Guest Post by Willis Eschenbach
(Image above shows the Cray Ecoplex NOAA GAEA supercomputer, which can generate improbable future climate scenarios far faster than we simple humans …)
Someone on the web was touting the abilities of the early climate models by referencing the study Evaluating the Performance of Past Climate Model Projections by Zeke Hausfather, Henri F. Drake, Tristan Abbott, and Gavin A. Schmidt.
And yes, per their study, the climate models are just wonderful in their ability to hindcast the past and forecast the future. They put up a variety of tables and graphs to show that.
But I was interested in something different. I wanted to see what the transient climate response (TCR) of each of the models was. The Intergovernmental Panel on Climate Change (IPCC) defines the transient climate response (TCR) as the global temperature change at the time of a doubling of carbon dioxide (CO2) in a 1% per year increase experiment.
In other words, it’s the change in temperature corresponding to a 1 W/m2 increase in forcing, times the 3.7 W/m2 increase in forcing that the IPCC says will occur from a doubling of CO2.
So I looked at the model data, which the authors very responsibly posted online as an Excel spreadsheet at GitHub, and calculated the TCR for each model that they analyzed.
Then in addition to the TCR I wanted to calculate the ECS, the Equilibrium Climate Sensitivity. To do that I looked at the TCR and the ECS of 23 models from this paper. Here’s that data, along with a LOWESS smooth of the data.
Figure 1. Scatterplot, equilibrium climate sensitivity (ECS) versus transient climate response (TCR), 23 CMIP models.
So I converted the TCR values from the model performance study into ECS values, using the LOWESS line to do the conversion. Details in the endnotes.
Here are those results for the various models.
Figure 2. TCR and ECS for the 12 models investigated in the model performance paper.
Now, the interesting part of this is that the climate sensitivity (ECS) for the models covers a very wide range, from two to four degrees for a doubling of CO2 … and yet, they all do a whiz-bang job of hindcasting the temperature history of the planet.
And IF they are all “physics based” as is always claimed, I’m sorry but that’s not possible.
I call this “Dr. Kiehl’s Paradox” because back in 2007, in a paper published in GRL entitled “Twentieth century climate model response and climate sensitivity” Jeffrey Kiehl first noticed this oddity. He said:
The question is: if climate models differ by a factor of 2 to 3 in their climate sensitivity, how can they all simulate the global temperature record with a reasonable degree of accuracy?
A good question indeed. In response to that paper, I wrote a post I entitled “Dr. Kiehl’s Paradox“. In that post I discussed Dr. Kiehl’s answer to that question, as well as my own answer to that question. I followed that up with a couple more analyses entitled “Zero Point Three Times The Forcing” and “Life Is Like A Black Box Of Chocolates“.
In those posts, I showed that despite their immense complexity, the global temperature output of the climate models can be emulated very exactly by a one-line equation that simply lags and scales the forcings used as the input to the models.
Oh, yeah, a final oddity. In the Hansen 1981 model, the TCR and thus the ECS are not constant. From 1981 to 2024 the ECS goes from about 1.6°C/2xCO2 to about 2.5°C/2xCO2. The ECS then continues to increase to 2040, decreases slightly to 2075, and by 2100 it’s all the way up to 2.8°C/2xCO2. That’s almost twice its starting value.
Say what?
Figure 3. TCR and ECS for the Hansen 1981 model. Each year’s TCS is calculated using the change from the 1950 forcing and temperature to the forcing and temperature for that year.
So I have to say that I’m not impressed by the apparent agreement of the models with reality. They are not physics-based. Instead, they are simply tuned to match the past … and in the case of Hansen 1981, tuned to exaggerate the future warming.
And as my bonafide genius older brother used to say, “It’s easy to predict the future … as long as it’s just like the past”.
Unfortunately, as I showed in my last post entitled Now You Sea Ice, Now You Don’t, the climate is occasionally and quite unpredictably very unlike the past …
Regards to all,
w.
Yeah, you’ve heard it before: When you comment, please quote the exact words that you are discussing. It avoids endless problems.
For the math folks: The function I used in Excel to convert the TCR into the ECS is based on an approximation of the LOWESS curve in Figure 1. It is:
= IF(TCR < 1.6, 1.39 + 0.8 * TCR, IF(TCR < 2, -1.04 + 2.32 * TCR, 1.4 + 1.1 * TCR))
This defines three approximately straight-line sections of the LOWESS smooth, and uses their formulas to calculate the ECS depending on the value of the TCR
Discover more from Watts Up With That?
