Climate Audit

Larry

I’m not sure whether I’m surprised or not.

Tim

“Although there are established data for the time evolution of well-mixed greenhouse gases, there are no established standard datasets for ozone, aerosols or natural forcing factors.”

Wow. While I’m just a layperson – that statement seems to me to be a huge admission of the weakness of current models.

My takeaway from this whole thing is that this pretty much opens the door for the modelers to tweak the inputs to get the output they are looking for. Is this what “science” has become?

Larry

But can they predict the future?

Joel McDade

I just paid the $9 for the article. That’s a first, but I think it is going to be worth it.

No I won’t be crass and share the PDF. To save you 2 seconds of google time, if not 9 USD:

http://www.agu.org/pubs/crossref/2007/2007GL031383.shtml

Wansbeck

Re #9 Joel

I had seen it on offer for $9 which is no big deal but will I paying that to hear someone state, IMHO, the obvious?

steve mosher

Well,

It has always bothered me that the guys in charge of the data, both historical and current,
are also in charge of the models. That means you can match the model to the “data” via several methods

1. goose the paleo record
2. goose the instrument record
3. goose the model and make her laugh.

Old school, we would never let a model runner near the data he had to calibrate against.

Jim Clarke

I agree with Wansbeck (#9). The conclusions of this paper have been stated countless times by climate crisis skeptics for at least 10 years. If this paper helps the rest of the climate change community come to terms with the obvious, then that is great, but why is it taking so long?

Steve: I haven’t seen any prior paper specifically showing that model climate sensitivity had an inverse relationship to the version of aerosol forcing history. People may have hypothesized something similar but this is more substantial.

bender

Finished reading the paper. Pardon the earlier comments, but I was worried that maybe this was a really important one. Not. I think I’ll catch the end of the Missouri-Oklahama game. Comments later.

Pat Keating

10 Edward

They cannot predict the future. It has been said here many times that many economic models tuned to a period of past (some sort of stat issues) history fall apart when predicting the future.

Have you read Gerald Browning’s posts over at the Exponential thread, where he says even the near-term weather-forecast models blow up trying to predict past 36 hours?

Philip_B

My comment referred to above was that the IPCC relies heavily on the models and their accuracy. It happens I know a lot about testing large complex software. Any piece of software is only as accurate/correct as the data used to test it. IMO even the temperature record pre-satellite isn’t accurate or comprehensive enough to adequately test the models. Even if the science was perfectly understood, the lack of good test data means the models and their outputs will contain significant errors. Wansbeck makes essentially the same point.

Tim G

Shouldn’t we doing “climate model report cards”?

I apologize, I am not a climate scientist and I don’t keep up with the details of the research. But as an engineer, I always wondered why we don’t compare the predictions with the measurements after the fact. We’ve had “accurate” climate models for at least a decade. And each of those models predict out over more than the next decade. Why not have a “report card” each year to compare the measured results with the predictions made by the models in the previous years?

Is there some reason we don’t do that? Or, do we and I just never heard of it. If the modelers are confident in their models, they should be the ones pushing this idea.

Anyway, maybe I’m missing something. Sorry for interrupting 🙂

–t

Philip_B

Previous thread on the accuracy (or should I say innaccuracy) of the models from 2006. Tim G, I think it answers your question.

Joel McDade

#24 Tim G

I’m not sure I agree with all of them, but Warwick Hughes has a running scorecard of GCM predictions v. observations:

http://www.warwickhughes.com/hoyt/scorecard.htm

The main one IMHO is tropospheric warming relative to surface. It seems to me that if the Team is inflating the surface measurements they are falsifying their own hypothesis. Weird, but true.

steve mosher

RE 29. It’s more like this. You run the model and it outputs 42.

historical says it should be 37.

So, you “adjust” the model.. more cow bell!

You run the model again.. 37.3. There you got just the right amount of cowbell

The temp record is 37. The Cowbell is the secret sauce.

bender

basically I would put the whole process in a loop and overfit the SOB like degrees of freedom came for free.
Right Bender?

Yes, this is my take on what they do. Gavin Schmidt insists that I am mistaken, however. Or rather, that I am misrepresenting the statistics of it. I do not think so. We’ll go there in a bit; the Mizzou game ain’t over yet. [Meanwhile, read “exponential growth” #1,#2,#3. Will pull out some relevant quotes tomorrow.]

Jon

Steve, #27 writes:

2. The temperatures are not used as inputs. They are used to “check” outputs

At least that’s the way he explained it to me.. plus if you get the Educational version of GCM
you will see that temps are not inputs.

None of which is relevant. ModelE runs on its own of course. The question is how was the model developed, not what happens when you get a code drop and press go.

The source code is riddled with non-physical fitting factors. I encourage you to look yourself: http://www.giss.nasa.gov/tools/modelE/

AFAIK, modelE is a very complex program; sadly it lacks complexity precisely in those areas more relevant to the debate and instead dives into what is arguable minutiae. Take this passage from their own summary:

“The soils are modelled separately for bare and vegetated
soil. Additionally, there is a canpoy layer over the
vegatated portion. The soils are modelled using 6 variable
depth layers. Evapotransiration takes water from the surface
pools, and from below as a function of rooting depths.
Runoff is calculated using a TOPmodel approach.”

(spelling errors preserved)

Reading stuff like that makes you wonder how the model manages to do anything meaningful at all. Too Many Variables.

James Lane

I’ll briefly chime in to echo comments made by others working with models. I also have extensive experience working with models in a commercial environment. It’s obvious that you can tune your model to replicate historical data, and that there are any number of ways to do it. To that extent I think Kiehl’s conclusions are banal. The interesting part is that they are being published by a mainstream climate sciientist.

Jon

#35 and others. Kiehl’s article is not generalized venting against overfitting but a specific empirical article linking model climate sensitivity to the selected history of aerosols.

Yes, he points out that some sort of fitting must be going on. This is what we’re harping about. His assertion that the fitting is a result of different aerosol datasets may well be true, but I doubt it is the aerosol datasets that are being manipulated. Rather, people are designing the model to get the desired match given whatever (honest) data they feed into the front-end.

bender

Bob Stoops is genius; this paper is not.

The paper starts with a premise that “climate forcing” and “climate sensitivity” are two (of three) important factors shown to determine Earth’s global mean surface temperature. It goes on to illustrate that “climate forcing” and “climate sensitivity” estimated in nine GCMs and two EBMs is negatively correlated. Is this a scientific discovery, or is it a tautology – true merely by definition? It seems to me to be an absolutely trivial result of the initial premise. If f=ma or d=vt or temperature=forcing*sensitivity, then any plot of the two factors against one another is going to lead to a hyperbolic curve of the exact shape produced in Fig 1. The result is predetermined by the premise.

I will let JEG decide whether tautology as proof qualifies as pseudo-science.

The author refers in the Introduction to the “curious result” that so many GCM parameterizations can lead to such a narrow range of outcomes. This also is trivially obvious. With a large number of free parameters and a number of different model formulations, this MUST be true, if all the models are being tuned to match a single output. Again, a tautology presented in the guise of a “result”.

The work involved in putting the paper together is not trivial, but the results are. This is pure pedagogy.

After all the effort putting this paper together, the author misses the most important point of all – which is the point made by Gerald Browning, and one made by several commenters here already – that multiple models all giving the same result using different tunings is not a sign of strength; it’s a sign of weakness. I won’t go further with that line of argumentation because it’s not my argument. (But we can go there tomorrow if you like.)

I close by asking the experts – like JEG: is this paper more “novel”, more “innovative” and more “publication-worthy” than, say, Loehle (2007) in EE? Who would like to step up and argue that case with me? Finally, where are the critics of EE and the defenders of GRL? Explain to me: who peer-reviewed this paper, and decided it was worth $9? I missed the game winning touchdown by Oklahama for this?

Save your $$$ folks.

bender

Kiehl makes a serious mistake in confusing actual climate forcing and climate sensitivity with modeled climate forcing and climate sensitivity. The mistake is made starting in paragraph 6 and carries through paragraph 8, to the conclusion in paragraph 17:

These results indicate that (i) the range of uncertainty in anthropogenic forcing of the past century is as large as the uncertainty in climate sensitivity and that (ii) much of forcing uncertainty is due to aerosols.

He is assuming that the true uncertainty is represented by the range of variable parameterizations used in the models (Fig. 1). This assumption is not necessarily true.

The true uncertainty on these “factors” is determined by the propagation of error through all the parameters (free and fixed, tuned and untuned) and equations that lead up to the computation of the “climate sensitivity” and “climate forcing” factors. Which he ignores. The “uncertainty” represented by the variation in the Fig 1 parameterizations is merely an indicator of random drift in arbitrary and subjective choices made by independently evolving modeling research groups. It is not the true uncertainty. It is human/institutional uncertainty, not physical/statistical uncertainty.

The argument advanced in the conclusion is somewhat dangerous. He argues that aerosols are the lynchpin, suggesting the community ought to target this aspect of the models for improvement. Sounds innocent enough. But what improvement for what purpose? The purpose IMO is not to improve individual model performance, but to shore up a consensus that has obvious weaknesses (e.g. Fig. 1 variability). The goal IMO is to reduce the human uncertainty evident in Fig. 1, with intent to infer that this is equivalent to a reduction in the physical uncertainty in the models.

The pea under the thimble.

Watch out for modelers that mistake their models for reality. See the hidden tautology. See the pattern of groupthink emerge. See the premium that is placed on getting this very mundane paper published. See the house of cards being buttressed.

Leonard Herchen

I think this paper is merely stating the obvious. Climate modelling is a history match effort similar to the reservoir modeling used by petroleum engineers (no irony there). To match the same data set of history, there will be a natural correlation between the way different modellers tune variables to get a match. It is not unscientific to tune a model to get a result. That is the only way to refine a model.

The fundamental problem with simulations is, as many readers know, is that there is no unique solution, so, even when we get a good history match, we can’t be sure if it is coincidence or not, and the value of a particular model as a predictive tool is nearly impossible to measure. A model with a perfect history match can go badly wrong as a predictor very quickly and there is no way to know in advance.

The anthropogenic climate change hypothesis is based on the observation that history matched models do not match if the CO2 forcing is turned off. It looks pretty convincing when you read the IPCC work or even one of Hansen’s first papers (Science 28th August, 1981, Volume 213, Number 4511, “Climate Impact of Increasing Atmospheric Carbon Dioxide”)

One way to test the hypothesis is to see if a good history match can be done without the CO2 Forcing using parameters that are physically believable. The “mainstream” climate scientists say it can’t be done. I remain unconvinced. I’m not sure how hard they are trying.

