I’m going to have to write a year end temperature post. Meanwhile… open thread!
Oh… if anyone has a tape of Mann-CEI court cases, I would like a copy. Is it available?
371 thoughts on “New Open Thread!”
Comments are closed.
I’m going to have to write a year end temperature post. Meanwhile… open thread!
Oh… if anyone has a tape of Mann-CEI court cases, I would like a copy. Is it available?
Comments are closed.
Here is a physical example.
These data are from a cylindrical 1″ thick steel tank filled with air. The tank is 42″ in diameter and is 84″ long, including hemispheric end caps.
The temperature is monitored on the exterior of the tank, and in the center of the tank. The tank is aligned with its major axis horizontal, and a positive temperature gradient is maintained vertically across the tank (this temperature gradient makes the air in the tank stable against convection).
Temporal record
There is a heat source in the interior of the tank (waste heat from sensors inside of the tank), which result in the interior tank temperature being slightly larger than the exterior temperature. But you can see clearly there is a time delay between external temperature fluctuations and the measured response in the interior of the tank.
This delay in response shows up clearly in: the lagged correlation plot , and corresponds to a delay of about 8 hours.
We can also look at the thermal relaxation time, where the interior of the tank is deliberately heated, the heat source is turned off, and then we track the temperature change inside of the tank as it returns to its previous thermal equilibrium:
Thermal relaxation curve.
As you make the object larger, the amount of time it takes to reach equilibrium depends quadratically on the dimensions of the object. That is, scale up the dimensions by a factor of 2 and time needed for equilibration will increase by a factor of 4.
Some complications: This relationship assumes you are warming/cooling the body uniformly. Objects that are not driven uniformly can respond much more quickly to a change in forcing. If the air circulates the rate of heat energy transfer is much faster.
Both non-uniform heating and fluid motion are present in the Earth. As a result the thermal relaxation time is the relatively short ~2000 year period I mentioned. Were the fluid stable against convection and the Earth uniformly heated and cooled, the relaxation time would be much, much longer (millions of years?).
Here is something that might help keep the discussion on heat and cooling on track.
https://msd23.haikulearning.com/ndonohue/earthscience2014-2015/cms_page/view/5850468
SteveF, I think the problem is laypeople don’t always realize that words are no substitute for equations. The theories and physical frameworks of science are arbitrated through the language of math. The words provide a means for us to describe the equations and justify our assumptions, but the ultimate arbitrator of true in the physical science is what the math tells us is valid, not our handwaving verbal explanations of them.
If you’re to arrive at a true understanding in the physical science, that necessarily involves grappling with the physical laws (expressed mathematically) and grappling with their implications and what they tell you, rather than verbal descriptions of how they behave.
Paul_K has complained that the output from TCR simulations shows an accelerating rate of warming, making assessment of TCR problematic. Why does this happen?
http://rankexploits.com/musings/2012/the-arbitrariness-of-the-ipcc-feedback-calculations/
While manipulating the equations used to abstract ECS and TCR from energy balance models, I arrived at an interesting relationship:
1/TCR = 1/ECS + (1/F_2XCO2) * (dQ/dT)
where dQ is the heat flux into the ocean and dT is the surface temperature change. The quantity dQ/dT is known as the ocean heat uptake efficiency (kappa). If dQ/dT changes with time (and ECS is a constant), then TCR changes with time.
As the mixed layer warms and expands faster than the bulk ocean below, one might expect vertical advection in the ocean to slow. Surface warming increases CAPE and (by inverse analogy) should increase the stable stratification of the ocean. Assuming AOGCMs take the density gradient into account when calculating heat transport into the deeper ocean, one might anticipate that their dQ/dT will decrease and TCR will increase with time.
Figures 4a and 4b of the following paper shows exactly this kind of behavior:
http://onlinelibrary.wiley.com/enhanced/doi/10.1002/grl.50541
In Figure 4a, dN is dQ and the slope is the ocean heat uptake efficiency.
I saw that David Appell claimed (on judithcurry) that this rather warm year is pulling the measured temperatures back into the p>5% range of the GCMs again. Comments?
MikeR
Maybe so. I’ll have to crunch numbers. But this would be pretty faint endorsement — as dancing in and out of ‘rejection’ is… well….
Also, when this is done by temperature anomalies rather than trend, it’s not an entirely fair test. The anomalies automatically force agreement of the mean during the baseline periods. So that part of any comparison is not any sort of test– and periods close in time to the baseline aren’t much of a test either.
MikeR, the 95% CL itself has uncertainty so I wouldn’t put any special meaning over magically crossing the 5% threshold. Anyway, as DeWitt discussed above, there is risk in relying too strongly on trends with large excursions near one boundary (the excursion ends up getting over-weighted).
If we see a trend, where the margin continues closing after a few years, that would be meaningful to me.
Mark+Bofill (Comment #133400)
why do you say that the energy going out on the nightside equals the energy coming in on the sunny side?
This might be the case, but there is nothing inevitable about it and absent some specific reason to think so I rather doubt it.
OK , take 1 constant * energy source, the sun.
Take 1 irradiated object at a constant distance, a blackbody *, the earth.
Estimate the energy in,estimate the energy out.
It is inevitable * , according to the laws of physics that we use for blackbodies * that the energy in will equal the energy out.
Now spin the said blackbody and check the energy equation.
Energy in equals energy out.
Not my say so, the science.
Rotate the axis of spin, spin it backwards , Stop and start it.
The energy in equals the energy out.
I cannot complain about snide put downs, when arrogant [feeling right] about an issue] I have often included “put downs” including of commentators here.
A common one on this site is to say, I am a scientist, I understand science and maths and the other person does not.
Carrick (Comment #133408) SteveF, I think the problem is laypeople don’t always realize that words are no substitute for equations.
SteveF (Comment #133405) DeWitt, Good explanation, but I bet he will not get it. If someone does not grasp the behavior of the simplest systems then they can’t possibly grapple with a complex system with a wide range of heat capacities which are only evident when a range of frequencies are applied.
I am educated, I have a bachelor of science, I have 60 years of trying to understand Mathematics and constantly improve myself and am quite happy to be educated further.
Please explain, Why you disagree with this answer or I may suspect you of being “someone does not grasp the behavior of the simplest systems”
“Warming or cooling of any body takes time, and is never instantaneous”, fantastic, but we are actually discussing continual steady state warming.The heat coming in takes time to work and the heat going out takes time to dissipate but in the example I give the laws of physics decree that both events take place at the same rate.
So stop your hand waving.
Carrick (Comment #133406) A physical example. What you are arguing is that if the material, the air, changes in texture, ie more CO2, that makes that material heat up a a little more before it releases the outgoing energy. SteveF is almost right [so close] on part of the process taking time and not being instantaneous but only when a basic condition is being altered,so it is not in the steady state.
That extra heat has to share with the surrounding materials , here we are talking about the oceans [though no one mentions the heat capacitance of all the water on land and in the air].
As per the distribution with your tank model I love your comment “Were the fluid stable against convection and the Earth uniformly heated and cooled, the relaxation time would be much, much longer (millions of years?).”
The earth is a basting spit but the degree of change in temperature we are talking about being distributed is so minute that your estimate might not be so bad.
After all heat conduction in the ocean is so slow!
Thank you both and Mark for trying to answer this puzzle.
The energy in = energy out, if true, would be a rather special case since it seems highly unlikely that Venus follows that.
Temp = temp out fails also on considering how the Earth is not a perfect black body: oceans, axial tilt, weather, non-symmetric landmasses, land surfaces of differing albedo, composition etc:
All act to change how radiant energy is received, re-radiated and stored in different ways.
angech,
You may be educated, but that education was plainly ineffective; you don’t have a clue what you are talking about. I will not waste any more time on you. I sincerely suggest that you concentrate on WUWT, where your crazy ideas will be warmly received. Adeus.
Angech,
No problem. Riff raff’s like us gotta stick together homey. 🙂
But I still think you’ve got a wrong idea about blackbodies. I read this:
There are lots of other interesting things I’ve read about blackbodies, but energy in == energy out isn’t one of them. If you insist that energy in == energy out in all cases for .. anything, really, any object in existence, then you’re effectively saying that that thing can never change temperature.
But itz all good.
angech,
Fewer words, more equations. Ein = Eout doesn’t count. Perhaps if you could do that, I might actually bother to read your posts.
You’re beginning to remind me of someone who posted frequently at Science of Doom. He wouldn’t write equations either and ended up being banned.
Mark+Bofill (Comment #133416) ” But I still think you’ve got a wrong idea about blackbodies. I read this:
A black body is an idealized physical body that absorbs all incident electromagnetic radiation, regardless of frequency or angle of incidence.”
True but it also emits black body radiation.
Black-body radiation is the type of electromagnetic radiation within or surrounding a body in thermodynamic equilibrium with its environment, or emitted by a black body (an opaque and non-reflective body) held at constant, uniform temperature.
The radiation has a specific spectrum and intensity that depends only on the temperature of the body.