Subscribe to get the latest posts sent to your email.
A good argument about Kiehl’s paradox.
But I must question the effort using the LOWESS smooth of the data. That spread of results (from n = 23) is all over the place. There does not seem to be any meaningful pattern to smooth, in the first place.
Which would, of course, be another outcome of Kiehl’s paradox.
Yes, with an OLS regression one is provided with a goodness of fit, the r^2 value, which tells one how much of the variance is explained by the fit. What does the LOWESS tell us about the goodness or fit or how good the prediction is?
The R^2 of the LOWESS model is 0.7. The R^2 of a standard linear regression of the data is 0.66.
Regards,
w.
Says it all right there. I don’t think physical functions are parameterized in the real world. Just my opinion though.
Reminds me of some combined Monte Carlo and least squares optimization calculations I used to spend a lot of time on. I used test cases to prove the validity of the calculations where I new the answer, like hind casting here. Small changes in the “constants” made huge differences in the output. But all was well for my validation calcs as I tuned to get the right answer. After a couple years, I found a new reference case..it all blew up. Never got a set of constants that could match the references after that. Might as well have used a ouija board.
Willis,
Your descriptions of the various sensitivity values leading to the same hindcasting “skill” made me immediately think of Neumann: with 3 variables I can create an elephant, with 4 I can wiggle its trunk. Don’t the climate sims have hundreds to thousands of variables?
I may be wrong (and am too lazy to look it up) but I think he said, “with 4 variables I can create an elephant, with 5 I can wiggle its trunk.” I think with 3 variables you only get an anteater.
The story of Freeman Dyson, Fermi, and von Neumann is here. Enjoy.
Another story about Planck:
Once, when he [Planck] forgot which room he was supposed to lecture in, he stopped by the department office and asked, “Please tell me in which room does Professor Planck lecture today?” He was told sternly, “Don’t go there, young fellow. You are much too young to understand the lectures of our learned Professor Planck.”
–page 147, The God Particle by Leon Lederman
Our digital computers are von Neumann machines. He promoted the idea of stored programs. That allows loops–a very powerful programming technique. Von Neumann also advised his friend, Shannon, to name his new information theory property: entropy. He said, “Nobody understands what entropy is and in an argument, you’ll have the advantage.” Unfortunately, people think information theory entropy is the same as thermodynamic entropy. They aren’t–for one thing the units don’t match.
As smart as Dyson was, his spheres are nonsense. He forgot that spherical shells involving inverse square law forces have zero force anywhere inside the sphere. The can’t orbit an object or star. That feature is what makes Van de Graaff generators possible.
Unfortunately, people think information theory entropy is the same as thermodynamic entropy. They aren’t–for one thing the units don’t match.
Information entropy is more fundamental. Gibbs arrived at essentially the same conclusion right before he died around 1902. Check out the work of Jaynes.
By written CMIP rules, model parameters are tuned to best hindcast 30 years. The CMIP spaghetti graph results always show decent hindcast agreement among the various models. Or, as WE said, ‘tuned to match the past’. BUT that is only because displayed as model anomalies to themselves.
In temperature ‘model’ reality, expressed in K or C, the models disagree between themselves by about +/-3C. Awful disagreement, over just 30 years. A root cause of Kiehl’s paradox.
Separate comment to figure 2. The original and revised Lewis and Curry EBM observational TCR and ECS range from ECS=1.23*TCR to 1.3*TCR (depends on which paper and which time period, but those are the upper and lower bounds—just went to Climate Etc, got both papers, and double checked). It is obvious from inspection that the 12 models depicted all over egg ECS compared to TCR.
Thanks, Rud. If I take a simple linear regression of the data shown in Figure 1, I get
ECS = 1.75 * TCS
That’s actual data from 23 CMIP5 models.
In addition, I find this graphic from the IPCC:
It shows a TCR of 2°C/2xCO2, and an ECS of 3.5°C/2xCO2. That’s also a ratio of 1.75.
Do you have a link to the Lewis and Curry paper?
Thanks,
w.
Easy to access both. Go over to Judith’s and type ‘Lewis and Curry’ into her search bar. You will need to scroll down the results a fair bit since she posted them in full (IIRC) in 2014 and 2016. Avoids paywall.
Thanks, Rud. The uncertainties in their paper “The implications for climate sensitivity of AR5 forcing and heat uptake estimates” have huge uncertainties. Interdecile range for the most recent (1930 – 2011) ECS calculations are 0.9 to 9.45°C, and the same range for TCR is 0.8 to 3.33°C. That gives a ECS/TCR range of 1.85 to 3.33.