It should be noted that Hansen’s predicts in his 1981 paper that world temperature will increase by a factor outside of natural variabiity (he uses 2 sigma) by the end if “this century”, that is by 2000. If the surface temperature record is correct, that has occurred and his 1981 model has some support from predictive success.

The reality is that the range of outcomes suggested by Hansen in 1991 hasn’t actually been refined much by 20 years of climate modeling effort.

trevor

I have spent much of my career developing financial models that are expected to provide a “value” for a mining project. The standard methodology requires that analysts estimate the most likely operating costs, capital costs, and revenue streams, derived from the underlying tonnage throughput, grades, recoveries etc and then use the Capital Asset Pricing Model (CAPM) to derive the NPV of forecast revenues and so arrive at a “value” for the project.

I soon discerned that by judicious manipulation of key inputs, ie, just shifting most of the inputs 1-2% in one direction or the other, one could ‘push’ the resultant number wherever one wanted, and over quite a wide range.

In an effort to address the problem properly, I investigated Monte Carlo simulation and applied it to these models. This involved applying a probability distribution to key input variables and then running, say, 1000 iterations using a marvellous off the shelf program called @risk – an add-in to Excel.

The results are illuminating. To illustrate, 25 years ago, I was required to develop a value for a particular resource asset. I applied the CAPM approach (as is still the standard approach) and obtained a result of A$1100 million. At about the same time, I was asked to value another, much smaller, resource project. I applied CAPM, and achieved a result of $33m.

The essence of CAPM is that you establish the “Beta” (the volatility of the share-price or a proxy for the share price) and that helps you to calculate the appropriate discount rate. Since the discount rate for the first project came out (using CAPM) at around 9%, and the second project around 14%, the CAPM theory says that it is safe to pay $1100m for the first asset, and $33m for the second.

However, I then applied Monte Carlo, running 1000 iterations for each model. The outcome in the case of the first asset was that the Median NPV was $1100m, as expressed earlier, but 1 Standard Deviation was plus/minus $55m, or around 5%. This is very tight for a resource asset, and clearly shows that (subject to other due diligence) an investor can afford to pay $1100m for the asset, since it is readily financeable with that low SD, and of course, the nature of NPV calculations is that they don’t consider the “option” elements, of which there were many, most importantly the possibility of resource additions.

The outcome in the second case was wildly different. The Median NPV was still $33m, but 1 Standard Deviation was plus/minus $75m million, giving a range of “value” from -$58 million to $108 million. I learned that the term for this distribution is “flat kurtosis”. Of course, any of us faced with that additional information would be very careful about paying the CAPM valuation of $33m, since it is demonstrably not financeable. An astute investor might pay around $5m for the option value, but not much more. My question is, if this is the outcome, then how valuable is CAPM is valuing such assets?

To this day, so far as I know, Monte Carlo simulation is held in low regard in corporate finance circles, but, as I hope my example shows, it can be extraordinarily valuable in providing additional key information that can dramatically change the outcome.

Further analysis illustrated why the outcomes were as they were. The first project had a very stable revenue stream (for reasons that I won’t go into) and very strong operating margins. The outcome is that varying key input factors in the likely rangese did not have much effect on the margin, with the result outlined above.

By contrast, the second project was a high cost producer in its industry with narrow operating margins, and was very vulnerable to fluctuations in key input factors, particularly exchange rates and price assumptions.

In the event, the second project went belly up, while the first project thrived, and delivered much more than the $1100m value to its investors.

The point about going on at some length here is to point out that Monte Carlo simulation, properly applied, gives a direct measure of the merit of the valuation. I suggest that if similar approaches were used in climate forecasting models, that the results would show very flat kurtosis, ie pretty much equal probability of any outcome.

bender

#28
Yes, there you go. You have expressed it exactly but without resorting to the offending word: “in”. Much better.

Philip_B

“The station data are not used *in* climate models, and they are not used to predict future climate.”

This is a patently ludicrous claim. Data is used all models. That’s pretty much the definition of a computer program. Data comes in, process occurs, data comes out (ie whether there is any data or reference to the data in the code is irrelevant). Climate models must use initial values in their processing. So is he saying they start the models from random values. Of course they don’t. They use station data however processed pre-satellite and it appears post-satellite as well).

What he means is current station data is not used. Well, duh. The point of a model is to predict. And to think these people get paid for this kind of analysis.

Ferdinand Engelbeen

Re #22,

Christopher, the next course can be found at: http://cpd.conted.ox.ac.uk/env/courses/modelling.asp. The individual fee is quite high, but if you are a member of any NGO (no matter what, doesn’t need to be climate related), it is bearable. We combined it with a nice round trip in Ireland and spent a few days in Oxford…
The Excel spreadsheet is propietary of the Oxford Environment Department and may be used up to 1 month after the course. I obtained a longer (life-time?) allowance, but may not share it with others.

But in fact, you can build it yourself, as the data series are available (besides that you can use different series for aerosols and solar forcing), and the effect of a change in any forcing on the ocean heat content (the main short to long-term influence on temperature) is known too…

The nice part is that you can change each individual effect factor for any individual forcing. That means that one can change the effect of e.g. CO2 vs. solar changes in forcing. In current climate models, they have similar effectiviness for 1 W/m2 solar as for 1 W/m2 GHGs (+/- 10%), while that is certainly not true (solar and volcanic have their highest influence in the stratosphere, GHGs and aerosols in the troposphere).

Willis has done some multivariable atttribution experiments in that sense, but I have seen that he differentiated a little too much between many variables (like black aerosols and sulfate aerosols). If he may repeat that with only four variables (all GHGs lumped together, all types of aerosols lumped together, solar and volcanic), the “best fit” might be more clear. Even that doesn’t mean that the “best fit” is what really happens in nature, as most data series themselves have rather high error margins and as you can see in the one can change the GHG/aerosol tandem within wide margins with little influence on the correlation of the model result with temperature.

Ferdinand Engelbeen

Steve McI,

I noticed a discrepancy between graph 2 of Kiehl and the graph used as reference at RealClimate:

The graph at Realclimate is right: if the influence of aerosols (forcing or sensitivity or both) is low, one need to lower the sensitivity for GHGs too, or the 1945-1975 temperature dip can not be reproduced and/or the increase after 1975 is too fast.

I have not purchased the Kiehl article yet (and didn’t go to the library either), thus I don’t know why they show opposite graphs…

And of course, graph 1 of Kiehl is only true because the models use similar sensitivities for all kind of forcings…

Tony Edwards

Pat Keating #19 says “Have you read Gerald Browning’s posts over at the Exponential thread, where he says even the near-term weather-forecast models blow up trying to predict past 36 hours?”
Shouldn’t that be postdict?

bender

Do we know which models are associated with which points on the two graphs?

Anthony Watts

RE46, Thanks Bender. I always suspected as much, but scanning old Fortran code is about as pleasant as having a tooth pulled, so I never got around to checking Gavin’s challenge, having a project to run.

Henceforth I shall say then, that “surface data is used to tune the models that predict our climate future” where such a description may be appropriate. I now beleive that statement to be fully accurate, unless evidence is presented otherwise.

bender

Re #59 Where’s the code? I’ll audit it.

bender

#60
More of the same. See the objective here?: Get the models to converge on one another (“reduce uncertainty”) to make it appear that there is consensus on the various forcings and sensitivity parameters. Re-read #41. That ain’t prediction uncertainty, folks. That’s institutional uncertainty about how to model climate. Ask yourself: what is the purpose of “reduced uncertainty” on aerosol forcing? Better forecasts? Increased physicality of model structure & realism of model behavior? No. Lower variation in institutional opinion. Why? New clothes for the emperor. The policy makers are embarrassed by Kiehl Fig 1. Why? Because they don’t understand what it means. They don’t understand that it’s inevitable. That:

The parameterizations are extremely crude in the climate models.

Has anyone asked why there are so many climate models and NWP models if the
numerical errors are small? The obvious answer is that there are so many different eddy viscosity types (that are too large) and so many different parameterizations that are used to overcome the inappropriate
viscosity in order to obtain a spectrum that kind of looks reasonable.

It has been known for some time that the highest resolution global model (ECMWF) has not converged to a better forecast exactly for this reason.

This is science?

Let’s ask JEG what kind of science he thinks this is.

[DISCLAIMER: Do not take my word on these issues. Investigate them for yourself. I am not a climatologist. My hunches are purely speculative.]

Jon

Bender, #61:

I posted the link to modelE earlier.

Anthony Watts

RE66,

Bender there’s also “model E light” aka EdGCM at: http://edgcm.columbia.edu/

Which unlike Model E, does not require you to install a FORTRAN compiler on your machine to run.

lucia

@Bender #31.
With regard to how the fitting is done, you may or may not be correct. Gavin describe the way tuning is supposed to be done in principle. In practice, there are several difficulties:

a) individual laboratory experiments may give a range for a parameter. The range may be large or small, depending on what you are trying to parameterize.

b) various different forms are suggested for parameterization. That is, entirely different sets of equations might be proposed. All, in principle, may looks ok compared to a laboratory experiment. (Generally, more complicated forms can give the illusion of capturing more physical processes, but they generally also create more tunable-constants, take more computational power, and have been tested in fewer experiments.)

c) laboratory experiments are more or less “clean”. One might do one experiment with only forced convection. Another with pure free convection and so forth. One might try mixed convection over some particular range, but with a very simple geometry. Sometimes, there really aren’t experimental data that fully test a parameterization outside a particular code.

Mind you, some parameterizations suffer from few of these difficulties and are bullet proof. (But, I’d guess the ones in GCM’s are NOT bullet proof. That’s likely why JimD mentions that tuning is an art when discussing tuning on the exponential thread. )

But, some parameterization is required to create a code, so the person assembling the full model picks a code and a parameter.

In principle, that would be that. You picked the ‘best’ values based on lab experiments.

But in practice, model are run by humans who do want their results to look good. Any modeler remembers they had to make a choice about the magnitudes of “constants” and also about the specific closures to select. If the model or tunable-constants results in poor predictions for the global model, people do start adjusting the constants to see which best match the test data (in some sense.)

This isn’t necessarily a bad thing to do, but you have identified the general problem. The knobs do get adjusted someone to fit the existing data. In most fields the modelers would list the magnitude of all values of “tunable constants” one might have selected, and the papers describing what the magnitudes are in lab experiments should be mentioned.

In the thread on exponential growth, you’ll notice I’m asking Jerry B. to explain precisely why he thinks the closure is unphysical. That’s because I’m trying to figure out how far people are turning knobs.