The thermal radiation spontaneously emitted by many ordinary objects can be approximated as blackbody radiation. As its temperature increases further it eventually becomes blindingly brilliant blue-white.
Although planets and stars are neither in thermal equilibrium with their surroundings nor perfect black bodies, black-body radiation is used as a first approximation for the energy they emit.. Black-body radiation is also called complete radiation or temperature radiation or thermal radiation.
“For a body in radiative exchange equilibrium with its surroundings, the rate at which it emits radiant energy is equal to the rate at which it absorbs it:
In other words, given the assumptions made, the temperature of a planet depends only on the surface temperature of the Sun, the radius of the Sun, the distance between the planet and the Sun, the albedo and the IR emissivity of the planet.”
These are not my words . They are straight off good old Wikipedia
Black Body radiation
Mark Bofill is correct in his most recent comments:
Radiation equilibrium is expressed as radiative power in = radiative power out.
The “energy in = energy out” condition that angech keeps asserting is clearly just nonsense.
As Mark points out, if you really required “energy in = energy out” a body could never heat up or cool down. If the energy of an object is fixed, then so is its temperature.
[Throwing angech a bone, note the Wiki article talks about the rate at which energy is absorbed or emitted. ]
Thanks Carrick.
ok
I suspect the food fight over the instrumental surface temperature vs the satellite MSU based lower troposphere temperature is going to warm up again if 2014 is labeled as the warmest year in the instrumental record or as close as never mind. Neither RSS at 0.253°C for the first 11 months nor UAH at 0.270°C are going to be particularly close to the records set in 1998 of 0.550 and 0.420 or even to 2010 at 0.476 and 0.399.
For RSS, the OLS trend of the 1979-2014 calendar year annual average is 1.03E-03°C/yr with an R² = 0.48 while the slope of 2000-2014 is 1.20E-05°C/yr with an R² = 3.9E-05. The trends are closer for UAH at 0.955E-03 and 1.16E-03 for 2000-2014 and 1979-2014 probably because the 2000 anomaly for UAH was much lower at -0.059 compared to 0.091 for RSS.
The “Rosetta” probe has measured the Deuterium to Hydrogen ratio of the comet to be much higher than Earth’s water’s ratio. Rather than speculate on asteroids being the source for terrestrial water, I was wondering if you would care to analyze/ponder/speculate on whether the coma formation and disappearance cycles may not change the isotopic nature of the comet ice found near the surface? Is this worth pursuing?
I have a copy of the court tape. I’ll see that it’s available to you in a day or so.
RSS does really seem to have issues. It is way different than any of the other series, including UAH (which it is supposed to resemble the most).
On a side note I got a kick out of Monckton looking exclusively at RSS . Sure looks like cherry picking to me.
Here’s the difference between RSS and UAH. Clearly something strange is going on after circa 2000.
https://dl.dropboxusercontent.com/u/4520911/Climate/Satellite/rss_minus_uah.png
Carrick – Yes, Spencer commented (here and here) about the RSS-UAH difference a while ago.
P.S. Just noticed this post on the topic as well.
Thanks for the reminder to those posts. Worth including this bit of text:
The difference between UAH and RSS is the gift that keeps on giving. There’s a delicious irony in watching the stumbling walk-back on the part of certain scientists and activists from the position that UAH was obviously flawed (when UAH was running colder than RSS) And in watching the same people, who would have used UAH 10 years ago because it ran cool, endorsing what I think and have thought is a substandard RSS product.
Carrick –
Cherry-picking can be done in any orchard.
The difference between UAH and RSS is the gift that keeps on giving.
In fact looking at the UAH/RSS graphs put up by Dr Roy Spencer the remarkable thing is how in step the 2 graphs actually are. They show that if using unbiased data collection from space that the surface trends are quite similar for the last 30 years.
I repeat quite similar, just look at them.
There is a very slight drift or anomaly between the two sets but the changes that occur up or down are very much in step.
. Unlike the GISS and NOAA Carrick thinks are wonderful with their urban heat effect and doctored statistics otherwise known as guessed or filled in anomalies.
There’s a delicious irony in watching people try to criticize good data sets that very little is wrong with while being unable to see the log in their own eye when it comes to the ridiculous manufactured data sets or in C and W models that are used to postdate and change historical data.
The effect of lowering past historical temperature data in the models is that one can then impute a lower past average and claim the present is warmer.
Obviously this cannot be done with satellite models.
With earth temperatures one is stuck with not being able to change the actual real thermometers hence there is an inbuilt limit to the amount of obfuscation and warming that can be inflicted by the models.
AS C and W have found. When you stuff all the warming you can manufacture into your algorithm and the temperatures start to go down the Kriging has to go down as well. Even in a pacifically warming year.
When this episode of Pacific warming settles, watch the temperatures with all data sets sink and the pause hits home.
What was the average temperature back in 1979-89 and what is it now?
A new post by Steve McIntyre:
“Unprecedented†Model Discrepancy
Given your discussion here, it’s interesting that McIntyre chose RSS. I think most of what he said wouldn’t have changed, though.
Carrick,
“endorsing what I think and have thought is a substandard RSS product.”
Some support here from … Carl Mears:
“A similar, but stronger case can be made using surface temperature datasets, which I consider to be more reliable than satellite datasets (they certainly agree with each other better than the various satellite datasets do!).”
Since Obamacare was a significant topic on the original open thread, I can’t resist commenting on Kathleen Sebelius’ comments about Obamacare. First, it should be noted that her father was John Gilligan a former governor of Ohio and she was a past governor of Kansas in addition to being Secretary of Health and Human Services. You would think that she would know how uninformed many voters are. However, she is so uninformed herself that she had no idea of how lacking in information many people are about insurance or anything involving finance. Here is what she said:
“One of the things that we have learned with the passage of the law [Obamacare], and certainly with open enrollment in 2014–and I think this will be true again in 2015–is that a lot of Americans have no idea what insurance is about, and have no idea, even if they have coverage, what it means, you know, what a deductible is, what a copay is, how to choose a network. Those are complicated terms,” Sebelius told USA Today’s Susan Page.
Sebelius added that Americans’ inability to understand health insurance was a “stunning revelation.”
“I think the financial literacy of a lot of people, particularly people who did not have insurance coverage, or whose employers choose their coverage and kind of present it to them, is very low,” said Sebelius. “That has been a stunning revelation.” http://www.breitbart.com/Big-Government/2014/12/02/Kathleen-Sebelius-Americans-Have-No-Idea-What-Insurance-Is-About No wonder the roll out under her directions was so botched.
None of this is stunning to me. The simple things that people don’t understand are sad and a partial indictment of our educational system. For instance, I had a tenant who left her heat on at 77 degrees in the winter even when she was gone for 8 hours and even though she had a programmable thermostat. (and she was always late with her rent) Another tenant was paying his ex-wife in Pennsylvania child support even though he was taking care of his son in Ohio. (This is not a misprint)
The fact that Sebelius was surprised by the financial illiteracy of many people shows how little she learned from her father and her own stint as governor. She must have lived in a bubble. In light of her cluelessness about our country’s level of education it isn’t surprising that she falls hook, line and sinker for AGW propaganda. For intance, on this video, she states that greenhouse gasses implicate national security concerns. http://www.wellcome.ac.uk/About-us/Policy/Spotlight-issues/Health-impacts-of-climate-change/Public-health-benefits-of-reducing-emissions/WTX058075.htm
JD
For what it’s worth, I didn’t do a graph with UAH and choose the more dramatic. I just did the one. The reason I used RSS was because it has been the preferred satellite measurement by Real Climate and such. If I re-did it with UAH, I doubt that the differences between graphics would be material
Nick Stokes (Comment #133434)
You mean this Carl Mears, I presume ?
Benjamin D. Santer, Céline Bonfils, Jeffrey F. Painter,
Mark D. Zelinka, Carl Mears,
Susan Solomon, Gavin A. Schmidt,
John C. Fyfe, Jason N. S. Cole,
Larissa Nazarenko, Karl E. Taylor & Frank J. Wentz
Authors of
Volcanic contribution to decadal changes in tropospheric temperature
where the conclusion is
“Despite continued growth in atmospheric levels of greenhouse gases, global mean surface and tropospheric temperatures have shown slower warming since 1998 than previously 1, 2, 3, 4, 5. Possible explanations for the slow-down include internal climate variability 3, 4, 6, 7, external cooling influences 1, 2, 4, 8, 9, 10, 11 and observational errors”
So Santer, Solomon and Gavin A. Schmidt are all listed 13/11/2013 as supporting the notion of a pause.
He is in good company, no bias there.
Nick+Stokes (Comment #133434)
December 11th, 2014 at 8:59 pm
Some support here from … Carl Mears:
Lifted straight from Sou at Hot Whopper, Dec 4th
good one Nick.
Us Australians have to hold together.
I agree with you though, it is the man behind the RSS system himself stated as saying this so your point holds water though I disagree with his statement.
though he also says
“there is not much doubt that the rate of warming since the late 1990’s is less than that predicted by most of the IPCC AR5 simulations of historical climate. This can be seen in the RSS data, as well as most other temperature datasets.”