On the other hand, the full range of ECS of the 23 models that I used is 2.1 to 4.7°C, and the TCR range is 1.1 to 2.6°C. And the range of the individual model ECR/TCS is 1.4 to 2.3.
Not sure what to make of all that …
Best to you and that good lady,
w.
And I would consider Lewis and Curry as the upper bound for warming effects, as it used a database (GISTEMP, I recall) that poorly accounts for UHI and poor siting in general.
‘The CMIP spaghetti graph results always show decent hindcast agreement among the various models.’
Hindcasting utilizes the same physical principle as a shotgun barrel. Once the load clears the muzzle, it spreads out.
“‘tuned to match the past’.”
But which “past”… the real past… or the past that has been deliberately adjusted to suit the agenda. ??
The hindcast is tuned to match the adjusted past, by definition.
Doesn’t matter, since the past is assumed all anthropogenic, when it clearly isn’t. Natural variation did not magically stop in 1975.
The oft heard assertion that models are ‘physics based’ simply isn’t true.
Moreover, it can NEVER be true because of the computational CFL constraints on numerical solutions to partial differential equations. (As a rule of thumb, UCAR says halving grid size—doubling resolution—requires 10x the computation.) The example used to illustrate this in old guest post ‘The Trouble with Climate Models’ was convection cells (thunderstorms). A physics based convection cell based on Navier-Stokes needs a resolution of 4km or less.
That is possible in a regionally bounded weather model going out 3-4 days, as illustrated in the old post.
That is NOT possible for a global climate model run out a hundred years. The finest resolution in CMIP6 was 100km, the average was about 180km.
All such important stuff has to be parameterized. Parameters aren’t ‘physics based.’ They are tuned guesstimates. Examples of guesstimates and the two basic ways climate modelers tune them are in the old post.
“That is NOT possible for a global climate model run out a hundred years.”
___________________________________________________
IPCC TAR; Chapter 14; Page 774 pdf 6; ¶ 14.2.2.2
In sum, a strategy must recognise what is possible. In climate
research and modelling, we should recognise that we are dealing
with a coupled non-linear chaotic system, and therefore that the
long-term prediction of future climate states is not possible.
Yes, they acknowledge that the energy exchanges in clouds have to be parameterized. Thus, even though everything else may be calculated ‘exactly,’ the results are corrupted by using the best guesses on how clouds behave and interact with everything else.
I have heard that none of the climate models are consistent with Conservation of Energy. Is this true? If so, the entire effort to model the Earth’s climate is pointless.
They are not models
They can not predict anything
Long term climate trends can’t be predicted.
They are climate confuser games used for propaganda. Accurate predictions are NOT a goal (Russian IMN may be an exception)
With RCP 8.5 and strong water vapor feedback you get a scary range of predictions in 400 years
With RCP 3.4 and a modest water vapor positive feedback, for the next 70 years, you can roughly cut the 400 year RCP 8.5 warming rate in half.
The 1970s models programmed that way made a 70 year prediction similar to the actual warming after 1975
So what?
So Zeke Hausfather wrote a study and article showing if you use the right assumptions the models can be made to appear accurate … and never mind the RCP 8.5 assumptions the IPCC publicizes. Climate Howlers treat the Zeke Hausfather article like their bible. Who names their kid Zeke?
Evaluating the Performance of Past Climate Model Projections – Hausfather – 2020 – Geophysical Research Letters – Wiley Online Library
Explaining the Discrepancies Between Hausfather et al. (2019) and Lewis&Curry (2018) | Climate Etc. (judithcurry.com)
Zeke also published a paper comporting to prove UHI isn’t real, even tho it most definitely is—easily proven by driving on a summer evening from a rural area into a city and back out to rural. The car outside ambient temperature gauge will change by 5 or 6F each direction. Explains no kids named Zeke.
Oh.. the temperature change is real…. it just doesn’t affect the urban thermometers.. 😉
Rud
I’ve done that experiment many times with the same result. Once, the difference from a very developed shopping mall area and the golf course 1 mile away was 5 F. A very real phenomenon.
“Who names their kid Zeke?” A Jew perhaps?
Stop with the ugly anti-Semitism! That’s a slimy personal attack that has no place here! This is a science site, not a place for you to display your nasty childish prejudices.
C’mon, Phil., you’re better than this.
w.
My first thought with models is always, Now do Venus. And by that I mean, put in 96% CO2 in the calculations and see what you get out the other side. If it’s more than 460 deg. C. then you don’t have the physics anywhere near right.