From my point of view, getting the answer for on particular parameterization would be enlightening, even if it’s not the most important one.

The answers to the questions I’m trying to pose would tell me whether the truth of what modelers are doing is closer to what Gavin says (which is possible) or what you suggest (which is also possible.)

For better or worse, both methods of getting parameters have been known to occur. It should be possible to tell which is happening if someone who runs GCM (possible Gavin) will answer with sufficient detail. (As in: for example, our boundary layer model uses a blah, blah blah type closure proposed by joe, fred and holly — give reference. Model constants were obtained from experiments conducted by Marvin- give reference. )

Boris

Why? Because they don’t understand what it means.

Oh, come on now. This thread is getting ridiculous.

bender

#71 It doesn’t bother me that you don’t agree with my hypothesis as to how Kiehl’s Fig 1 might be interpreted by policymakers. It’s a side point that is fairly irrelevant to the overall argument, which is that there is a move afoot to shore up the “consensus” surrounding the models.

Which leads to the fairly robust, testable prediction that the literature is about to get silly with efforts to band-aid that which is not easily fixed.

bender

The genuflecting is nauseating.

Ivan

Second, many of the emission scenarios for the next 50 to 100 years indicate a substantial increase in greenhouse gases with associated large increase in greenhouse forcing. Given that the lifetime of these gases is orders of magnitude larger than that of aerosols, future anthropogenic forcing is dominated by greenhouse gases. Thus, the relative uncertainty in aerosol forcing may be less important for projecting future climate change.

But, problem is high sensitivity for “projecting future climate change”. Aerosols may or may not become less important as a forcing in the future, but none can establish “desired” high GHG sensitivity without attributing very high cooling effect to aerosols in XX century. So, it’s not clear how Kiehl can eat the cake and to have it in the same time: if you didn’t establish properly large cooling effect of XX century aerosols (as you didn’t), you also didn’t establish high CO2 sensitivity – from here to eternity, whatever happens in the future. Period. Whether aerosols will became less important factor in future is therefore completely irrelevant.

bender

From RC:

bender Says:
18 May 2007 at 12:42 AM
Parameterized models typically fail soon after they are used to simulate phenomena in domains outside those in which the model was parameterized. As climate continues to warm, the probability of model failure thus increases. The linear trend fit of the Hansen model to observations during the last few decades thus does not imply that we can expect more of the same. This is why I am interested in the uncertainty associated with the parameters that make up the current suite of parameterizations. How likely is it that those parameters are off, and what would be the consequences, in terms of overall prediction uncertainty? Is moist convection the wild card? Why or why not? Tough question, I know. Possibly even overly presumptive. Maybe the best way to answer is through a new post?

[Response: You need to remember that the models are not parameterised in order to reproduce these time series. All the parameterisations are at the level of individual climate processes (evaporation, cloud formation, gravity wave drag etc.). They are tested against climatology generally, not trends. Additionally, the models are frequently tested against ‘out of sample’ data and don’t ‘typically fail’ – vis. the last glacial maximum, the 8.2 kyr event, the mid-Holocene etc. The sensitivity of any of these results to the parameterisations is clearly interesting, but you get a good idea of that from looking at the multi-model meta-ensembles. One can never know that you have spanned the full phase space and so you are always aware that all the models might agree and yet still be wrong (cf. polar ozone depletion forecasts), but that irreducible uncertainty cuts both ways. So far, there is no reason to think that we are missing something fundamental. -gavin]

Bold mine.

I wasn’t suggesting they are “missing something fundamental” in their models. I was asking what was their confidence was in their parameters. i.e. What’s the probability they’ve inadvertently built in a warm bias? I asked for a post on the topic of model tuning. e.g. What are these “out-of-sample” scenarios? How many knobs? How do you handle the potential non-ergodicity of oceanic flows (i.e. non interchangeability of time-series and ensemble-series and ensemble-“ensemble” series)? etc. I did not get the post that I politely asked for. Instead, I got snowed.

There were two other times when the same thing happened. On equally important questions of how GCMs are tuned.

Why don’t they answer questions, if they are the high priests of climate science? I don’t get it. Steve M answers questions asked of him. Moreover, the tougher they are, the more likely he is to answer.

bender

Ken, it’s #79 that gets my goat. Kiehl (2007) appears to be functioning as a whip to the modeling community. A call for even more groupthink.

bender

Appeal to authority #78 is fine with me, although it’s usually more helpful if you can be specific about what it is there that would help to advance the discussion.

I’ve had my say for the day, and I’m sure Steve will agree: that’s quite enough.

Frank H. Scammell

I find it interesting that a lot of the shouting and hollering by the AGW/RC crowd is about catastrophes, tipping points etc., because these events are clearly nonlinear and probably chaotic. When you are talking about climate sensitivity (to a doubling of CO2, for example), that is a linearized estimate. The models do not permit chaotic solutions because they are highly damped (atmospheric viscosity, boundary conditions, etc.). Otherwise, there would be exponentially divergent solutions (some of which would show -horrors- lower temperature histories -predictions – projections – estimates for the future.) The models are seeking an equilibrium solution some time in the future when it is obvious that the climate is a non equilibrium entity (non-stationary). It may be that most of the equation set are real physics, but non physical elements are required to deliberately constrain the equation set, in order to produce a stable solution. Monte Carlo runs assume that equations have been linearized. It would be quite unsettling if a Gaussian distribution among the variables produced a bifurcated, clearly nonGaussian output. Hard to estimate a climate sensitivity !

Frank K.

“I wasn’t suggesting they are “missing something fundamental” in their models. I was asking what was their confidence was in their parameters.”

As I mentioned in an earlier response, there is no way, for modelE specifically, that you can know if they’ve “missed something fundamental” because they have NOT adequately documented their code. I REALLY is that bad – and I don’t think Dr. Schmidt will be doing anything soon to rememdy the situation.

I encourage everyone to go have a look at the FORTRAN. It’s nearly impossible as it stands to determine if they have everything implemented correctly. These kinds of research codes frequently break when you go outside some established range of parameters because the authors in many cases cannot anticipate every possible permutation and combination of inputs. Add to this the fact that the various modules interact with each other in a myriad of complex ways, and you have even more opportunities for problems. And yet it is the output from these codes that is being used to “predict” global climate distaster. Accordingly, it is NOT too much to ask for some due dilligence on the part of NASA/GISS with repsect to code documentation!

Follow the Money

#85, Frank,

I find it interesting that a lot of the shouting and hollering by the AGW/RC crowd is about catastrophes, tipping points etc., because these events are clearly nonlinear and probably chaotic.

Tipping points etc. are for political/public consumption to scare nations recalictrant to enact Kyoto financial market carbon offset schemes by creating a sense of crisis. Whether Hansen invented the “tipping point” talking point, or it was put into his hands by the wordsmiths at Oglivy Mathers is still unknown…

Australia has now elected Kyoto profiteers, the USA is holding out, hence the recent emanation from Bali about how USA joing Kyoto is “essential.”

henry

One way to test the hypothesis is to see if a good history match can be done without the CO2 Forcing using parameters that are physically believable. The “mainstream” climate scientists say it can’t be done. I remain unconvinced. I’m not sure how hard they are trying.

It sounds like the CO2 forcing could be TOTALLY REMOVED, and still the tuning could give a good match within the models.

Or am I seeing it wrong?

Wansbeck

re Al #92

In my view this is a sound point.
However, with the presently disputed models won’t people simply check for a match with the sequestered data, ‘tune’ accordingly, and then remodel?

JS

Responses about good ‘out of sample’ performance deny the reality of model making. The only good ‘out of sample’ data are data that were unavailable to the researcher when they chose the model. Thus, you find many models can break down 5 years after they are completed because that is when data truly unavailable to the researcher becomes available.

In most cases, ‘out of sample’ testing is no more than parameter stability testing. But robustness to out of sample data sounds sounds so much more impressive than ‘parameters are stable within subperiods’.

So, in #92, I must applaud your effort. Most researchers probably can’t resist the urge to peek at the data. But even then, you will only publish a result that demonstrates parameter stability – not one that can be guaranteed to have good forecasting performance. For that, one must wait a number of years. And in that light, so very few models of complex systems survive.

I only wonder whether or not Gavin/GCM modellers have enough introspection to realise the frailty of their process and enough balls to confront it head on.

JS

#96
I’m not suggesting incompetence. Just that the process of building models it very complicated and it is easy to present the model as having more robustness than it does. The models can be better (although obviously not perfect) if the practitioners acknowledge and respond to the inherent weakness of the process.

Interesting that you should mention LTCM. The underlying structure didn’t change – it has always been what it was. Rather, an assumption that they relied upon turned out not to be true. It was true enough for a while, but after a while it wasn’t. Keynes got it best “The market can remain irrational longer than you can remain solvent.” (Many models assume a liquid market to dispose of bad investments – markets usually turn illiquid right when you want to dispose of bad investments. Just look at the sub-prime meltdown; people don’t want to buy at any price.)

hswiseman

Could someone provide feedback on my thought process set forth below? I got an A- in differential calc. (my last math class) as a freshman 30 years ago. I may be completely out of my depth here using layman rationales and what seems like common sense instead of the actual mathematics and stats. Feel free to say so. Thanks in advance (strapping on bulletproof gear)…..

If the modeler is entitled to tune the forcings or feedback value of the model parameters, it can only be because the potential range of errors for the adjusted parameter supports the alternative value. This is really the best case scenario scientifically, as the modeler chooses an adjusted value within support range of the research error bars in order to produce a better fit. A less acceptable alternative is parameter adjustment outside the error bar range, supported by the proposition that “I know all my other values are pretty darn close, so the science must be wrong on this parameter.” Using the GCM to solve for a parameter value is AKA “just makin’ the s**t up” and the error range in this technique would be huge. I think this is essentially what Kiehl 2007 tells us. The error bars for the GCM itself should be some derivative of the number of parameters that are fair game for tuning and the latitude allowed within error ranges for such parameters (as well as the base level of certainty/error range in the science for the underlying error range). Any way you slice it, if you are taking liberties with the parameters, it must be because the underlying science of the particular parameter is uncertain or inadequate. The error bars for the GCM once overfitting stopped (i.e. for predictive future results) would be huge as you have layers of compounded uncertainty baked into the cake, and probably render the model useless as a predictive tool. A massive ensemble run might remove some of the outliers, but the model would still have error ranges far greater than the average individual parameter error ranges. As the model propagates over time, the initialization error range should follow some compounding function (the rule of 72’s comes to mind). A theoretical 10 percent initialization error range from GCM ensembles might approach 100 percent after about 30 years.