So Carl Mears just picks the data source whose documented problems, flaws and bias is well documented over the satellite record. And just happens to be able to explain away the pause with it.
lol.
JD Ohio
How to choose a network if you are given a choice is not a ‘complicated term‘. It is a complicated decision. Fraught with many consequences. It affects your choice of doctors. Different networks operate differently–with more or less use of the internet to provide you information and web tools might be better organized or worse. Some permit quicker or slower scheduling. Some are better at certain types of care. The choice can affect some aspects of how care are delivered. Worse: Until you experience a particular network, you don’t know how well they might interact with you given your preferences and abilities.
And of course people who didn’t previously have insurance don’t understand “co-pay”, “deductible” and so on. Many middle income people who are generally healthy also don’t quite grasp the sticker shock of ‘deductible”. These stories are just cases– not statistics– but we are reading stuff like this at the New York Times (not exactly an anti-Obama paper)
“Unable to Meet the Deductible or the Doctor”
http://www.nytimes.com/2014/10/18/us/unable-to-meet-the-deductible-or-the-doctor.html
Lucia,
I expect the progressives should be along at some point to tell you that the problem is we didn’t go far enough. Single payer!
It’s as if, having smacked ourselves in the face with a baseball bat of PPACA, we haven’t quite succeeded in achieving coma yet and a few teeth are still hanging in there so we need to go a little further with it.
JD,
Given the bias shown by divorce courts, I wouldn’t be at all surprised if he was paying child support under a court order with the court fully aware that the ex-wife did not have custody of the child.
Gil Garcetti, the former DA of LA county, once literally picked names out of the phone book that matched names given by unwed mothers on birth certificates and filed papers to force payment of child support. The notices of court action were mailed. If the defendant either did not receive the notice or failed to appear in court, the judgement was automatic. Once the judgement was in place, it was nearly impossible to get it lifted. DNA evidence was not sufficient.
I don’t consider that RSS, UAH, STAR and whoever else are independent data sets. They all rely on exactly the same raw data. The same is true for the surface temperature compilations of GISS, CRU, etc. I also don’t consider better agreement between GISS, CRU, etc. as being evidence that those compilations are somehow better, i.e. more accurate, than the satellite compilations.
DeWitt Payne:
That’s not technically correct. UAH uses AQUA. RSS does not.
Here’s a commentary on how the differences in data sources might be affecting RSS:
Steve McIntyre:
RSS has the negative most trend of the satellite series in the last decade. I would be cautious against only using that series on that basis alone.
I would think redoing the analysis with other data sets might be illustrative of how important the spread of trend between surface temperature data product is compared to the discrepancy between surface temperature data products and the climate models.
Looking at the most recent results, here’s the trends for 2000-2013 (inclusive).
rss_monthly -0.006
hadcrut4 0.042
ncdc.temp 0.046
gistemp 0.064
ecmwf 0.079
hadcrut_plus_merra 0.091
cowtan_way_had4_krig_v2_0_0 0.096
hadcrut_plus_krig 0.100
tltday_5.6 0.111
UAH (tltday) has the largest trend over this period, whereas RSS has the smallest value. RSS is the only series that shows a true “pause” in global warming since 2000.
Carrick,
It seems that none of the trends you just posted are showing anything significant at all.
Certainly nothing to support the idea of 2.0o increase in the next 50 or 100 years.
What would your trends look like if you went back 1998? Or even 1970 when the climate fear was regarding the already underway ice age?
Carrick,
UAH stopped using AQUA because of spurious warming in the AMSU data in June 2013. They are now using an average that includes NOAA-15.
RSS and UAH have been using identical data since then.
As I said above, a starting date of 2000 is going to increase the observed trend for UAH significantly because that was a big La Nina year.
Specifically, the trend of the UAH global annual average for 2001-2013 for UAH is 0.050°C/decade while for 2000-2013 it’s 0.110°C/decade.
That still leaves RSS as the outlier.
DeWitt,yes AQUA is no longer being used, but there is as period from 2002-2013 where AQUA was being used by UAH and not by RSS. It would be interesting to compare the behavior of the two series, but that period corresponds to the period where we’ve seen what appears to be a large negative bias in RSS.
Here’s 2002-2013:
rss_monthly -0.079
hadcrut4 -0.030
ncdc.temp -0.020
best -0.005
gistemp -0.004
hadcrut_plus_merra 0.012
ecmwf 0.013
cowtan_way_had4_krig_v2_0_0 0.023
hadcrut_plus_krig 0.029
tltday_5.6 0.029
ENSO of course is alway going to be an issue wrt to interpretation of trends unless you use really long duration windows (e.g., 30 years).
Carrick,
I look forward to your explaining why you picked the starting year you did in support of your claims on trends. The sensitivity of the sensors to things like you describe regarding the RSS problem with its platform is interesting: Yet another indicator that what is going on is subtle and small.
hunter,
2000 was picked because it didn’t include the obvious outlier year 1998 and it’s a nice round number. Unfortunately, 2000 was also something of an outlier in the opposite direction, at least for UAH. As Carrick pointed out, though, it’s difficult to avoid ENSO affecting the trend for periods much shorter than thirty years.
JD+Ohio (Comment #133435)
“The fact that Sebelius was surprised by the financial illiteracy of many people shows how little she learned from her father and her own stint as governor. She must have lived in a bubble. In light of her cluelessness about our country’s level of education it isn’t surprising that she falls hook, line and sinker for AGW propaganda.”
JD, that depends on your view of the political process. Sebelius, like other politicians, I would think is not above feigning ignorance or naivety when it suits their political purposes. It takes little fore thought to figured out that the stumbling block for people was going to be the deductibles in ACA. The politicians in the case of ACA wanted to get the government foot more firmly in the door in the case of medical decisions (or a lack thereof) of the citizenry. The idea was to get the bill passed into law and keep it there for a sufficiently long period so as to make striking it down an onerous and risky political task. Further, the promises made and the features supposedly contained in the law to keep the costs down once the law is firmly establish can be forgotten and the purse strings loosened.
We have members with congressional oversight responsibilities for the CIA’s handling of the captured terrorists very self righteously denying even a hint of what was going on while condemning what they were supposed to control. That response is very much in vein of Sebelius. I am no fan of the CIA or any other undercover government agency, but like the banks making bad mortgage loans with the encouragement of well place congressional members it makes these congress people with oversight duties look like fools or agents encouraging bad behavior by looking the other way.
hunter, what DeWitt said:
If you use ex ante rules for selection of boundaries (like aligning with natural boundaries), it pretty much eliminates cherry picking. However 2000 wasn’t ideal because La Niña is a significant player, and ENSO seems to affect the TLT data more than the 2-m surface data.
That seems to be a reasonable argument for shifting the starting point to 2002.
2013 is the last year with complete data for all of the series I considered.
IMO, RSS’s problems are not subtle.
To save arguing about the starting point, here is a plot showing trends from year x to present, for many of the datasets Carrick listed. It clearly shows how RSS is an outlier. There’s an active version here which shows how these trends have been changing with new months.
ps hunter, Carl Mears is the man behind RSS.
That would also explain why the current satellite anomaly isn’t close to breaking a record in 2014.
Which raises the obvious question of why. Perhaps the 2m surface data under weights or underestimates the change in tropical sea surface temperature during an El Niño or La Niña or that the MSU data overestimates it.
Nick, the graph is off the chart hence up interpretable. Have you got a better one.
UAH and RSS run in tandem with there trends and changes, the only discrepancy is a very slight difference in the degree of change.
They are measuring the same sort of change just with an added C (for a constant dependent on time).
There is no difference in the message they are sending about cooling and warming changes.
Including Cowtan really bugs me as it is not a data set. Is a model of what data could be using their algorithm.
Because it implicitly demands warming, warming is what you will get.
It therefore does not and cannot fit,match or resemble other real data sets.
If you put up a clearer graph this would be obvious to all of us. Cowtan goes up where all other models go down.
Fortunately when you overhype all the data into 3 or 4 basic stations, which is what they have done, and rule out other stations completely, which is what they have done, you get egg all over your face when your “specifically chosen and puny sample starts to cool.
Heck if we have a cooler Arctic summer, fingers crossed, Cowtan and Way will show a 1930’s dive in 2015
angech,
“Nick, the graph is off the chart hence up interpretable. Have you got a better one.
UAH and RSS run in tandem with there trends and changes, the only discrepancy is a very slight difference in the degree of change.”
The graph was prepared to show the comings and goings of the pause. All the data points cited by Carrick lie within the y-range shown.
UAH is at or near the top of the range, in trend. Till 2005, RSS is way below the others. From about 1996 to 2003 the gap is about equal to the entire range of the rest.
DeWitt:
My guess is this is a real feature of the data and reflects the difference in the physics of the boundary layer (of which 70% of is in contact with the oceans) and the physics of the lower troposphere (the part above the ABL and partially decoupled from surface physics).
Nick, thanks for the link to the plots. I added MERRA and NCEP reanalyses, as well as BEST (BEST appears to have stopped updating their time series data on their blog, so I wasn’t including it in my script; however I can still get it in a non-automated fashion using Climate Explorer).
