards. Losses due to conduction or radiation are insignificant by comparison.
What odd ideas some people have.
You are of course correct. And, in this case, it is not odd ideas that I have, but rather the odd way my brain communicates with my fingers. I meant “convection” and typed “conduction”. In terms of letters, I got 8 out of 10 correct…Nobody’s perfect!
—————
It’s strange to me, Joel, that you so readily concede this point. Surely, the greenhouse is warm because it prevents convection. It does not lose heat by convection unless the windows are left open!
As for the Earth’s atmosphere; it’s not a greenhouse.
a jones (17:34:44) :
Maybe I should once more on this board say how the GCMs are supposed to work, and how they cannot be trusted for unlimited iterations forward in time.
They make a three dimensional grid of the globe. 200*200*20km(height)
They take time steps from 20 minutes or more.
They use known/supposed solutions of equations for the box and keep boundary conditions physical.
Where equations are not available ( cloud generation, wind motion, convection, conduction, albedo, etc) they use average values.
This is about the way weather prediction works too.
Now the solution of the equations are not linear, they can be highly non linear. Average values used for unknown solutions of equations are also assuming linear behavior ( the first term in a putative perturbation series expansion). That is the reason why weather prediction cannot go further than a few days, and sometimes a few hours, if the non linear chaotic reality kicks in.
It is inevitable that GCMs for climate will fail after a certain number of time steps.
The fact that hindcasting may be working is just a multidimensional demonstration of what von Neuman said: “give me four parameters and I can fit any function. With five I can fit an elephant”. The number of parameters in these GCMs is n, where n is a large number ( as my math professor used to say). I suspect that albedo is the most crucial one.
I think I’ve never heard so loud
The quiet message in a cloud.
===================
anna v (and a jones): Using your logic, one could conclude that we would be unable to predict that the climate will be warmer here in Rochester in July than it will be in January due to the seasonal forcing.
The actual fact is that the climate modelers understand chaos..and in fact can easily demonstrate it in the climate models by making a perturbation on the initial conditions. And, while two different runs with slightly different initial conditions will diverge in their details (e.g., the “jiggles” up and down in temperature), they will both show approximately the same response to a forcing such as that due to increased GHGs provided the run is over a long enough period that the response to the forcing dominates over the climate noise.
Admittedly, non-linear dynamical systems can be full of surprises and there is always the possibility that driving such a system beyond a certain point will send it into a very different state … However, such potential surprises seem to me to be arguments for more, rather than less, prudence in how hard we drive the system.
While it is true that one can usually fit a curve with a well-chosen model containing only a few parameters, one can also devise a system with millions of parameters where you could not fit some simple piece of data. The distinction is what sort of model it is and what things the parameters control. To fit the global temperature record, the simplest thing to perform such a fit would simply be an empirical model that writes temperature T as a function of time t (e.g., a polynomial in t).
However, this is not in any way shape or form the sort of models that climate models are and the parameters in climate models are not designed in such a way that it would be easy to reproduce a given global temp vs time curve. Rather, the parameters tend to control microphysics such as the nucleation of clouds and other such stuff. You could probably tinker with these parameters all day and never fit the global temperature record as long as the history of the climate forcings vs time are not approximately correct. (And, in fact, I know of no skeptic who has claimed that he/she can take a GCM and play with the parameters to fit the global temperature record while including only natural forcings and not including any anthropogenic forcings.) Furthermore, there is a lot more data that a climate model can be compared to than simply the global temperature vs time. The parameters in the climate models are generally constrained by either getting the physics on smaller scales approximately correct or reproducing current climatology.
The one exception to the above is that the aerosol forcing is not currently well-constrained empirically, so it is possible to get reasonable fits to the global temperature record with a range of possible efficacies for the aerosol forcing coupled with a range of possible climate sensitivities in the model. This is a true limitation of the use of the 20th century temperature record to determine the climate sensitivity…and is one of the reasons why the range of estimates for the climate sensitivity remains rather broad.
Anna V.
I failed to express myself clearly. I am sorry.
It does not matter what goes on inside the black box, for all we know it could be daemons calculating the results on a blackboard with chalk.
We are only interested in the prediction that comes out.
But whatever does go in inside the black box we do know that if it obeys the second law of thermodynamics, which daemons may not but computer code certainly does, there are limitations on its predictive power.
Thus on the scale we are concerned with truly random outcomes can only be produced by mechanical artifacts such as the roulette wheel: where I can say that over time the trend will be to equal numbers of reds and blacks.
Weather and climate are not truly random which is why we use the term chaotic, but for all their considerable degree of randomness, we can deduce broad limitations on the likely range of variation: thus I can say that it is unlikely that the weather in London will either become tropical or Arctic in the near or indeed far, centuries, future.
Beyond that I cannot go.
And here we come to the nub.