Mhaze

#79 bender says:
December 2nd, 2007 at 2:33 pm

Part of quote from RC:
[Response: You need to remember that the models are not parameterised in order to reproduce these time series. All the parameterisations are at the level of individual climate processes (evaporation, cloud formation, gravity wave drag etc.). They are tested against climatology generally, not trends. Additionally, the models are frequently tested against ‘out of sample’ data and don’t ‘typically fail’ – vis. the last glacial maximum, the 8.2 kyr event, the mid-Holocene etc……. -gavin]

Such assertions are rampart regarding climate models. Where is the verification, the proof of such assertions? Where is a history of parameters for just one of the models?

bender

#98
Read what Harry said. He said you can change inputs by X amount and have no effect on ouput. He didn’t say that was always going to be the case.

bender

#104 You’re right; the context of his quote changes its meaning significantly. My mistake. I’d read #94 but not connected it with your quote.

I’m no climate modeling authority. I have no comments on anybody’s perceptions or misperceptions of what the climate modelers do or don’t do. I don’t know what they do. That’s what I’m trying to find out! My advice is to go talk to them. And report back to us what you hear.

Carl Gullans

#92: I think Bender has addressed this when referring to having seperate people control the data and the modeling process, respectively. I build models on training data and test them on validation data that’s been completely untouched (the exception being removal of bad records from both data sets) up until that point. The problem is that you have to trust that the modelers haven’t peeked at the validation data beforehand…

I think I asked this here once before, but is there anybody who has compiled many/most of the GCMs from the 90s and 00s and compared actual vs. predicted? It shouldn’t take long for the model to fail if it is incorrectly specified. Do any of them predict well for 5-10 years (a link to an ’06 McIntyre post in this thread shows that at least one did pretty poorly).

bender

Re #106

This is at least partly true.

Yes.

That said, there’s a bit of slight of hand here

Yes.

Jaye

A few years ago I was hired by an old boss to help with some problems he was having with some of this current staff members. The problems were related to a specific incident. These guys had developed what they thought was a very good prediction tool, it was a maximum likelihood estimator with a grab bag of techniques, some fuzzy logic, some neural nets, some of everything. They were trying to predict the oil futures market. So they got some data from a group of investors and SME’s in the oil future business. Of course, they split the data for a training and a validation set.

Ok so time for the presentation of the results to the potential investors, who had their experts in attendance. They all walked out after the first graph was presented. Seemed that our guys had “predicted” a spike in the prices, except that this spike was due to a refinery blowing up. The only way to do this was to cheat and use the training data to do the predictions. We shut down that project and went on to more useful things. Too bad the IPCC and others aren’t as decisive as these potential investors.

bender

Re #112 Who wouldn’t?! Sure, that level of documentation would increase the expense of the work, but I think it would be worth it.

DharmaHunter

On a philosophical note, I think the underlying question driving this discussion is: what is the utility of modeling?

On a short term scale, the global models are very important in predicting the short and medium term weather for specific locations. Due to their different parameterizations each of the most popular weather models have different biases. It is the skill of the individual meteorologists to digest the model outputs and current weather conditions to make specific forecasts, be it a snow storm next week or a hurricane in the Gulf. The utility of these models are tested on a daily basis, and are assessed and graded on their skill of prediction. The National Hurricane Center, especially under the direction of Max Mayfield, has instituted a fairly rigorous set of criteria to their models and predictions. One model under development is the very tricky prediction of hurricane strength, which was alluded to in a few of their discussions. The goal is to extend their prediction time and increase accuracy. The former is important for communities to prepare and the latter is especially important to avoid “chicken little” forecasts. The NHC discussions are especially informative with regard to the certainty of the forecast. They also refer to the models with the phrase “model guidance.”

The intermediate term scale, say the 3-6 month window, is important for agriculture, hurricane frequency, and water resource management. These of course are much less accurate, for good reason: it is difficult to predict storm paths and frequency when they have not formed yet. The best that you can do is to forecast general conditions which influence specific, stochastically occurring events. Although there is certainly room for improvement, the ENSO prediction models have done a decent job.

In the long term, the climate models attempt to take given a certain set of conditions (e.g., CO2, aerosols, etc.) and simulate the general climate, specifically temperature. However, the public has latched onto “predictions” of drought, floods, hurricanes, rising sea levels, heat waves, etc. All of these depend upon the influences of the general climate conditions on regional weather. From what I have read, the GCMs can capture the general features of climate, such as the ITCZ and Hadley cell, but perform poorly on specific features such as the ENSO. Since the ENSO is such an important driver of regional weather and temperature, there is certainly room for improvement beyond the projection of future global temperatures.

With respect to model predictions, have the current suite of GCMs performed better than a prediction from, say, 1971?

A Nature editorial (The great greenhouse scare, Nature, 229:514, 1971) reported on a publication by a study group, “1970 Study of Critical Environmental Problems” (MIT Press). Their predictions for year 2000 were for an 18 % increase in CO2 and a corresponding increase in temperature of 0.5 ºC – the actual numbers ~14 % and ~0.5 ºC. Now, I figure (can’t find the report online) these estimates were based on clear air effects of CO2 on temperature and estimation of CO2 increase from the existing trend and projected fossil fuel consumption. While these estimates verified as well as the as the GCMs despite their difference in sophistication, have the GCMs significantly increased our understanding of the workings of the atmosphere?

With regard to this last question, Isaac Held of GFDL has written an interesting essay entitled, “The gap between simulation and understanding in climate modeling.”

Click to access 21.pdf

In it he calls for development of a hierarchy of models, from large, comprehensive global models to more specific models of different climate processes. He asks,

What does it mean, after all, to understand a system as complex as the climate, when we cannot fully understand idealized nonlinear systems with only a few degrees of freedom?

When the comprehensive models give an output of a specific temperature, do we actually know why? The inverse correlation between anthropogenic forcing and sensitivity as well as the direct correlation between total and aerosol forcing suggest that the key underlying processes are not completely understood (clouds, precipitation) or there are processes that are missing. Held goes further to emphasize building a hierarchy of models to better understand the workings of the atmosphere. The need for such a hierarchy derives from his view:

When a fully satisfactory systematic bottom-up approach to model building is unavailable, the development process can be described, without any pejorative connotations intended whatsoever, as engineering, or even tinkering. (Our most famous inventors are often described as tinkerers!) Various ideas are put forward by the team building the model, based on their wisdom and experience, as well as their idiosyncratic interests and prejudices. To the extent that a modification to the model based on these ideas helps ameliorate a significant model deficiency, even if it is, serendipitously, a different deficiency than the one providing the original motivation, it is accepted into the model. Generated by these informed random walks, and being evaluated with different criteria of merit, the comprehensive climate models developed by various groups around the world evolve along distinct paths.

The value of a holistic understanding of climate dynamics for model development is in making this process more informed and less random, and thereby more efficient.

He points out as an example of hierarchal model building, the success of biological research (with which I am familiar). He rightly points out that nature has provided biologists with a natural pre-made hierarchy of systems, from simple to complex. The success of biological research has been in defining the circuits and interactions. Indeed, structural research on biological nuts and bolts informs larger complex systems such as cell cycle control. It has only been recently that numerical simulations have been applied to biological systems, with some very interesting results. It should be said that many biologists are distrustful of these models and question their utility. This is likely because the emphasis in biology has been on experiment with very little theoretical development of the workings of biological systems, though this is changing.

Conversely, climate science has emphasized theory over experiment, which is not surprising since its tough to isolate the ENSO and study it in the laboratory. Held’s view is that smaller processes have to be studied and understood before we will be able to judge the strength and validity of the models. For example he asks,

…if stratosphere-troposphere interactions in one comprehensive model result in a trend in the North Atlantic Oscillation as a result of increasing carbon dioxide, but not in other models, how does one judge which is correct? One can try to judge which model has the most realistic stratospheric-tropospheric interactions by comparing against observations, with the understanding that theoretical guidance is required to design these comparisons. One can also analyze more idealized models designed to capture the essence of the interaction in simpler contexts, within which the climate dynamics community can focus directly on the central issues. These idealized studies can then suggest optimal ways of categorizing or analyzing more comprehensive models. If we are to claim some understanding of this issue, our modeling hierarchy, from idealized to realistic, should tell a consistent story.

He goes on to suggest several systems that would be of value to study. I think that this is a worthy goal. In biological research (specifically molecular biology), the gains in understanding different systems have been through the development of techniques that reduced the number of degrees of freedom. I think this is what he is saying with regard to climate research. By concentrating on specific systems, such as moist convection, both experiments and theory will be better defined and understandable. In his concluding remarks he states,

..without the solid foundation provided by careful study of an appropriate model hierarchy, there is a danger that we will be faced with a Babel of modeling results that we cannot in any satisfying way relate to one another.

Scott

I want to reiterate, as a guy somewhat familiar with modeling…it’s all about parametrization.

Willis Eschenbach

Dharmahunter, thank you for your most interesting post.

I had to laugh, though, at the last sentence you quoted from Held …

..without the solid foundation provided by careful study of an appropriate model hierarchy, there is a danger that we will be faced with a Babel of modeling results that we cannot in any satisfying way relate to one another.

Will be faced? Will be? We’ve faced it for years …

Figure 1. Notched boxplot of the Santer data and model hindcasts. Gray horizontal lines across the figure are the 95% confidence intervals on the medians and quartiles of the data. Notches in each boxplot show the 95% CI of the median of that data. If notches the notches of two datasets do not overlap, they are outside the 95%CI, and are statistically distinct.

These are boxplots of the observational data and the models from the Santer study, I could find the URL, hang on … OK, Science mag, subscription required …

Now, there’s some things of note here. First, the pattern of the two observational datasets is very similar. They are both compact, ocean temperatures don’t move around much, it takes a lot of power to warm or cool the ocean. They are both asymmetrical with regard to the median, being longer on the top. They both have numerous outliers, with many more outliers above than below the median. It is a curious and distinctive pattern.

Now, compare that to the model hindcasts. There is one, and only one, model that produces what I call “lifelike” data. The rest are all quite unlike the real stuff. And clearly this, the simplest of comparisons, has already produced what, in Isaac Held’s lovely turn of phrase, is assuredly “a Babel of modeling results that we cannot in any satisfying way relate to one another.” What can we make of model CM2.1, which has the temperature flailing all over the place? Or CCSM3, which periodically hindcasts an oceanic cold swing twice as deep as anything in the record?

If it were me, I’d likely toss out all but one of those models, the one that is lifelike, and go find another batch to test. And that’s on the first and simplest of tests, that the modeled and observed datasets have to be similar. The tropical ocean has no mountains, no deserts, no ice or snow … man, if you can’t get the temperature of the tropical ocean right, good luck with the rest.