You had noticed on your blog that MERRA is running cool compared to the other reconstructions.
I’m seeing this here too.
merra -0.131
rss_monthly -0.079
hadcrut4 -0.030
ncdc.temp -0.020
gistemp -0.004
hadcrut_plus_merra 0.012
ecmwf 0.013
cowtan_way_had4_krig_v2_0_0 0.023
ncep 0.025
hadcrut_plus_krig 0.029
tltday_5.6 0.029
best 0.047
Curious, because if I would have picked, without looking at the answer, which reanalysis series I thought was most reliable, it would have been MERRA. According to their documentation MERRA is supposed to not have issues with baseline drift (the putative origin of the artifactual cooling in RSS). Maybe they need to check again.
C&W + Merra does seem to be a “value added” product.
On another note, Steve McIntyre has generated a curve for UAH TLT, and added it to the bottom of his post.
figure.
Carrick,
Since most of the energy radiated away by the atmosphere comes from the surface, the naive view would be that the LT would be less sensitive to changes in SST rather than more. Off the top of my head, only a change in lapse rate would have the observed effect. Shouldn’t someone have noticed this by now?
Carrick (Comment #133446),
Are those values per decade trends, or overall trend for the 14 years?
DeWitt,
” Off the top of my head, only a change in lapse rate would have the observed effect. Shouldn’t someone have noticed this by now?”
Not sure I understand what you are getting at. The expected (modeled) mid tropospheric warming is way higher than measured (satellite or balloon). Does that not qualify as at least a discrepancy in lapse rate?
SteveF,
Satellite measurements appear to show a larger ΔT during ENSO events than the surface temperature as evidenced by the average anomaly for 2014 almost certainly being nowhere close to the peaks reached in the El Nino years of 1998 and 2010. So maybe the tropospheric warming in the models just has the magnitude too high. That wouldn’t be too surprising.
SteveF, they are per decade number (OLS trends).
It therefore does not and cannot fit,match or resemble other real data sets.
If you put up a clearer graph this would be obvious to all of us. Cowtan goes up where all other models go down.
Fortunately when you overhype all the data into 3 or 4 basic stations, which is what they have done, and rule out other stations completely, which is what they have done, you get egg all over your face when your “specifically chosen and puny sample starts to cool.
##############
Nope. Its been validated against Bouys.
I validated it against Airs.
all data series estimate the arctic.
nope
“all data series estimate the arctic. nope.”
True, so far we are on the same page,
all global temperature data series estimate the arctic. nope.
That would be what you meant to say,
True.
Would you care to let Nick et al in on how many sets actually estimate it? Out of their 12.
One.
You would say Cowtan and Way as you said
“I validated it”
The it is very informative, not them, for 2 or more is it.
I would say none.
As I said, “It” is not a data set, A model.
And a possible bad one. Why.
“As you know the number of land stations in the World Meteorological Organization (WMO) list above 80 degrees North is very small indeed (a handful),”
So five stations, remembering that C/W removed three nuisance stations that were too cold from their calculations [37.5%] and one guess is all they are going on.
Nope. Its been validated against Bouys.
In water. not many of them.
I validated it against Airs.
Atmospheric Infrared Sounder (AIRS) on the NASA Aqua satellite.
Would you count this as a data set estimate of the arctic then or does it miss areas in it’s flight path.
Curious to know.
Sorry to seem dismissive of your answer, You are the best credentialed person here and you have put the work in.
I am looking for flaws in the warming argument in the same way that others are looking for flaws in the skeptical argument.
Steve Mosher: “Nope. Its been validated against Bouys.
I validated it against Airs.”
angech: “In water. not many of them.”
And the buoys don’t report in winter, which is where the controversy (for me) is for HadCRUT + UAH hybrid.
angech: “Atmospheric Infrared Sounder (AIRS) on the NASA Aqua satellite.”
If that ‘s what he means, then its a useless comparison. AIRS doesn’t measure 2-m above surface. It is missing the ABL entirely.
On physical grounds, using a simple linear model to combine HadCRUT + UAH is almost certainty wrong. But I happen to like HadCRUT + MERRA and HadCRUT + NCEP. The reanalyses are based on physical models + data, which is must be better than ad hoc statistical models (like BEST).
DeWitt:
I think this is a good point.
I think issue with the comparison though is the the different contributions (e.g., ENSO 3.4 region) show up with different magnitudes and phases and latencies in the TLT data than in the 2-m surface data.
What you get in the global average depends intricately on the exact phasing relationship of the components you are averaging together to produce a global index, so its not surprising to me that the short-period (less than 5-year period) fluctuations don’t line up with each other between TLT and 2-m data .
Some of what we don’t see in the global mean average are signal components which tend to average out in the global average due to the wavelike (oscillatory) nature of their propagation across the Earth. Change the weighting of the components, and what survives in the global average get shifted too.
Does the bigger picture here with regards to temperature trends, or at least as I see it, tend to get lost in the weeds of the detailed discussion we have with regards to the “pause”? I would think that we need at least 35 years of comparison of the model trends with the observed trends to avoid most but not all of the uncertainty of the weather noise. My 35 year or greater time comparison have shown statistically significant differences between the observed and modeled temperature trends. Even those comparisons can suffer from not finding and using the proper trend line (which most certainly is not linear) in order in turn to find the residuals that are required for adjusting the confidence intervals. The longer time period for comparisons might also be used to overcome some of the differences in trends estimated using various observed temperature data sets. If we talk about differences in 15 year or so trends from the various observed data sets should we not be including the uncertainty of those data sets or at least how one would go about estimating those confidence intervals so we could talk about statistically significant differences? If we judge that there are significant differences in the observed data sets then we should perhaps consider the observed data and the adjustments of it a work in progress and not set in stone as might be implied by some of these modeled to observed comparisons made.
While I think that most parties who make the modeled to observed temperature trend comparisons take into consideration that the observed estimate is a single realization of chaotic system and the individual model with multiple runs simulates several realizations any of which could hypothetically be that realized by observation, I am sometimes disappointed to see the detail of the individual model (and observed) noise left out of the analyses. I am currently plotting the Pre-Industrial Control runs from CMIP5 model runs in order to get a better and detailed picture of the noise and cyclical character of the model temperature series. I plan to link those plots here along with some ARMA model data I have from the residuals. My analysis to date shows some very different character in these individual model control series.
More to the point of these discussions, and again in my view, is attempting to estimate the modeled versus observed effects on temperature changes caused by AGW related forcing which is currently estimated in terms of Equilibrium Climate Sensitivity (ECS) and Transient Climate Response(TCR). Nic Lewis has done a considerable amount of work in this area of late and has a detailed review of the published papers that have attempted to separate the observed from the modeled estimates. A number the observed estimates have used model data to make their estimates and are thus not completely independent of the model results. Nic has zeroed in on using separate parts of the observed historical instrumental record where the AGW effects, or at least GHG part, is mainly absent and where it has affect the temperature changes and is of a known quantity. The time period choices also are made in attempts to avoid differences due to volcanic activity. Nic Lewis review is linked here:
http://www.thegwpf.org/content/uploads/2014/02/A-Sensitive-Matter-Foreword-inc.pdf
I find it interesting that the model data used to estimate ECS and TCR can for some models be sufficiently disconnected from the temperature data to not fit a correlation. Also Nic Lewis has not been shy about pointing to variable uncertainty that can affect the results of an independent estimate of the observed ECS and TCR.
I am more inclined in my layperson approach in these matters to look in detail at the CMIP5 individual models and differences in models and determine, at least, whether the majority of these models are ready for a prime time comparison with the observed results – and with the given that the observed results are not set in stone.
Kenneth Fritsch, I guess it’s a matter of perspective what you think the “bigger picture” is.
I think it’s very important to understand the instrumentation if you’re going to try and make use of them to delve into the agreement (or lack there of) between model and data.
Saying we need to have 35-years is a bit of a punt, IMO. It’s not obvious this is long enough, especially given the impact the putative 60-year cycle has on the trend you get:
Figure.
Ideally you’d like 200 years of measurements since the start of the anthropocene, but we need answers sooner than that, so reliance on models that allow us to shorten the required observational period is a requirement for policy makers.
Carrick (Comment #133471)
Carrick, I agree in essence with what you say here. The bigger picture in my view is estimating that part of the temperature increase that we can attribute to AGW and primarily GHGs and in turn use that in an attempt to evaluate the climate model results both historically and into the future. To accomplish that does require a good understanding of the instrumental record. I also think that finding valid temperature proxies for looking back in time would help in evaluating climate models.
I do think the most direct route to portioning out the warming due to GHGs and the future effects of increases/decreases in those gases is estimating ECS and TCR from observed data (where we need reasonably accurate instrumental data) and comparing it to the climate models. Like you note in your post we have uncertainties that can cloud the comparisons that would otherwise allow a reasonable evaluation of the models.