Venus is so hot because of the enormous atmospheric pressure (ca. 90 bar), not because of the Greenhouse Effect.
Maybe so, but if a model designed for Earth’s atmospheric pressure comes back with 1200 deg. C. at 96% CO2, it’s even more reason it’s not right.
“Oh, yeah, a final oddity. In the Hansen 1981 model, the TCR and thus the ECS are not constant.”
I’m pleasantly surprised to see that revealed from such a date. I’ve said here before that there is no a priori reason to think that it should be constant in such a complex system.
The numbers are, after all, just a human abstraction from the models: ‘This is what happens when I feed in this set of numbers into my model under these conditions I have chosen’.
Of course, if they admitted that they were full of it then the funding and attention might go away and the human race might be better able to focus on some real problems.
Hi Willis,
IPCC suggests natural cycles cancel out. Atlantic oscillation is around 70 years so can’t determine if cycles do cancel out without testing over a number of full cycles?
Using model output and a fourier analysis, do the models show the Atlantic and Pacific cycles?
If not, then I suspect significant manipulation to hind cast the past. Given both cycles were at or near their coolest points in the 70’s, before their getting-warmer phases, I suspect manipulation of outputs to get expected results.
Apologies, my statistical knowledge is poor so unlikely to get reliable results on my own!
If the 50-year-old models are so good, why are we building gazillion-quid super pooper scooper computers to produce new models? Do we even need climate scientists any more?
Last time I checked the numbers (a few years ago) US alone spent $2.4 billion on climate models. That same year it spent $85 million on weather models. So lots of jobs and careers at stake in climate modeling, even tho weather models (tornados, blizzards, hurricanes) are much more valuable.
Separate observation. The climate science is supposedly settled, so no we don’t.
But somehow climate scientists never spot the contradiction.
AGW is a work-making employment scheme that Western governments have glommed onto.
It’s really a dole system for people who were told that a university degree would guarantee them employment and a career.
Let’s not forget that the “temperature history” has been specifically mal-adjusted to help validate their fake models.
If you can change the past temperatures to match the models.. hindcasting is easy.
If you also have many model “parameters” you can wiggle… hindcasting becomes even more easy.
Let’s keep it simple.
Does the multi-model mean resemble observations?
Seems to me it does.
WRONG, the observations you refer to are NOT actual real observations.
They have been manically adjusted and “homogenised” to suit the agenda.
They are NOT REAL… they are fabricated.
…. so if the models “match” them, the models are WRONG !!
It can resemble anything you want it to.
Even down to hundredths of one degree C.
Apparently.
Real atmospheric temperature are not remotely a “match”.
To 2015 ? It’s 2024, you’re missing a lot of data points….
“Let’s keep it simple.”
“Simple” is about all you are capable of being.
A non physical (ie fitted) representation of a physical process has no predictive power.
No. One variable (GST) has been fitted to represent observations with varying degrees of success in the various models. The rest of the models’ variables are generally rubbish.
Seems to me you are wrong
The average 400 year RCP 8.5 prediction from the 1970s reflects a warming rate roughly double the actual UAH warming rate since then.
Those were the IPCC publicized assumptions
When people claim the models that were publicized were accurate, they are lying.
When using surface GAT, rather than UAH, the error is less than 2x but still large.
Ah yes, does the composite average of junk outputs look believable…
Typical climate alarmism at work.
Resemble?
That relation is far worse than the “correlation is not causation” relationship maxim.
Leaving the programmer’s axiom, ‘garbage in, make work garbage processing means absolute garbage out’ as proven correct.
“Resemble”, the international standard for climate activists; i.e., is it good enough to fool the ignorant…
Why does the multi-model mean have any significance? It seems to be made by averaging a bunch of models, the worse with the better, using different parameters and theories of what the climate drivers are.
Why are we not just throwing out the ones that don’t work and getting to one model, with some uncertainty bounds around its predictions? If this was public health or aircraft design or drug treatment evaluation that is what we would be doing. Why is climate different?
Why should we put any faith when designing public policies in the mean of a bunch of mainly failing models? What happened to falsification? Popper must be turning in his grave.
It gets even stranger if you add in the bunch of CMIP6 models that now have ECS greater than 5 degrees.
Also my understanding is that some, maybe most or all, of the models exhibit sensitivity to initial conditions. I do not see how such models can be said to have a specific CS, T or E.
And conversely given that climate is a far-from-equilibrium (that is chaotic) system the concept of ECS has no physical meaning.