And yes, I know that being “lifelike” is not a statistical test of any kind … but it is the test that comes before all the fancy statistical measurements. You don’t need statistics when a supposedly “lifelike model” of a princess actually looks like a frog … you go “green skin … check … webbed feet … check … NEXT!” …

Here’s another look. This one compares skew, kurtosis, and power. Skew and kurtosis measure departures from normality. Skew refers to the asymmetry I spoke of before, more data on one side of the mean than the other. Kurtosis is how clumped the data is around the mean compared to a normal distribution. Power is the amount of energy moved to warm or cool the ocean. Here’s a chart that compares the models and the observational data:

Figure 2. Skew, kurtosis, and power for the models (white) and the observational data, shown as probability clouds. The center of the cloud is the data point. The width, height, and depth of the cloud represents one standard error in the respective direction.

As you can see, the observational data are distinctly non-normal, with both large skew and kurtosis. The natural system also expended significantly less power than any of the models. The ocean is simply not doing, and has never been recorded to do, what the models say it is doing. Natural systems, in accordance with the constructal law, are not profligate of power. The models haven’t figured out that part of the climate system yet. The ocean doesn’t get real hot one month and real cool next month, never happened. Nor does it hardly vary from month to month, as some models have it doing. And whatever it is doing … the models are a long ways away.

So, while I commend Isaac Hurd on his ideas and his suggestions, I would respectfully suggest that things are far, far worse than he thinks. Babel is upon us.

w.

Hans Erren

aerosols are the ultimate fiddle factor.

Geoff Sherrington

Re # 12 Ferdinand Englebeen

The influence of aerosols in the 1970s alleged cooling period is more than a discomfort in trying to explain the cooling. What worries me is that the cooling component became greater and greater to the present, so although the temperature graph rises recently, it throws out of balance the attributed causes of the rises and their magnitude. It’s like motion on a waterbed. Put your foot down in one place and it bulges in another. Now try to set up a comprehensive model of waterbed motion during a romantic interlude.

I find it hard to ascribe quantitative effects to qualitative processes. For the same reason I am VERY suspicious about oxygen isotopes in ice cores as temperature proxies.

Tom Vonk

I wasn’t suggesting they are “missing something fundamental” in their models. I was asking what was their confidence was in their parameters. i.e. What’s the probability they’ve inadvertently built in a warm bias? I asked for a post on the topic of model tuning. e.g. What are these “out-of-sample” scenarios? How many knobs? How do you handle the potential non-ergodicity of oceanic flows (i.e. non interchangeability of time-series and ensemble-series and ensemble-”ensemble” series)?

Bender you should have affirmed (not suggested) that they missed something fundamental .
Actually 2 HUGE somethings fundamental .

1) The proof of convergence

Of course if you model , even very wrongly , a system that you submit to strong constraints , the numerical run will not blow up because of teh constraints .
For instance energy , impulsion and mass conservation are extremely strong constraints that will make sure that your model will not go wild and unphysical .
However that is not convergence .
Convergence is that you prove that the computed dynamical states of the system are uniformly converging to a certain trajectory in the phase space when the size of the step goes against 0 .
For the sake of the argument imagine that you take a spatial step of 1000 km and a time step 1 month .
A very crude model indeed that could run on your PC and behave very stably .
Yet even this model will get some things “right” – Arctics will melt in summer and freeze in winter and with the hypothesis of well mixed gases you can even implement an extremely sophisticated radiation transfer .
Now the problem is that everything below this size is unresolved and that means that even if you know the laws of physics on the sub grid scale , you have to substitute in them monthly averages .
And of course that is a big sin against physics because the laws of nature are local and instantaneous and being mostly
non linear as they are , they never work for averages .
So now you take 100 km and 4 hours . Many of the previously subgrid phenomenons must now be explicit and you obtain a new dynamical state that will be rather far from the previous – even if Arctics stil melt in summer and freeze in winter .
What would happen with 1 km and 30 minutes ?
Nobody knows because the smaller gets the scale , the greater the sensibility and the risk of wild divergence at longer times .
So not only the convergence is not proven but it is not clear how to tackle the problem .

2) The definition of metrics and stochastical behaviour

The problem above is a problem of a trade off – with crude resolution you get stability and are far from reality , with fine resolution you are (supposed to be) near reality but get unstable .
Unfortunately the problem cannot be treated mathematically because the GCMs do NOT solve differential equations in which case it would be possible to speculate about the existence , regularity and features of the solution and the ability to approach it with numerical simulations .
What the GCMs basically do instead is to simulate the system by having N equations tying together N variables .
The computer solves for t and moves to t + 1 .
That’s why it has no problem with the initial and boundary conditions that are crucial for every physical problem .
Now what could mean this series of computed numbers ?
Certainly NOT a solution to the uderlying differential equations because they are partly unknown and if they were known , their number would be astronomically huge .
The modellers then make the following assumptions : unresolved size is random fluctuations and even if one simulation is not equivalent to a prediction , the difference between 2 simulations with different values of a parameter represents the sensibility of the system to this parameter .
Such bold assumptions need proofs .
First there is no necessity than anything subgrid be random and indeed it mostly isn’t – the reason is the already mentioned fact that the law of nature are local and are not valid for averages .
So everything subgrid (with a 100 km grid !) dismissed by hand waving and saying that it’s random and that it “averages out” must be proven what is again not possible because the exact solution of the system is unknown – the ergodicity you mentioned is also part of this kind of problems .
Secondly to interpret the difference between 2 runs as a sensibility , you need metrics .
What is the metrics for the validity of GCMs and/or for comparisons between 2 GCMs ?
Temperature not and “global” temperature surely not .
R.Pielke has written n times that the only relevant metrics for the climate is the energy content of the oceans and I think that he is right .
It doesn’t seem that the “climatologues” have ever cared to benchmark their models against reality and among themselves in this metrics .
A rather big fundamental something in my book .

Richard S Courtney

Dear Steve:

At point#15 you say;

“I haven’t seen any prior paper specifically showing that model climate sensitivity had an inverse relationship to the version of aerosol forcing history. People may have hypothesized something similar but this is more substantial.”

I published a paper that showed this is true some years ago, but it was only demonstrated for one model.
Ref. Courtney RS, “An Assessment of Validation Experiments Conducted on Computer Models of Global Climate (GCM) Using the General Circulation Model of the UK Hadley Centre’, Energy & Environment v.10 no.5 (1999)

All the best

Richard

bender

#124 I suspect Dr Schmidt would have then quibbled over the meaning of “fundamental”. But I think the thread is still open if you would like to follow up, Tom. Actually – I did bring up the ergodicity issue at a later date. But then I was assured that the internal variability in terawatt heat engine Earth’s climate is miniscule compared to the external forcings. I believe it was “mike” who told me that. Unfortuantely, he did not supply a proof of that assertion. I have no way of determining how true or false that statement might be.

Steve has mentioned Dr. Demetris Koutsoyiannis work in the past, and I thought I would provide a link to his website. In particular, there is a paper there, titled Climate change, Hurst phenomenon, and hydrologic statistics, from which I would like to quote:

The intensive research of the recent years on climate change has led to the strong conclusion that climate has ever, through the planet history, changed irregularly on all time scales. Climate changes are closely related to the Hurst phenomenon, which has been detected in many long hydroclimatic time series and is stochastically equivalent with a simple scaling behaviour of climate variability over timescale. The climate variability, anthropogenic or natural, increases the uncertainty of the hydrologic processes. It is shown that hydrologic statistics, the branch of hydrology that deals with uncertainty, in its current state is not consistent with the varying character of climate.

In his view, statistical hydrology is in its infancy. This is an important statement because it suggests we have only begun to describe ocean hydrological behavior in statistically robust ways. If that is true, it means that GCM tuning is far more art than science.

I have not read all of Koutsoyiannis’s work, but I like what I have seen so far. It supports the hypothesis that an ASA Journal of Statistical Climatology is badly needed.

Harry Eagar

Yes, Pat, when I said ‘can’ I meant, ‘the real world already has tried it.’

1 ppm CO2 would mean no life on Earth any more, with resultant effects on, among other things, aerosol production, so I expect the models would be far outside their parameters by then.

kim

I laughed at another blog when someone dismissed my authority as a ‘little known water hydraulics guy’.
===============================================================

kim

Better:

The Wexmen Cometh Whence?
===================

bender

Suppose each “knob” is a “card” …

bender

If I were to “huff” and “puff” on those “dots” …

bender

Discerning cause-and-effect is very hard when you have a delayed feedback loop A<->B that is being forced by other agents, C.

Frank K.

Re: 136

I poked around the climateprediction.net site. Not a differential equation in sight – in fact very few equations at all!! Are there links to actual documentation of their code? Maybe I was overlooking something…

Craig Loehle

It is disengenuous to say that the models are not “tuned” to trends. If they were purely from basic climatology, then you would fix the parameters from studies of clouds, el nino, PDO etc and then let it rip—but in fact early versions of the models gave an ice covered earth or other crazy behavior. The obsession with aerosols in the models is because without them the models won’t replicate the cooling from 1945 to 1980 (or whatever). This part of the model was NOT calibrated from the ice age model comparisons or other such paleo-tests. To prove that the models are NOT “tuned” to 20th century trends, an audit trail would be required: where, when, why and by whom each model change and parameter was derived. Guess what: no audit trail. Sorry Bender—you will never get your answer because no one wrote it down.

Bob Koss

RE: 142

Frank K.

I doubt there is any documentation there. I think they have a forum of some sort though. Don’t know if it’s public posting. You might try asking there.

Went there quite a while back. Knew I wasn’t going to be too impressed when on their home page I saw they post the number of model years run to three decimals. Had to look twice to read the figure correctly. I was off by a factor of 1000 with my first read. An 8.76 hour resolution. That’s nutty. Probably did it intentionally. 😉

Willis Eschenbach

The Koutsoyannis paper bender mentioned above is a very important one. The models do not give reasonable Hurst coefficient results. Here’s the range, standard deviation, and Hurst coefficient of the models used in the Santer paper

Note the models either have Hurst coefficients of about 0.5 (no long-term structure), while the data is quite a bit larger. Some of the models even have a Hurst coefficient below 0.5 (the mid-line on the graph), which I have never seen in a surface temperature dataset. A veritable Babel, indeed.

w.

Wansbeck

Re #150, Thanks.

I have just found some information on the BBC website. It says “The BBC Climate Change Experiment is a climate model, designed and produced by climateprediction.net researchers based at Oxford University” not Cambridge.