In the meantime I think we can gain at least a qualitative view of what the climate models have to offer by looking at the individual model output in some detail and making comparisons amongst models. In my layperson view, which I admit may result from my own lack of comprehensively reading the topical literature, what I see from literature is more of a black box nature and without the detail I am most interested in – something like I see with the lack of proxy detail for temperature reconstructions.
Kenneth,
My view of the models is that they have not evolved as much as they might have due to political considerations…. no one in the field seems willing to even discuss the need for winnowing out the obviously wrong models. And I don’t mean just those models which grossly overstate warming; some models yield variability which is either comically too high, comically too low, or with much too high or too low a frequency of variation. I find it very strange that the obviously silly models continue to be funded, in spite of inviting ridicule.
.
The other political influence is the apparent refusal to discount those models that are so high in rate of warming compared to reality that they couldn’t possibly be accurate. Yet they are dutifully included in the ‘model ensemble’. The only plausible explanation IMO is the desire to maintain plausibility of very rapid future warming.
SteveF:
I think that pretty much nails it. Not just green politics, but I think politics at the academic level too, have prevented the exploration of the full parameter space available to climate modeling (to channel Steven Mosher for a second).
By the way, some good comments on Steve McIntyre’s blog on his Unprecedented model discrepancy post. Worth the read if you get a chance.
“remain in ore”? Does this refer to models successfully predicting the location of valuables in the mining business?
Re:”obviously silly models” .. “Model IAP-FGOALS1 […] is known to produce nearly perpetual ice cover out to almost 40°S.” Even if it predicted global average temperature pretty accurately, this model would hardly seem to be a reliable simulation of reality.
“If that ‘s what he means, then its a useless comparison. AIRS doesn’t measure 2-m above surface. It is missing the ABL entirely.”
wrong.
AIRS gives you estimates of the surface.
AIRS gives you estimates of SAT.
AIRS gives you estimates at 24 pressure levels as well.
So, its pretty easy.
with the arctic you have these choices.
1. dont infill. This imputes a trend in the arctic that is equal to the trend of the whole global. check this trend against the AIRS
trend for the surface and for SAT and for 1018 hPa or whatever
you like
2. extraplolate. Assert the trend at 80N is constant up to 90.
3. Infer the trend from Satillite trends.
check 1 and 2 against merra, check then against bouys check them against AIRS (LST and SAT), check them against the new 30 year Ice surface trends.
Guess what.. #3 does better
the bottom line is this.
There is NO independent evidence that suggest #1 is a better estimate. NO independent evidence that #2 is a better evidence.
#3… has multiple supporting data sources.
perfect? nope.
just better.
Steven Mosher—AIRS gives you skin temperature. It does not give you 2-m above surface temperature.
Sometimes the relationship between skin and 2-m temperature is very good, other times very poor (this is well studied in boundary layer meteorology). This is why there is no 2-m temperature product in any of the satellite series.
Secondly, the issue with satellite is that there isn’t a linear relationship between 2-m and TLT…instead there is a complex transfer function that relates them (which probably isn’t constant).
Also, using AIRS to validate a satellite surface temperature hybrid, well that seems to me to be circular reasoning. Regardless, you don’t have to choose to use a linear model, which is guaranteed to be unphysical.
Next, as C&W have realized you don’t need to use satellite, you can use reanalysis products instead. This seems more likely to give a correct result than the hybrid UAH method they original proposed.
Finally, nobody is arguing that no-infilling is better. But if you have infill methods, and you can fix some of the the problems with them, you should.
As to anything is better than nothing, depends on your definition of “better”. My definition would be “closer to the truth”. Obviously not all infill methods are equal, so some will be closer, others may be worse than the default infill method (replace with global mean temperature).
AIRS has a 2 metter product
http://catalog.data.gov/dataset/aqua-airs-level-3-8-day-standard-physical-retrieval-airsamsu-v006
note surface air temp which is a 2meter product
one of the validations of the 2meter product
http://www.aoml.noaa.gov/phod/docs/Dong_etal2010.pdf
Bouys .. you need to read more and get the data
an example
http://iabp.apl.washington.edu/pdfs/InoueEtal2009_ImpactOfBuoyOnPressureFields.pdf
SteveF (Comment #133473)
“My view of the models is that they have not evolved as much as they might have due to political considerations…. no one in the field seems willing to even discuss the need for winnowing out the obviously wrong models.”
One question has remained begging for a long while. The Climate sensitivity to CO2 changes inherent in these models appears to be the problem, not the models themselves.
Why will not Mosher or Zeke or Nick run some of these [their] models with a CO2 sensitivity a half or a quarter of what they expect and see what happens.
My guess is with good models [good data] that this one change would bring most of these models into line.
Given the ongoing work put into them [new data as it arises] I would also suspect that the predictive results are a lot better than most expect.
Or have they already done it and been burned by the amazing closeness to our actual weather, hence too ashamed to publish it.
What about it Steven, not your time, the departments’.
Surely in the name of validation these experiments have already been done, why not show the results.
Carrick, thank you for your links ,here and at CA, appreciated
Steven Mosher, my apologies. I stated what I meant imperfectly: AIRS has a product but they don’t have data. They measure skin temperature, and have a model that relates it to near surface temperature. This works very well some places, not so well other places. The validation paper you linked is for the Southern Ocean, one of the places, in my understanding, the boundary layer people would expect to work well.
Unfortunately the Arctic is precisely one place where the satellites fails to validate. I believe Judith Curry has some work in this area too (she has also stated reservations about the ability of satellites to accurately characterize near surface air temperatures in the Arctic.)
I have played with the bouy data. In my memory, there are big wintertime gaps. Is this a faulty recollection?
angech—you’re welcome.
Here’s a more recent paper which addresses the issues with measurements in the Arctic. Note this is looking at skin temperature for MODIS vs AIRS. Skin itself can be a difficult measurements.
Of course, if there are issues for the primary measured quantity that will propagate to derived quantities as well.
Comparison between MODIS and AIRS/AMSU satellite-derived surface skin temperatures
The arguments over the reasons for small fractions of a degree of possible change- real or inferred, modeled or measured- distract from the real point that no climate catastrophe is taking place.
Steven Mosher (#133477) –
I know you don’t like this option, but you did not mention option #4 — mask. Mathematically, it’s the same with respect to observations as #1 (don’t infill). The difference is that the observational mask is applied to the models as well. As you’ve stated before, global average surface temperature is merely a metric, and a metric which covers 90% (or whatever) of the surface is as useful as one which covers 100%. And we don’t have to extrapolate or make assumptions about the behavior in places where observations are not made.
“Surface Air Temperatures above the Melting Ice in the Arctic Summer – Most of the area above 80N is (currently) still covered in permanent sea ice. In the Arctic Summer when the surface ice is melting, the air temperature close to the surface is limited by this ice melt temperature to just above zero degrees C”.
I had not realized this very curious fact.
Problem.
So how does the Arctic warm at 8 times the rate of the rest of the world given this curious fact.
Particularly as it cannot warm [for the last 2 years in particular] beyond the freezing point at the hottest part of the year!
And why do we need Krigging to infill an area that cannot get hotter in Summer, when all you need is a satellite area map x 0 degrees for the most important 3 months.
You do not need weather stations!
It raises serious doubts that claimed x8 warming in the coldest part of the year in the Arctic can possibly generate enough heat to explain away the pause.
angech,
go here: http://data.giss.nasa.gov/gistemp/seas_cycle.html
Enter date range from 1979 to 2013
Observe that the max warming in the Arctic is from Autumn through Spring.
Observe that the min warming in the Arctic is during Summer.
This is because the thinning of Arctic Sea Ice allows more ocean heat into the atmosphere.
Now enter the date range from 1910 to 1945
Observe that the same pattern occurred earlier
Periods of Arctic Sea Ice decline are not new
The question is, how much of the current decline, and resulting pattern, are natural variation, and how much are due to ‘global warming’?
angech,
It warms because the surface temperature in winter increases. That temperature is on the order of -30C if the sky is clear and the air is calm so a temperature inversion can develop. But the winter temperature when the sun is below the horizon is sensitive to radiation from the air above the inversion layer as well as heat transfer through the ice from the water below. The thickness of the ice is also a factor. A warmer surface temperature implies thinner ice.
Increasing greenhouse gas concentration has the same effect at high latitudes as it does at low latitudes, for a given rate of power transmission, the surface temperature must be higher if the ghg concentration is higher. Air in contact with ice at low temperature requires less energy to change temperature because there is little change in specific humidity.
Increasing global average temperature for whatever reason tends to flatten the latitudinal temperature profile. The means the poles must warm faster than the equator. Fifty million years ago, Antarctica was covered with a boreal forest rather than an ice cap. Meanwhile, the Tropics were only a little warmer than now and home to a variety of life, not uninhabitable. The main difference between then and now was not geographic location but that the connection between Antarctica and South America still existed.
Also, remove the Antarctic ice cap and the Arctic sea ice in the summer and the melting ice constraint on summer temperature no longer holds. Note, I’m not saying that could happen any time soon, where soon could be measured in millenia for the continental ice caps on Greenland and Antarctica.