There are those who would have you believe that although these black boxes cannot predict the weather beyond a few days or weeks they can predict decades ahead, climate is not weather we are told, so things will average out.
Not so. The fallacy here is to assume that over a sufficiently long period of time the results must, like the reds and blacks, revert to some trend line so that the variations even out. Really! why?. The system is not truly random so equal outcomes are not necessarily certain. And even if they were what is that period of time? it could still be longer than the lifetime of the universe itself.
The reality of this is reflected in the very problems of a collapse in the outcome that these boxes show, which collapse then has to be fixed by imposing other constraints on them. It does not matter whether that is done by computer code or by simply chucking away the results of every run that does collapse within the selected time frame and only using those that manage to stagger on long enough to survive.
Either way the result is about as much use as a fish on a bicycle.
I should also add I entirely agree with your views on hindcasting.
Kindest Regards
Joel Shore (14:27:50) :
And, in fact, I know of no skeptic who has claimed that he/she can take a GCM and play with the parameters to fit the global temperature record while including only natural forcings and not including any anthropogenic forcings.
Albedo,varied between its error band can do the job nicely, for one other parameter than aerosols.
Well, I could take your word that GCM modellers understand chaotic dynamics, but, from the last IPCC report it is clear that they do not understand statistical propagation of errors in their models, ( chapter 8) so you will allow me to be doubtful as far as their expertise in chaos is concerned.
IPCC AR$WG! chapter 8
8.1.2.2 Metrics of Model Reliability
…..
The above studies show promise
that quantitative metrics for the likelihood of model projections
may be developed, but because the development of robust
metrics is still at an early stage, the model evaluations presented
in this chapter are based primarily on experience and physical
reasoning, as has been the norm in the past.
This means in simple language that for all I know a true error propagation could show the model outputs anywhere, from zero anomaly to humongous, and the outputs have no predictive value at all.
When using a simple model that contains albedo and one varies albedo within its errors , for example in http://junkscience.com/Greenhouse/Earth_temp.html
varying the albedo from 0.31 to 0.30
the temperature goes from 15C to 16.1C
and do we really know the albedo so well? measurements say NO.
http://bbso.njit.edu/Research/EarthShine/literature/Palle_etal_2006_EOS.pdf
correction, that was 16.04 C , not 16.1
no edit here
Joel,
http://rankexploits.com/musings/2008/effect-of-including-volcanic-eruptions-on-hindcastforecast-of-gmst/
anna v:
Albedo vs. time is not an adjustable parameter in the models. It is a quantity that comes out of the models. And, even if there is uncertainty in the precise absolute value of the albedo, what matters is how it changes in runs with increasing GHGs vs control runs.
steven mosher:
I think Lucia is worrying about issues that are basically within the current error bars…especially when the uncertainty in the aerosol forcing is taken into account.
Joel Shore (10:24:50) :
Albedo vs. time is not an adjustable parameter in the models. It is a quantity that comes out of the models. And, even if there is uncertainty in the precise absolute value of the albedo, what matters is how it changes in runs with increasing GHGs vs control runs.
Thanks for the clarification. Seems the GCMs are not doing that well on that anyway:
http://www3.interscience.wiley.com/journal/118587917/abstract?CRETRY=1&SRETRY=0
A comprehensive comparison of characteristics of the planetary albedo (α) in data from two satellite measurement campaigns (ERBE and CERES) and output from 20 GCMs, simulating the 20th-century climate, is performed. Discrepancies between different data sets and models exist; thus, it is clear that conclusions about absolute magnitude and accuracy of albedo should be drawn with caution. Yet, given the present calibrations, a bias is found between different estimates of α, with modelled global albedos being systematically higher than the observed. The difference between models and observations is larger for the more recent CERES measurements than the older ERBE measurements. Through the study of seasonal anomalies and space and time distribution of correaltions between models and observations, specific regions with large discrepancies can be identified. It is hereby found that models appear to over-estimate the albedo during boreal summer and under-estimate it during austral summer. Furthermore, the seasonal variations of albedo in subtropical areas dominated by low stratiform clouds, as well as in dry desert regions in the subtropics, seem to be poorly simulated by the models.
Not surprising since clouds are badly simulated by the models too.( at least in the IPCC report figures).
“Indeed, global warming stopped and a cooling is beginning. No climate model has predicted a cooling of the Earth, on the contrary. This means that projections of future climate is unpredictable, writes Henrik Svensmark.”
For this kind of reasoning the professor should be sacked or the Copenhagen University should close its doors.
It seems Svensmark believes that next years have cooled down considerably (I said, like Svensmark: ‘have cooled down’ instead of ‘will cool down’!) and taking this as a fact denounces all climate models and claims that future climate is unpredictable. Except for Svensmark, of course. He knows ‘a cooling is beginning’.