Frank K.

Re: 150
“There’s an educational version of the GISS Model II at http://edgcm.columbia.edu/
You can play with some of the parameters, but not many of them.”

I went to this site and, while you do have access the FORTRAN source code, alas, there again is NO formal documentation for this code outside of its use by third parties – that is, no equations, no mathematics, no numerical analysis, no theory – nothing. I detect a pattern here…

Well, there was this…

Click to access Hansen1983.pdf

The terse description of the mathematical formulation contained therein does NOT substitute for code documentation since it is unlikely that all of the models provided by the source code are completely covered in the paper.

And of course, the purpose of the edgcm software is to give users that great GCM experience!!

“The software allows users to experience the full scientific process including: designing experiments, setting up and running computer simulations, post-processing output, using scientific visualization to display results, and creating scientific manuscripts ready for publishing to the web.”

Software for cranking out scientific manuscripts – what a concept!

Gerald Browning

Frank K. (#153),

The NCAR documentation on the continuum equations and numerical approximations for their CAM3 model is fairly complete and well done.

However, the mathematical properties of the system are not discussed and for that info I refer you the the Exponential Growth threads.

Jerry

Bill F

If the models are not tuned to trends and are tested against out of sample data, where are the documented results of the tests? Where are the published papers reporting on “lessons learned” from failed model runs where valuable information about flawed process modules was gained by unrealistic model output? Surely somebody’s model failed badly enough that it was newsworthy to study “why” and report on it? Whats that? All the runs are successful? Nobody ever had to go back and adjust the modules to make the outcome realistic? Why do I have trouble believing that?

Did I just write an entire post composed of questions?

Frank K.

Re: #156

Yes I agree – I provided a link to NCAR’s CAM3 documentation in comment #38 in this thread…That is the kind of documentation I think is acceptable at a minimum.

However, researchers at NASA/GISS, for some reason, appear to believe that proper code documentation is not worth their time…and that is unacceptable!

erikG

re # 161

That sounds like a fallacy. I mean, they can do that, but they have not proved anything about the accuracy of the model. what is url?

Yorick

They just run thousands then pick the best.

Kind of like looking for temp proxy BCPs…

OT but Steve M mentioned looking for the Law Dome borehole data, here is some.

http://www.gfy.ku.dk/~www-glac/data/ddjtemp.txt

DocMartyn

The cut off for photosynthesis is about 90 ppm [CO2]
Whiteman and Koller, 1967…This file is very slow to load
http://www.blackwell-synergy.com/doi/pdf/10.1111/j.1469-8137.1967.tb06025.x

Low levels of CO2 at mid-day can get to about 190ppm, but that is rally low (no ref to hand, I can post one when I am at work).

lucia

>>You continue to assert that you have read our manuscripts, but continue to ask questions that seem to indicate that you have not read them very carefully.

No. I did not assert I had read all your manuscripts.

I said I read the comments in these three threads. I said I read “Silvie’s Manuscript” (once I was able to identify which manuscript you meant by this.) I told you Kress and Browning 1984 is money walled (see above). I don’t have it. I have not read it. (I read the abstract)

I said I have not read the book chapter you advised I read after I asked you for a reference that might list the full set of the equations you call “the hydrostatic system” and “the non-hydrostatic system”. I wanted a reference so I can be certain I know precisely what set of equations you mean when you use these terms. (I can guess, but wouldn’t it be better if you told me?

That chapter is also money walled. I plan to get it from the library, as there appears to be no other way to get you to specify the precise specific set of equations you use these terms to describe. (Why you can’t scan the page with the equations, and ask Steve to upload the page is beyond me. It would fall well within fair use of copyright. But, for now, until I get the chapter, I am somehow left to guess. )

As to any other “manuscripts”, you will need to be more specific about their titles authors dates etc. Otherwise, neither I, nor anyone reading this thread, is likely to read the manuscripts.

>>It is time that you start to answer some questions so I can see what you have learned from the manuscripts.
>>Before we continue, have you or have you not read the multiscale manuscript and if so please describe the system of equations cited in (2.1) so we can both agree on the system described in the earlier comment (inviscid, compressible NS with Coriolis and gravitational forces).

1) I have no clue which manuscript you have given the nickname “the multiscale manuscript”. I know this is a blog, and we don’t have formal citations rules, but, since I don’t know which paper that is, I can’t answer your question.

2) When you say “the earlier comment” which specific earlier comment do you mean?

>>What are the horizontal, vertical, and time scales and dependent variable scales.
>>Why is dissipation left off of these equations?
>> What does the S_1 scaling parameter mean. Where is the heating term?

Are you asking me where the heating term is in equation 2.1 or “the multiscale manuscript”? Beats me! Even if I have read the manuscript, I still wouldn’t know which one you call “the multiscale manuscript”. So, clearly, I can’t know which is equation 2.1.

If you mean to ask if there is dissipation in inviscid compressible NS with Coriolis and gravitational forces: No. Viscosity is neglected. Consequently this set of approximate equation has no viscous dissipation. ( That said, I’m not sure I would ever use the term “inviscid NS”. Why not say Euler equations? )

Did I answer the question you meant to ask?

If not, ask your question more clearly, and I’ll try to answer it. If showing the equation is necessary, then get a scanner, scan the equations, and upload the pdf. Doing that wouldn’t violate copyright, and it’s easy. I could answer more quickly.

Now, that I tried to answer the question I think you meant to ask, will you answer the ones I asked?

Are you saying that that one can or one cannot get smooth bounded numerical solutions to:
a) The hydrostatic system (full NS) when molecular viscosity is included?
b) The non-hydrostatic system when molecular viscosity is included?

These are simple questions I am asking because I am you are claiming in your things you say here in comments. The answers are either “yes one can”, “no one can’t”, or “I don’t know.”

I already asked these questions, and I know you can answer them without saying “Please read my manuscripts”. These questions can be answered without regard to your manuscripts because you are not the only person in the world who has done DNS!

Demesure

#162: Wansbeck: ALL the results of their climate experiments are at climateprediction.net
Go have a look before they sweep the inconvenient results under the carpet.

Besides the how to reconcile different models, there is the question of how to reconcile different results of a SAME model by other things than the pseudo-scientific probability density functions.

Tom Vonk

Steve has mentioned Dr. Demetris Koutsoyiannis work in the past, and I thought I would provide a link to his website. In particular, there is a paper there, titled Climate change, Hurst phenomenon, and hydrologic statistics, from which I would like to quote:

Thanks Bender , I got it and read it .
Excellent !
I didn’t know this work but it corresponds exactly to what I have been saying for some 10 years – the climatic system is self similar at all time scales .
Follows that you can’t use standard statistics hence I mentioned in the post above the second fundamental missed point – the stochastics of the climate system .
And that means that it suffices to take a meteorological time series (f.ex for the humidity) with a scale unit of 1 hour ,
unzoom to a scale unit of 10 years and obtain the same kind of variability .
Self similarity implies of course much more than only different statistic and needs completely different tools than what the modellers use today .
Of course as the modellers are deterministic and standard statisticians , they are 2 times wrong .
As the Hurst variability is “wilder” than the standard variability , using it would make several hundreds (thousands ?) modellers jobless 🙂
You wouldn’t want that , would you ?

On a particular note I seem to remember this individual S.Bloom from the R.Pielke website .
I have never met on the net something more slimy and obnoxious than this thing (robot ? lunatic computer programm ?)
As he has no clue about science and posesses a skull that has been so perfectly brainwashed that the vacuum within contains only a kind of pulsion to bite and snarl , his “posts” have always been unpleasant , insulting and irrelevant .
Your quote shows well the Bloom thing at its finest .

Jan Pompe

Tom

didn’t know this work but it corresponds exactly to what I have been saying for some 10 years – the climatic system is self similar at all time scales .

Is this, you think, why we see things like the same thirty degree lag in diurnal warming (hottest part of day is 2hrs after noon) and a thirty degree seasonal lag (the hottest day usually a month after the longest day) and so on, or is that something else?

Michael Jankowski

A somewhat out-dated but still humorous application of GCMs – the 2000 US Nat’l Assessment http://www.usgcrp.gov/usgcrp/Library/nationalassessment/overviewlooking.htm.

Both the Hadley and Canadian models of the day project warming, mostly in the midwest/west, but to varying degrees. They do differ on the direction of precipitation for much of the US, however. And when combined to determine soil moisture temp, forget it…they have different directions for about half of the nation.

Gavin will tell you that GCMs aren’t supposed to work on a resolution below global/continental scale, but obviously policymakers do use them as such.

steve mosher

RE 172. Guys I suggest you read the whole experimental approach first to the BBC thing.
It might focus the discussion. No judgements. Actually I was a participant in the experiment
early on ( researching distributed stuff) but they had some issues and had to reset.

Anyway, http://www.climateprediction.net/science/scientific_papers.php

steve mosher

RE 172. Thanks JohnV. Back inthe day we used the same approach in our modelling. We would
use a low resolution model ( effects based) to map out the boundaries of the parameter space
and to judge the sensitivity of given parameters so that when we ran the full blown physics
models we could devote more runs to the most sensitive paramaters. So there was a hierachy of models.

One thing I would like to hear bender Opine on is the stats behind this hierachical approach.

I’m sure it depends on the exact system of equations. Also, Much to our suprise we would often find
that the highest level models had a mind of their own.

Jan Pompe

Pat Keating says:
December 4th, 2007 at 10:03 am

Interesting. The CO2 lag during de-glaciation is roughly 8% (30 degrees in your terminology) of the rise-time of the temperature.
What does it mean physically, though?

Yes I tend to think of these matters in terms of control systems and while the tradesmen I used to work with understand circles and degrees their eyes glaze over when starting to talk in radians and to them ‘lag’ is the insulation on hot water/gas pipes. Old habits die hard.

I really haven’t a clue but I did notice the similarity relative to each of the cycles but in terms of actual time constant they are wildly different from a few hours to millennia. This of course points different systems with different time constants and it will be interesting to try and sort out the interplay. The only thing I can’t think of presently is that it’s a coincidence and I’m reading too much into it, but on the other hand it’s tempting to think there is a connection. I’m wondering what lag there is in the response to the sunspot cycles something for another time it’s 4 am and I’m off to bed.

SteveSadlov

Notice also the obvious cooling bias, expressed during the control phase. That is not at all surprising – we are after all in an interglacial, and not in an overall time frame such as 100 M years ago (Greenhouse World). We are in Ice World. Experiencing a temporary thaw.