More recently, the Antarctic ice core isotope ratio data indicates that the ΔT between glacial and interglacial epochs is two to three times the ΔT in the Tropics.
angech,
“The Climate sensitivity to CO2 changes inherent in these models appears to be the problem, not the models themselves.”
.
You do not understand what the models do. The sensitivity to CO2 forcing is an emergent property, not a parameter that is an input to the model. The sensitivity to GHG forcing is determined, within a narrow range by the basics of radiative heat transfer. The value is close to 1.1C for the radiative forcing equal to that for a doubling of CO2. This value is not seriously in dispute among scientists ( including well known skeptical scientists). The increase in radiative forcing due to increased water vapor at that slightly higher temperature ‘amplifies’ the direct warming from man made GhG’s by a bit less than a factor of two, bringing the expected warming to about 2C for the forcing from a doubling of CO2. While there is some (modest) uncertainty about the contribution of increasing water vapor, that this contribution is real and significant is also not in serious dispute.
.
What is in serious dispute is the additional amplification in the models due to other ‘positive feedbacks’, the most important of which is the influence of clouds. The influence of clouds is one of the least certain influences, and is the difference between mild warming (slightly negative feedback) and catastrophic (strongly positive feedback). In fact, most of the difference in diagnosed sensitivity from different models is demonstrably due differences in how cloud behavior is parameterized in different models.
.
You might wonder how different models with very different diagnosed sensitivity (2-4.5) can all (sort of) match historical warming. The answer is that each modeling group is free to choose historical ‘offsetting’ cooling from man made aerosols and ocean heat uptake….. the assumed aerosol offset is nothing but an arbitrary kludge, and the ocean heat uptake, especially before ARGO, little more than a wild guess. The use of GCM’s to ‘determine’ climate sensitivity is IMO intellectually corrupt. The models are grossly (on average) too high in sensitivity…. for political reasons. Which is why empirical determinations of sensitivity like Nic Lewis and others have done are IMO much more important than GCM in guiding rational public policy choices.
.
When you say something like “The Climate sensitivity to CO2 changes inherent in these models appears to be the problem, not the models themselves.” people who DO understand what is disputed will either giggle or ignore you. You should understand the issues…. before spouting nonsense.
Toto,
I suspect we’ll have some sort of handle on that by 2020. The most sensitive measure will probably be ice volume. My guess is that the August PIOMAS volume anomaly will be well above the 1979-2013 OLS linear trend line by then.
I think it is semantics. The models should all have an “effective climate sensitivity” which is derived from what SteveF has documented, but could be measured at the high level by running the different RCP scenarios and doing some calculations.
If you want to make a lower sensitivity model than you can alter the cloud parameters or whatever else controls the emergent property.
All the models may be running hot because they all make similar wrong assumptions somewhere. I think the question is whether they have allowed for a large enough diversity of models to make it easier to determine which ones are working better for reasons that can be determined by comparing models.
SteveF (Comment #133490)
“The sensitivity to CO2 forcing is an emergent property, not a parameter that is an input to the model”.
contradicts your second sentence.
“The sensitivity to GHG forcing is determined, within a narrow range by the basics of radiative heat transfer. The value is close to 1.1C for the radiative forcing equal to that for a doubling of CO2. This value is not seriously in dispute among scientists ( including well known skeptical scientists).
You are right.
“The increase in radiative forcing due to increased water vapor at that slightly higher temperature ‘amplifies’ the direct warming from man made GhG’s by a bit less than a factor of two, bringing the expected warming to about 2C for the forcing from a doubling of CO2. While there is some (modest) uncertainty about the contribution of increasing water vapor, that this contribution is real and significant is also not in serious dispute.”
OK,
some caution about multiple bites of the cherry, If Babe Ruth hits a homer,the next guy does not get a free swing at the same ball. Lucia cautions about zero climate sensitivity in that feedbacks should not normally be greater than the original force, particularly negative ones. I would think this argument also holds for positive feedbacks for the same reason
What is in serious dispute is the additional amplification in the models due to other ‘positive feedbacks’, the most important of which is the influence of clouds.
Nailed it on the head but your use of positive feedbacks, slightly negative feedbacks and strongly positive feedbacks seems to emphasize a slightly positive approach. The problem here is that the failure of the models all suggest the feedbacks of clouds, aerosols etc might be strongly positive.
The influence of clouds is one of the least certain influences, and is the difference between mild warming (slightly negative feedback) and catastrophic (strongly positive feedback). In fact, most of the difference in diagnosed sensitivity from different models is demonstrably due differences in how cloud behavior is parameterized in different models.
Exactly what some skeptics are saying.
You might wonder how different models with very different diagnosed sensitivity (2-4.5) can all (sort of) match historical warming. The answer is that each modeling group is free to choose historical ‘offsetting’ cooling from man made aerosols and ocean heat uptake….. the assumed aerosol offset is nothing but an arbitrary kludge, and the ocean heat uptake, especially before ARGO, little more than a wild guess.
Agreed, They have not chosen enough ongoing offset have they?
Does historical mean non ongoing in your definition?
But this is what we are both arguing in different ways.
If you choose these offsets, particularly if they are ongoing, you are arguing for a lower climate sensitivity input, They have just not made it low enough as a low sensitivity causes less alarm.
The use of GCM’s to ‘determine’ climate sensitivity is IMO intellectually corrupt. The models are grossly (on average) too high in sensitivity…. for political reasons. Which is why empirical determinations of sensitivity like Nic Lewis and others have done are IMO much more important than GCM in guiding rational public policy choices.
Agreed
When you say something like “The Climate sensitivity to CO2 changes inherent in these models appears to be the problem, not the models themselves.†people who DO understand what is disputed will either giggle or ignore you. You should understand the issues…. before spouting nonsense.
Back to your first comment which the rest of your excellent answer seems to contradict as I basically agree with everything you are saying
DeWitt+Payne (Comment #133488)
angech,
It warms because the surface temperature in winter increases. That temperature is on the order of -30C if the sky is clear and the air is calm so a temperature inversion can develop.
I know what you are trying to say, but it comes across like the chestnut of “there is more ice in Antarctica in the winter because it is warming” theme for increasing ice at the other pole.
“It warms because the surface temperature in winter increases”
No, it gets colder because the surface temperature in winter decreases, Winter is colder.
What you meant to say and said is that the cloud cover which is 80% of the time in winter, the coldest months, traps and radiates more infra red back thus preventing the Arctic from getting as cold as it should. The second reason is heat from the sea particularly through cracks and vents which stop it dropping much below minus 35 degrees.
It is much warmer in the Summer, 30-35 degrees warmer, but cannot get above zero basically because the extra heat goes into melting ice rather than heating the air.
Temperature inversions are more a feature of Greenland and Siberia where the land present with high ice sheets and valleys creates a true temperature inversion that can make the air warmer higher than lower. Hence Greenland can have high land surface areas much colder than the less insolated Arctic sea ice surface. My comments are more directed to the Arctic sea ice surface temperature.
A warmer surface temperature implies thinner ice. true but it also implies it is in summer, not winter.
Agree with the rest of your summary.
angech,
I admit my phrasing was not precise, but you seem to be purposely misreading what I wrote. To be more precise, the winter temperature anomaly increases much more rapidly than the summer temperature anomaly with increasing global temperature for whatever reason. Obviously it’s colder in the winter than in the summer.
If you think that temperature inversions occur only over land based ice and snow, try reading this: Atmospheric Inversion Strength over Polar Oceans in Winter Regulated by Sea Ice
Also, cloud cover percentage is 60% in the Arctic winter, not 80%.
From another reference:
[emphasis added]
DeWitt+Payne Thank you for the reference and clearing up the inversion issue. I agreed with most of what you and Steve F have written.
I have been looking in detail at the CMIP5 model runs from KNMI in attempts to find any differences amongst the models and model runs and the four observed temperature series also from KNMI. I decided that my best choice in starting this analysis and showing the results, at least for now, was to perform a Singular Spectrum Analysis (SSA) decomposition of the annual global mean temperature series from CMIP5 Pre-Industrial Control (PIC) and Historical (Hist) model runss along with the observed series. I used the ssa function from the R library(Rssa) and plotted the principle component combinations of PC12 which is imputed to estimate the secular trend and the combinations of PC34, PC56 and PC78 which represent some of the cyclical properties in the series. I also plotted the resulting SSA residual on the same plot. I also did an ARMA model of SSA residuals and put those results in a table.
There is an obvious problem in attempting to present in one place or link all the results from 4 observed, 41 PIC and 149 Hist runs. I settled on using the plots of the above described decomposition components using the line colors of red, dark green, blue, purple and black for the PC12, PC34, PC56, PC78 and residuals, respectively.
Since I wanted not only model to model and model run to model run comparisons but model to observed comparisons, I limited the model runs to the 150 ending years for the PIC runs and 1850-2005 for the Hist runs. For CW Infilled HadCRUT4 I used 1850-2014; for HadCRUT4 I used 1878-2014; for GISS 1200 km and GHCN I used 1880-2014. My primary interest at this stage of my investigation was to compare the noise levels in the model runs and observed series.