Surely one answer to the paradox is that other factors, boundary conditions perhaps, can be set to properly tune the models to the past. For example both SO2 levels and ozone depletion can easily vary over time (as needed).
Look, we can use computers to perform physics-based numerical weather prediction! Great, now we can benefit from improved short term forecasts.
< HUGE LEAP, NEVER FOR A MINUTE JUSTIFIED TO DIAGNOSE THE CLIMATE SYSTEM RESPONSE TO INCREMENTAL CO2>
Look, we can use GCMs to match observations and project the warming scenarios from CO2 emissions!
The huge leap is obvious, and so is the circular reasoning beginning with the “forcing” + “feedback” framing of the core question itself – where should we expect the energy involved in the incremental static radiative effect of rising CO2 to end up? Will it be a.) accumulated on land and in the oceans, or b.) dissipated to space by the general circulation, just like all the rest of the absorbed energy?
Pick b.) Incremental CO2 adds no energy of its own to the atmosphere+land+ocean system, and the concept of the Lorenz energy cycle implies dynamic self-regulation.
https://youtu.be/hDurP-4gVrY (please read the full explanation in the description box.)
There was never any good reason for the IPCC, the UNFCCC, the 2009 Endangerment Finding, none of it.
Well said David.
except you missed a word or two…
“TINY incremental static radiative ABSORPTION effect “
This line bares repeating
“Incremental CO2 adds no energy of its own to the atmosphere+land+ocean system”
Willis,
I hope that you do not mind a rough and ready different interpretation of your Figure 1 scatterplot. I am not inferring that your graph was wrong – I’m just suggesting a further factor. That factor might be named “Compliance with emerging dictates”. (These ECS numbers were published before the IPCC was formed in 1988.)
My graph splits ECS estimates into those authors taking the low road to sensitivity and risking retribution from colleagues pushing global warming alarm and the high road of eventual IPCC policy preferences.
Cheers Geoff S
>> The question is: if climate models differ by a factor of 2 to 3 in their climate sensitivity, how can they all simulate the global temperature record with a reasonable degree of accuracy?
Aww, I do not at all understand this focus on older models by you and others.
CMIP6 models have a 25% higher climate sensitivity for CO2 after improving the cloud parametrization.
This mean older models did not get the clouds right and it matters.
A systematic error of 25% or more will put an end to any model in any field of science.
It is that simple.
But you have not answered his key question:
If you think that the climate alarm was raised on broken and incorrect science, and that it matters, yet… it was still accepted anyway… why do you think that the science is any better now?
The gatekeeping is still just as flawed.
😂 🤣 😂 😂 🤣 😂 🤣 😂 😂 🤣
The article and especially the references to recent alarmist model analysis that alarmists use to claim their models, today, produce accurate results.
Then look at the picture of the NOAA’s Cray super computer that computes and recomputes the current slate of models used in the analysis.
Where are these alleged but unnamed unspecific “old models” that are different than the unspecific “CMIP6 models”?
“Where are these alleged but unnamed unspecific “old models””
Willis names a number of models in Fig 2, dated from 1970 to 1988
There are over a hundred models “based on physics” and none of them agree with each other. Do we really need to dig any deeper into this conundrum?
Seems obvious that the “physics” should result in ONE model, eh? We either don’t know the physics well enough, or we are incapable of capturing it in any model. I suspect both of these possibilities are at play.
There is one thing the models can assuredly tell us! We don’t know enough.
Mirrors and prestidigitation bait & switch by alarmists whose incomes and glory are dependent upon the climate scam continuing.
In contemporary teaching feedbacks depend on the climate state, not CO2 per se. In warmer states the global climate system is said to be less stable. A linear increase in sensitivity with warming. Same with cooling. In cooler climate states the climate system is also less stable. This latter (cooler; high sensitivity) case is often ignored.
The optimal climate state from a thermodynamic perspective is one with maximum stability (lowest sensitivity). All else being equal, the system will strive towards this optimum.
When ignoring the higher sensitivity in cooler states while using proxy data to infer sensitivity, such as temperature change since last glacial maximum, e.g.. in Sherwood and more recently in Hansen, they can only result in an ECS value that is too high.
This additional complicating factor, i.e. the non-constancy and non-linearity of climate sensitivity with temperature, is often misunderstood.
This reminds me of the good old days when there were two groups of fission track geochronologists, who had different values for standards and different values for the decay constant for spontaneous fission of U-238. Despite this, they would get the same ages on samples. When you have multiple “Cook’s Variable Constants”, there’s nothing you can’t do.