Anthony Watts

RE132 Bender,

You may find this tidbit interesting. Here is a paper: Falsifcation Of The Atmospheric CO2 Greenhouse Effects Within The Frame Of Physics
Gerlich and Tscheuschner, July 24, 2007 available here: http://arxiv.org/PS_cache/arxiv/pdf/0707/0707.1161v2.pdf

In this paper they make a very interesting observation about GCM’s:

Computer models of higher dimensional chaotic systems, best described by non-linear partial differential equations (i.e. Navier-Stokes equations), fundamental differ from calculations where perturbation theory is applicable and successive improvements of the predictions – by raising the computing power – are possible. At best, these computer models may be regarded as a heuristic game.

and also this:

The temperature rises in the climate model computations are made plausible by a perpetuum mobile of the second kind. This is possible by setting the thermal conductivity in the atmospheric models to zero, an unphysical assumption. It would be no longer a perpetuum mobile of the second kind, if the “average” fictitious radiation balance, which has no physical justifcation anyway, was given up.

but this one made me laugh:

The choice of an appropriate discretization method and the definition of appropriate dynamical constraints (flux control) having become a part of computer modelling is nothing but another form of data curve fitting. The mathematical physicist v. Neumann once said to his young collaborators: “If you allow me four free parameters I can build a mathematical model that describes exactly everything that an elephant can do. If you allow me a fifth free parameter, the model I build will forecast that the elephant will fly.” (cf. Ref. [185].)

bender

John V, do you care to address these specific points raised in the paper? Failing that, can you link us to a rebuttal? Or are you just going to wave your hands?

bender

I caution all (myself included) to stay on topic, or else take it to the rubber room.

Tom Vonk

Jan

Is this, you think, why we see things like the same thirty degree lag in diurnal warming (hottest part of day is 2hrs after noon) and a thirty degree seasonal lag (the hottest day usually a month after the longest day) and so on, or is that something else?

It is similar but not exactly as quantitative .
What I mean is that if I show you a chart with a climatic parameter X drawn on a scale of 1 hour and then the runing average of the same parameter over a certain period T but with a scale of 10 years (or 50 or whatever) , you would be unable to make the difference and specifically if the time scale is not on the chart , you would be unable to distinguish which one is the short term and which one is the long term .
Of course the values of X and the running average of X (aka scale of the vertical axis) are not necessarily the same on both charts .
That’s also called scale invariance because self similarity leads to scale invariance and fractional dimensions .
In other words if you have problems to correctly quantify/estimate the short term (weather) , any amount of averaging won’t help you and you will find the same problems possibly with different/additional causes on the long term (climate) too .
That’s exactly the feature displayed f.ex by the Mandelbrot set – only an example because this set doesn’t correspond to a physical system but it gives the idea .

At least as long as the climate is defined as a temporal average of the short term dynamics what is the widely accepted definition up to now .

John V.

#188 Anthony Watts:
The flying elephant quote is funny because it’s true.

#187 bender:
I’m not going to bother getting into an argument on the Gerlich paper. Even if I did, it would get snipped for being off-topic.

bender

#191 Self-similarity is one of Spence_UK’s favorite subjects. (I say that only because he’s not here, but may wish to visit at some point.)

Cliff Huston

#190 Anthony Watts,

You may be on to something here. Only the British would design carburetors that could only be tuned by a Zen master with a good ear and soda straw. Since the ICCP is largely a British invention, I am wondering how the Zen master needs to be applied in this case. I’m not sure a soda straw would work with the climate models. Do you think an A.M. radio would help? (ref. Daisy et al)

Cliff

SteveSadlov

RE: #196 – slightly reworded:

Problem Statement: chaotic weather is continuously folded into the climate trajectory, yielding a long-term product whose time-series realizations over short time-scales appear to exhibit well-defined statisical probability distributions, but whose ensembles do not. Climate is non-ergodic, and the climate modelers are blind to the problem because they are divorced from mainstream statistics.

SteveSadlov

Junk parameterization is a key issue. This is how clouds are tap danced around, to name but one example.

RE: Kiehl – check out his full bibliography, it says a lot about his overall frame of reference and his attitude about GCMs. You mentioned pea and thimble, whether conscious or subconscious. And that, to me, is also a microcosm of how this whole GCM / climate prediction mess ever got so ugly. Final note about Kiehl’s misplaced faith in parameterization / GCMs (and again, this is a microcosm of “The community”) – it is that same faith in parameterization / GCMs as they are presently implemented that led him to a highly questionable cause – effect depiction of GHG concentration and the late Permian “die back.” These models tend to assume GHG as a lead rather than lag in nearly all cases. I’ve seen enough examples that seem to fly in the face of that, that I no longer take this as a given.

SteveSadlov

Unbelievable:

http://www.proquestk12.com/curr/snow/snow1095/snow1095.htm

I know this is, as Bloom would quip, “old,” but it reveals much. Here’s the money … quote:

Jeff Kiehl is a scientist who doesn’t believe in the scientific method. “If science proceeded the way the textbooks say it does, every scientific problem would have been solved long ago,” he quips. Instead, he points to creativity and intuition as critical elements in solving the complex puzzles the natural world poses.

SteveSadlov

Click to access radiative_kernel.pdf

SteveSadlov

http://www.ccsm.ucar.edu/models/

Anyone up for some serious auditing?

E&M

bender: #132 “One more comment to come – if I can dig it out – on internal vs external variability in Earth’s climate – an issue at the core of what they are tuning the GCMs to produce.”

I would very much like to see this comment. I have wondered for some time if climate variation must be
closely tied to driver variation or if the climate system has internals modes that can play out for years
after some external perturbation.

bender

William Connolley may be “movin on” from RC, but his legacy lives on. Check out this rebuttal to Roger Pielke Sr., in a thread on chaos and climate at RC:

Roger A. Pielke Sr. Says:
4 November 2005 at 11:11 AM

James and William- your post, unfortunately, perpetuates the use of climate to refer to long term weather statistics. You state that

“The chaotic nature of atmospheric solutions of the Navier-Stokes equations for fluid flow has great impact on weather forecasting (which we discuss first), but the evidence suggests that it has much less importance for climate prediction.” and that

“Fortunately, the calculation of climatic variables (i.e., long-term averages) is much easier than weather forecasting, since weather is ruled by the vagaries of stochastic fluctuations, while climate is not. ”

This is incorrect.

First, the more appropriate scientific definition of climate is that it is a system involving the oceans, land, atmosphere and continental ice sheets with interfacial fluxes between these components, as we concluded in the 2005 National Research Council report. Observations show chaotic behavior of the climate system on all time scales, including sudden regime transitions, as we documented in Rial, J., R.A. Pielke Sr., M. Beniston, M. Claussen, J. Canadell, P. Cox, H. Held, N. de Noblet-Ducoudre, R. Prinn, J. Reynolds, and J.D. Salas, 2004: Nonlinearities, feedbacks and critical thresholds within the Earth’s climate system. Climatic Change, 65, 11-38.

[Response: Roger – you appear to have your own, curious definition. You’re welcome to use your own definitions, for your own purposes; but if you redefine words in that way you can expect to have confusing conversations. We’re using the standard definition of climate. See, for example, the IPCC glossary – William]

That the model simulations that you discuss in your weblog do not simulate rapid climate transitions such as we document in our paper illustrates that the models do not skillfully create chaotic behavior over long time periods as clearly occurs in the real world.

[Response: This is wrong. The current climate does not display rapid transitions, which is why good GCMs don’t simulate them. Past rapid climate change appears to be associated with the Laurentide ice sheet, which is why we don’t see it now. See this post for more – William]

That climate is an integrated system and is sensitive to initial conditions is overviewed in Pielke, R.A., 1998: Long-term variability of climate. J. Atmos. Sci., 51, 155-159). We show in this study that even short-periodic natural variations of climate forcing can lead to significant long-term variability in the climate system.

We need to move the discussion to studying climate as a complex, nonlinear system which displays chaotic behavior if we are going to provide scientifically robust understanding to policymakers. Readers of your weblog are invited to read my postings at http://climatesci.atmos.colostate.edu if they would like to read a different perspective on climate science.

[Response: Everyone studies climate as a complex non-linear system. Few however expect chaotic behaviour from it – William]

A pity that James Annan wasn’t the one doing the inline responses. The article he wrote had a much more realistic conclusion:

The climate of a model can be easily defined in terms of the limit of the statistics of the model output as the integration time tends to infinity, under prescribed boundary conditions. This limit is well-defined for all climate models. However, the real world is slightly messier to deal with. The real climate system varies on all time scales, from daily weather, through annual, multi-year and decadal (ENSO), Milankovitch, glacial-interglacial cycles, plate tectonics and continental configurations, right up to the ultimate death of the Sun. The average temperature, and all other details of the climate system, will vary substantially depending on the time scale used. So how can we talk meaningfully about “the climate” and “climate change”? Well, although there are interesting scientific questions to ask across all the different time scales, the directly policy-relevant portion is on the multi-decadal and centennial time scale. It is quite clear that the pertubation that we are currently imposing is already large, and will be substantially larger, by up to an order of magnitude, than any plausible natural variability over this time scale. So for the policy-relevant issues, we generally focus on the physical atmosphere-ocean system, sometimes with coupled carbon-vegetation system, and treat the major ice sheets, orbital parameters and planetary topography as fixed boundary conditions. It’s an approximation, but a pretty good one.

Annan suggests that chaotic internal climate variability is not a problem for simulating responses to external forcings. In principle, this makes sense to me. In practice, I wonder how well those parameterization schemes will work when they are being tuned to non-persistent features of the cliamte system – a system which we all agreet “varies on all time scales”.

bender

#209 Alas, I cannot find my comment with the inline reply by “mike”.

But this will do. It is Gavin Schmidt on the thread Planetary Energy Imbalance, May 3 2005, discussing a graph of his simulation model output matched to the ocean’s heat content over time (10y):

Firstly, as surface temperatures and the ocean heat content are rising together, it almost certainly rules out intrinsic variability of the climate system as a major cause for the recent warming (since internal climate changes (ENSO, thermohaline variability, etc.) are related to transfers of heat around the system, atmospheric warming would only occur with energy from somewhere else (i.e. the ocean) which would then need to be cooling).

Secondly, since the ocean warming is shown to be consistent with the land surface changes, this helps validate the surface temperature record, which is then unlikely to be purely an artifact of urban biases etc.

Thirdly, since the current unrealised warming “in the pipeline” is related to the net imbalance, 0.85+/-0.15 W/m2 implies an further warming of around 0.5-0.7 C, regardless of future emission increases.

This exposition is much clearer than the one “mike” provided.