The results are linked below to a drop box in an Excel Worksheet with the plots of the SSA decomposition of the observed and Hist model runs appearing at the top of the worksheet and starting on the left side while those plots for the PIC runs are presented in the worksheet directly below. The table with the ARMA model results is presented to the right of the observed and Hist model plots.
I have presented this information for any interested posters here to peruse and make comments. In general in the observed to Hist comparisons the 4 observed series decompositions look much the same although not exactly the same and more like one another than the modeled series do in general. The model runs vary in noise levels and in general have a higher level than the observed series. The observed series fitted ARMA models as a group are different than those of the models. The models show variations in ar coefficients while the observed series appear to show less with the exception the CW HadCRUT4 series which appears further from the ar1 average than the other observed series.
The secular trends vary more for the model runs than the observed series and for some runs it appears that that PC combination of 1 and 2 is being affected by what is showing for the PC combination in 3 and 4. I do not know whether there is some interaction between the trend and cyclical noise that causes this result or is simply a limitation of the SSA to make the proper separations. Regarless it is a difference between models.
Also of note is that the decomposition components for the PIC runs are much reduced from those for the Hist runs. There is evidence for this in the SSA plots and the ARMA models of the SSA residuals. The decomposition of the PIC runs shows little net trend as it should over a 150 year period for this control run. A notable exception is the plot for GFLD-CM3 which shows a trend and also more noise in the other decomposition components. That might indicate an interaction between PC12 and the other PC combinations. However, there are plots in the Hist model plots that have normally large secular trends and little noise from the other PC combinations.
Please note that all plots have the same y axis scale.
https://www.dropbox.com/s/d10imd5x7l9fj8u/SSA_Obs_CMIP5_Models.xlsx?dl=0
I should have noted in my post above that the ARMA models of the SSA residuals were selected by first determining that the first 5 orders of ar produced at least one model with a Box test p.value greater than 0.20 (to insure reasonable independence of the ARMA residuals) and then the aic score was used to determine the best model with independent residuals. If none of the first 5 order ar models produced a p.value greater 0.20, the table will have an NA entry.
“I know you don’t like this option, but you did not mention option #4 — mask. Mathematically, it’s the same with respect to observations as #1 (don’t infill). The difference is that the observational mask is applied to the models as well.”
That is ALREADY DONE when you do comparisons with models.
For example in Ar4 and 5.
I’m not talking about that. I’m talking about estimating the arctic.
you got 3 options.
when face with those options you have to test which is the best.
the test shows that C&W is superior.
but folks can still wave their arms about what might be.
ya monkeys might be flying out of my butt
JD Ohio – Don’t worry about the Sebelius’s financial literacy problem, the Federal Government has a new agency, the CFPB, that is working very hard to fix that right up.
Steven Mosher: “[masking] is ALREADY DONE when you do comparisons with models. For example in Ar4 and 5.”
No, it is rarely done. Ed Hawkins’s well-known chart (not in AR5) is one of the few cases where I’ve seen it done.* AR5 WG1 Fig. 11.25 (=TS.14), for example, isn’t masked. Nor is AR5 WG1 Fig. 1.4. As for AR4, it’s been a while since I’ve looked at it, but WG1 Fig. 1.1 doesn’t appear to use masking, nor does Fig. 10.26.
* Note that his up-to-date version does not mask.
.
It’s fine to discuss the best method for estimating the Arctic. I’m sure there are contexts where it’s important. But it’s not necessary to do so, in order to compare models and observations. I see no reason to make that task more difficult.
Steven Mosher: #133501) “[masking] is ALREADY DONE when you do comparisons with models. For example in Ar4 and 5.â€
No, it is rarely done. Ed Hawkins’s well-known chart (not in AR5) is one of the few cases HaroldW (Comment #133503).
So, who is right??
Angech,
I don’t much care. Regardless of the question of masking & whether or not it’s done, I think there’s still value in C&W’s approach. From my perspective it’s simple and obvious that it’s going to give a better estimate that otherwise.
Could it be wrong? Sure. Do I have some specific reason to think it is? Not really. Although to be fair I seem to remember Judith Curry wasn’t impressed and either I never understood or have forgot why that was the case. People who are dead set on disliking Cowtan & Way might look there for solid ground to stand on I guess.
In the first link below I have tables showing the 15 and 30 year linear trend confidence intervals (CIs) derived from 10000 and 5000 simulations of the ARMA model which in turn was fitted to the SSA residuals from the CMIP5 Pre-Industrial Control(41) and Historical (149) and the Observed series(4). In the second link below I have tables of the portion of the variance explained for these same groups from the first link.
The ARMA coefficients standing alone are rather abstract and therefore to put these data into more understandable terms I did 15 and 30 year linear trends and calculated the CIs. Linear trends where used as mere convenience in producing a metric. Recall that the ARMA models were fitted for each individual series from the SSA residuals remaining after spectrally decomposing the series with the first 8 principle components (PCs). The ARMA model is essentially the red/white noise remaining after removing much of the variance due to the secular trend and the periodic components of the series. The portion of variance explained by the 8 PCs is obviously going to be different where there is a relatively large secular trend as occurs in the Historical model and Observed series. The Pre-Control model runs with little or no secular trends should then better show how much red/white noise remains after removing the first 8 PC components.
I have wondered for a time about whether climate models for comparison to observed series for temperature should be eliminated if the noise and periodic components differed significantly from that reasonably expected from the observed series. The differing noise levels in the models can confuse these comparisons by adding and subtracting to the trends estimated. Now that I have a reasonably good feel for the noise levels in the models (and the observed series), I want next to look at the Equilibrium Climate Sensitivity (ECS) and Transient Climate Response (TCR) for the climate models. These emergent properties of the climate model can be determined from the data presented at KNMI.
I have assumed that these models had more or less invariant trends, related to the models ECS and TCR, in multiple runs of the same individual model that was only varied when estimated by random white/red noise and periodic structure.
http://imagizer.imageshack.us/v2/1600x1200q90/909/ZBD9jr.png
http://imagizer.imageshack.us/v2/1600x1200q90/540/GIZRvA.png
Kenneth,
If we took up a collection and sent you the money, would you write a guest post covering this work?
HaroldW (Comment #133503)
“It’s fine to discuss the best method for estimating the Arctic. I’m sure there are contexts where it’s important. But it’s not necessary to do so, in order to compare models and observations. I see no reason to make that task more difficult.”
I have not thought through what you say, Harold, about excluding the Arctic from a model to observed comparison (which is what I have assumed you are suggesting here). If there is interaction between the more extreme periodic warming/cooling in the Arctic with the remainder of the globe, I would think that the observed to model comparison might change depending on timing of the warming/cooling period. That would in turn be problematic since the models are not able to time period events well.
Kenneth –
There are broad claims that the ice coverage in the Arctic is responsible for effects in the northern temperate zone, so yes, it’s possible that the comparison will look different. And if you’re looking into, for example, the “stadium wave” hypothesis, you’d definitely want to make an estimate. In which case, C&W may well be your best bet — Steven seems to think so, and I see no reason to doubt him on this.
To make a “50,000-foot” evaluation of GCMs, though, you’re interested in the secular trend. I don’t think that the exclusion of the Arctic detracts from your ability to make such a comparison. You have a problem to solve, and it has some hard parts. One of those hard parts is what to do about under-sampled regions. By redefining the problem (to a masked comparison) you can make that particular tricky bit go away. It’s an energy-saving device. [Saving the planet!]
SteveF (Comment #133507)
SteveF, you have presented the case for what I call the retired guy dilemma. If I am able to do my simple minded analyses without worrying about compensation (which in this case was not a serious offer to be sure nor do I expect any offers for my work in this area) or putting it out in good finished form, I am happy, satisfied, in a learning mode and most importantly not stressed.
What I have presented here is just a part of my own planned analyses of the climate model data that is available in good form at KNMI. In this case and point in time I have no hard conclusions to discuss other than the differences in noise levels which are rather self evident. I posted my results here for anybody interested to use or comment on and for criticism if someone finds errors in the analyses or my approach.
I have some fundamental questions about the climate models and the difference among models and model runs that are much like those I have with the proxies used in temperature reconstructions. I have a more comprehensive view of the temperature reconstruction literature than that for the climate models and perhaps the answers to my questions about models have some answers of which I am not aware in the literature. I have found that these answers can depend very much on the overall view one has on these issues and whether the approach is one of proving the worth of the results of a temperature reconstruction or climate model output or the more skeptical one of looking for discrepancies in the results.
HaroldW (Comment #133509)
“To make a “50,000-foot†evaluation of GCMs, though, you’re interested in the secular trend. I don’t think that the exclusion of the Arctic detracts from your ability to make such a comparison.”
Harold, in my view we should be most interested in how well the models capture that part of the temperature/climate change that can be attributed to AGW, and there primarily from GHGs, in the historical period and for the future. The question arising out of this proposition is how well can we separate the deterministic trend (and give it the proper shape) from the white/red noise and periodic events in these temperature series? While a secular and deterministic trend in these cases should be near the same I do not assume both are attributable to the same phenomenon.