Evan Jones

“seems to admit that the anthropogenic forcings in the 20th century used to drive the IPCC simulations were chosen to fit the observed temperature trend.”

Woof! That sounds about as bad as my historical BvG storyboard model for the Civil War. (Of course I was openly unashamed about it–with no pretentions.)

How well do I know (by analogy) what those so-and-sos were up to. The accuracy of these-here “models” seems about as “realistic” as your average wargame. (That’s a base insult, BTW.) Well, this crass game designer can only doff his hat to his fellow “modelers”. (We rough out a plasticine Pygmalion and then have the face to claim we’ve created mankind.)

I’d be honored to loan them my shoehorn!

Tom Vonk

#202

BOULDER—Scientists at the National Center for Atmospheric Research (NCAR) have created a computer simulation showing Earth’s climate in unprecedented detail at the time of the greatest mass extinction in the planet’s history. The work gives support to a theory that an abrupt and dramatic rise in atmospheric levels of carbon dioxide triggered the massive die-off 251 million years ago. The research appears in the September issue of Geology.

“The results demonstrate how rapidly rising temperatures in the atmosphere can affect ocean circulation, cutting off oxygen to lower depths and extinguishing most life,” says NCAR scientist Jeffrey Kiehl, the lead author.

This one made me explode laughing !
It is known that the Earth’s orbit , as well as any 3 (or more) body orbit is chaotic (see f.ex Malhotra,Holman, Ito) .
That of course doesn’t mean that the orbits are random or that they can do anything crazy in short times .
However it means that the orbits are never the same and that the distance between 2 orbits with as close initial conditions as one wishes will increase exponentially with time .
This exponential parameter is called Lyapunov coefficient and is for the Earth orbit between 5 – 10 millions years .
That means that any calculation , be it forward or backward in time for more than let’s say 100 millions year gives an orbit whose parameters have no significance and could be completely random for all that matters .
Therefore considering that one can’t know the orbit parameters with any reasonable accuracy for 200 millions years ago , one wonders if that “paper” is more relevant than throwing a die . Or could it be something so mundane as choosing the orbital parameters in a manner that the to be proven thesis holds ?

# 211

That climate is an integrated system and is sensitive to initial conditions is overviewed in Pielke, R.A., 1998: Long-term variability of climate. J. Atmos. Sci., 51, 155-159). We show in this study that even short-periodic natural variations of climate forcing can lead to significant long-term variability in the climate system.

We need to move the discussion to studying climate as a complex, nonlinear system which displays chaotic behavior if we are going to provide scientifically robust understanding to policymakers. Readers of your weblog are invited to read my postings at http://climatesci.atmos.colostate.edu if they would like to read a different perspective on climate science.

[Response: Everyone studies climate as a complex non-linear system. Few however expect chaotic behaviour from it – William]

To make this understandable , I will rephrase what Connolley wrote in order to see what he thinks .

[Response : Everyone studies climate as a complex non-linear system because it is unfortunately so . However every time we think that we can get away with it , we’ll linearize as much as possible . Never mind the validity . We are scared s…less that the system might be chaotic at all time scales . Unfortunately none of us has a clue about those fancy theories of the young people . They really talk crazy things – fractional dimensions , exponential divergences , bifurcations , state transitions – can you imagine ? Even without understanding we believe that all that talk means that computer models are not worth the silicium they are running on and that is a clear no no . So everybody who opens his mouth implying that there could be chaotic behaviour we make him shut up . As for us , we won’t certainly shoot ourselves in the foot and we certainly don’t expect any chaotic behaviour . It goes without saying that we don’t need any proof as our priority is to save mankind and it is in your interest to not expect anything chaotic too . And now go away please .]

“The chaotic nature of atmospheric solutions of the Navier-Stokes equations for fluid flow has great impact on weather forecasting (which we discuss first), but the evidence suggests that it has much less importance for climate prediction.” and that

“Fortunately, the calculation of climatic variables (i.e., long-term averages) is much easier than weather forecasting, since weather is ruled by the vagaries of stochastic fluctuations, while climate is not. ”

And here we are again in the heart of the matter . Stochasticity , ergodicity – the single most important questions that have to be asked to make any progress .
On a particular note Connolley and Anan show one more time very explicitely that they understand nothing about deterministic chaos .
They seem to assimilate it both to a kind of randomness and to fluid dynamics (Navier Stokes) .
When they say that ” … but the evidence suggests that it has much less importance for climate prediction .” then that MIGHT be true for N-S and specifically small scale turbulence and I say might because even that would need a rather sophisticated mathematical proof .
But they didn’t understand that it is not even the point .
If the climate is chaotic , then it is because of the multiple strong non linear interactions between the subsystems each with a different characteristical time scale (f.ex annual ice cover variations , ENSO , NAO , cloudiness , use of land etc etc) .
That has nothing or extremely little to do with turbulence and N-S .
Chaos theory applies to any dynamical non linear complex system – fluid dynamics and N-S within fluid dynamics is only a tiny part of it .
It seems they don’t even understand the difference between weather and climate or that they have a very restrictive definition of “climate” that would be only the time average of atmospheric parameters what would of course be dramatically inadequate to formulate anything relevant about the whole system (= Lithosphere&Biosphere&Hydrosphere&Biosphere&Atmosphere) .

Max P.

Alright, here is a really dubious idea from a lay person…

In the social science of 30 years ago we used to have fun with step-wise regression as a “initial” analysis. Of course, that requires a data set of the suspected independent variables that might contribute to the dependent variable (in this case temperature). So has anyone taken only the measured data (greenhouse gases, rainfall, etc.) and run a simple multiple regression analysis?

Naturally, given the number of unmeasured unknowns, I would expect it to find little correlation to greenhouse gases (or any measured variable). Still, it would be interesting to see the results, as well as the verification scores.

Us dummies in S.Science would have simply “moved on” after finding our hypothesis false. We didn’t know how to construct ‘naturalistic’ models of society with invented data for the unknown parts of the nations social environment. If we had, we could have fit federal election results to the historical record and “predicted” elections for the next 100 years…

Sigh. If only I had been born later…

bender

Digressions not so bad if you can label them as such, stop it there, and bank it for discussion elsewhere. (exponential growth in informational systems)

bender

#223 Sadlov
If it is the objective, independently verifiable truth one seeks, nothing is more powerful, or painstakingly slow, than the scientific method. Those who choose the fast track … the consequences are yours to live with.

Chas

Willis, Re 119 & 149 [Very OT, so apologies to everyone]: Those are very nice, clear, 3d graphics : Are they yours -if so, what did you do them with ?

Tom Vonk

This is an interesting point. Has anyone ever mentioned this before to G. Schmidt? I wonder what the Hurst coefficients are for the GCM runs that they weed out for lack of convergence.

Some little bird is whispering to me that chances are that they are significantly above 0.5 .
Not that they’d bother to calculate them .
But of course we can’t know unless Gavin publishes a peer reviewed paper explaining that the variability estimators had to be revised because the modellers and IPCC have been using wrong statistics for 25 years .
Somehow I think that once the paper comes , it will not come from that general area .
Now why should I think that … 🙂

steve mosher

I mentioned Hurst components on RC once.

I heard this:

[audio src="http://new.wavlist.com/soundfx/014/cricket-1.wav" /]

bender

As a fundamentally thermohaline process, what is the chance that ocean currents express BZ wave-like behavior as the ocean strives for thermal equilibrium? Could this be a source of long-term persistence? Would additional forcing by wind, solar, and other quasiperiodic agents not make the resulting LTP chaotic? How would a GCM parameterization process cope with these complexities? Is this why the GCMs fail to exhibit Hurst-like all-scale variability?

Wunsch dynamics :: Koutsiyiannis statistics?

[Steve M: snip if you wish. (Ocean thermodynamics)]

bender

heh heh 🙂
but wikipedia’s an authority!

steve mosher

RE 225. Best way to cure bad engineeritis is to make them eat their own damn dogs breakfast for a year.
You built that? use it.

lucia

@steve mosher in 236 (and 237)
You don’t even need to force engineers to eat their own damn dogs breakfast. Serve it to the dog. If he won’t eat it then, “Okay, Houston, we’ve had a problem here.”

Pat Keating

How do the GCMs handle the integral over time and space, of many similar situations?

Gunnar, I think that’s your cue.

Chas

Thanks Willis, I wont tell a soul, I promise. 🙂

Willis Eschenbach

PS – Seeing the lightning outside my window reminds me that there is a large body of observational evidence that the processes that the modelers are controlling with various applications of the TRH (e.g. cloud formation, precipitation) are greatly influenced by the electromagnetic properties of the clouds. A storm cell is not just a giant solar driven engine for removing heat from the surface and adding cold to the surface. It is also a huge van de Graff generator, producing unbelievable voltages and amperages. These monster electric fields modify all of the parts of the condensation – evaporation – coalition – precipitation cycle in poorly understood ways.

There is also evidence that the electromagnetic properties of the clouds are influenced by the electromagnetic properties of the ionosphere … and that the electromagnetic properties of the ionosphere are driven in part by a combination of the solar heliomagnetic field, the coronal solar wind, the solar flare proton storms, and the coronal mass ejections (CMEs, also called “solar storms”) of the sun.

I can tell you how the modelers deal with that obviously critical group of phenomena … they don’t include them. Simple and clean, wouldn’t you say?

w.

Jaye

RE: Hurst coefficients

In the late 90’s I was involved in some ATR algorithm development for targeting FLIRS. One technique we tried was computing hurst coeff’s, using a convolution like kernel approach, on each image frame. The result was an image of hurst values. Worked fairly well in high clutter but was extremely slow. Anyway, we found that the ability to accurately measure hurst values was limited to about the middle half, give or take, of the available scales in the image. Low res scales had too little data and high res stuff suffered from aliasing, etc. We tested on 1 and 2d fbm’s generated with know hurst coeff’s.

bender

This bump is for Eric McFarland who wants to illustrate his deep knowledge of the hard physics underlying the GCMs. I will ask him some questions shortly. Study up, Eric.

bender

1. If GCMS are ruled by hard physics, why the different parameterizations above? (Why many dots on a line? Why not a cluster?)
2. Why THESE parameterizations, and not some other intermediate ones? (Why are the dots located where they are?)
3. Why this RANGE of parameterization and not something wider? (Why not more dots on either end of the curve?)

[Hint: These are fudgings, err, tunings. They are the free parameters whose governance by hard physics is uncertain.]

The floor is yours, Eric. Maybe you have a nice YouTube dodge for us?

conard

Bender,

Give him a little time. He is currently debugging the last few lines of LaPlace’s demon 😉