I am becoming more and more convinced that looking at the climate model emergent properties of ECS and TCR is a more direct way of comparing models to models and model run to model run and models to observed with regards to the AGW/GHG component of warming. Nic Lewis has pointed to the improved observed data now available for estimating the observed ECS and TCR for comparison with that emergent from the climate models. Even so, I believe that the observed data has uncertainties due the white/red noise and periodic events.
In addition I am not so quick to embrace the ECS and TCR estimates coming out of climate models and ignoring the fact that some of these models do not emulate (excluding here timing issues) the observed white/red noise and period events well.
Mark+Bofill (Comment #133505)
“Regardless of the question of masking & whether or not it’s done, I think there’s still value in C&W’s approach. Judith Curry wasn’t impressed. People who are dead set on disliking Cowtan & Way might look there for solid ground to stand on I guess.”
There is the dilemma.
They have an approach to an area previously ignored or poorly done. They put up a method which inherently should work “to give a better estimate than otherwise”. Zeke likes it. I should like it.
My reasons for being “dead set on disliking Cowtan & Way” are
-They both belonged to the Skeptical Science crowd before they put out their papers.
-The papers were done at a time of needing to find a reason to remove the pause and the conclusion they drew fitted these aims perfectly.
-Mr Cowtan wrote a piece for Skeptical Science arguing that he could get more value from linking areas far apart than areas close together. [faulty scientific reasoning].
They also excluded 3 ? Soviet island data areas because they were too warm [biased use of data science].
Their results always went up on preexisting data and their current data at the time [Lack of some negative results indicates a biased warming algorithm or a confected data set and the data set was not confected by them].
Luckily, though set in stone, an algorithm based on finding increased Arctic warming with small increments in Arctic temperatures has a fatal flaw. When the temps do decrease by any significant amount C and W will go through the floor.
It is happening now and will only get a lot worse if there is one more increase in sea ice left in the Arctic in this current rise of 6 out of the last 7 years.
Of course there will be a review of their algorithm to readjust it.
I should read Curry as well for her reasons, these are mine and the personal two at the start should not count scientifically.
Angech,
It’s refreshing to hear somebody else be forthright about their biases, I applaud you for this. FWIW, I approached C&W with a greater expectation of finding a problem than normal because of this association. ~shrug~ IMO, not everybody affiliated with SkS is a villain. Just Dana. 🙂
I’m not conversant with either of these issues. Maybe I failed due dilligence.
Look, I don’t think anybody believes this method to be ideal. The real answer is to get a good sampling of measurements geographically distributed with good coverage over the Arctic. [edit, but not with my tax dollars. :p] That’s not disputed. Of course there are problems with not having them. It certainly is flawed to work that way without them. I’d expect real thorough measurements to get a better result.
The thing is, we don’t have the comprehensive well distributed good measurements, and in the face of that what do we do. I still think C&W helped with this; what they did is better than any alternative I can come up with off the top of my head.
Mark re ” I’m not conversant with either of these issues.”
How global warming broke the thermometer record
Posted on 25 April 2014 by Kevin C at Skeptical Science
Otherwise a good article to read.
“We found that when the HadCRUT4 data are extended to cover the whole globe, some of the apparent slowdown in global warming over the past 16 years disappears.
*The original motivation for the project came from the fact that different versions of the temperature record were showing substantially different short term trends. We assumed that in addressing the coverage bias in HadCRUT4 we would bring it into agreement with the GISTEMP record from NASA. *But what we actually found was a surprise* – our infilled record showed rather faster warming than GISTEMP.The GISTEMP conundrum
. When we looked at a map of differences between the GISTEMP trends and ours the main differences were in the Arctic. And the differences in this one small region were big enough to explain about two thirds of the difference in trend between our results and GISTEMP.
Other data sources were investigated, including the new Berkeley land-ocean temperature data, the MERRA weather model reanalysis, and satellite radiometer datasets from AIRS and AVHRR. All of them showed faster warming in the Arctic than GISTEMP, particularly in the *Barents and Kara seas off the Russian coast, and in the *Beaufort sea off of the Canadian coast
What was causing the differences? The difference between our series and GISTEMP had a much simpler explanation: it came from the input weather station data.
The weather station data for NASA’s GISTEMP come from the Global Historical Climate Network (GHCN-monthly version 3). The station data in our reconstruction comes from the Climatic Research Unit (CRU) data used in the Met Office record. If our version of the GISTEMP algorithm is applied to the GHCN station data, the resulting Arctic temperatures are a good match for GISTEMP. But if the CRU data are fed into the same calculation we see faster warming – and the pattern of warming is very similar to our own reconstruction.
How do the station records differ? two things stand out. Firstly, the CRU data include a lot more stations – about 4 times as many observations as GHCN. Secondly, the Arctic island stations which are present in GHCN show much less warming that their more numerous counterparts in the CRU data.
How can the observations differ? The CRU data are mostly collected from regional weather organizations, and in some cases corrections are made by the provider or by CRU to account for changes in station location or equipment. The GHCN data are subjected to an automated homogenization algorithm which has been shown to be very capable at resolving individual station errors in the continental US record (Williams et al 2012). Could the adjustments, which were designed to fix problems in the data, be responsible?
The GHCN raw and adjusted records for selected Arctic stations were compared with both the Berkeley Earth records and the MERRA weather model.
*Two Russian stations and one *American station were identified where the GHCN adjustments were contradicted both by MERRA and by other nearby stations not present in GHCN. while some of the homogenizations appear to be correct, it is unlikely that so many widely dispersed stations would require similar adjustments in such a short period. The more extensive Arctic station inventories of Berkeley and CRU and the reanalysis and satellite datasets also support this conclusion.
“We are not the first to suggest that the GHCN corrections might be introducing errors into some station records However when we use the GHCN data in a global temperature reconstruction, we find that the adjustments have been suppressing the warming signal over the past decade.”
So they scrubbed the islands and warmed the Antarctic.
***** “The GHCN adjustment algorithm makes the conservative assumption that neighboring stations should show similar trends.****
As a result it performs well when presented with discontinuities in individual station records in well sampled regions, however we might expect it to have problems when faced with the sparsely sampled and rapidly warming Arctic”
It did not fit with established scientific technique [ Mosher called it a basic fundamental scientific truth somewhere but he does not insist on it either since this article ]
so they dumped the science.
These two steps are enough to make one distrust the motives of the people involved.
Hi Lucia,
I know you have a Cuban heritage. Don’t know if you think this appropriate for your blog, but I would be interested in your take on Obama’s new Cuban policy. I don’t feel that I know enough to have a strong opinion. Would be interested in yours, if you wish to give it on this blog.
JD
JD Ohio,
Oddly, I haven’t read up on it a lot. My main reaction was: “That’s sudden”! Usually you hear a lot of debate and discussion over these things before they happen.
December 16, 2014 at Arctic Sea Ice Blog
“Finally, everyone here will be pleased to know that I caught an Arctic specialist carelessly using the r-word yesterday and got prompt agreement from him (with a couple colleagues listening in) that using it is bad, bad, bad.â€
Thanks Angech.
The r-word? I presume that’s some variation on recover. Talk about standing knee deep in a big river in North Africa (to avoid using the d-word).
angech (Comment #133514)
As I recall the CW Infilled HadCRUT4 global mean temperature makes a difference (and I am not sure it was significant) in trends measured over the past 15 years between their temperature data set and that of the other 3 more utilized data sets. GISS which extrapolates over the Arctic did not show a great deal less warming than CW Infilled.
CW is greatly affected by the Arctic polar amplification and more generally the more extreme historical changes in temperature trends in that region of the globe. That means that if one uses the last 35 years in comparing the global model and observed trends those extremes tend to cancel out and over that period of time the CW Infilled gives much the same trends as the other three data sets. As it turns out it is easier to show a significant difference in the CMIP5 climate model mean global temperature trends and that from the observed series over 35 years than 15 years.
There are lessons to be learned in all this with the first being that given the noise in the signals a 15 year long temperature series is not adequate for showing a significant difference in trends of models to observed. The second is that CW made a big deal out of their fluctuating-Arctic-affected 15 year trend being greater than supposed and that the pause was than less than supposed while if one looks deeper into the reason for the differences one continues to see a significant difference in model and observed trends. The third lesson is that I judge that kriging is a valid method for infilling missing area coverage as long as one can at least attempt to put some confidence intervals on the use of that method. The fourth is that any new methods of looking at the observed temperatures should be encourage and is something from which we can learn. My view of this comes from my judgment that the observed historical instrumental temperatures and the adjustments and estimating confidence intervals thereof are a work in progress.
JD+Ohio (Comment #133515)
You did not ask my opinion but that has never stopped me from commenting here.
As a practical matter I think the Cuban policy of creating economic stress on Cuba in hopes of bringing down the Communist regime has failed. The only way the regime currently in power will be removed is when the majority of Cubans in Cuba realize there is better way. Economic hardship has not evidently swayed the Cuban people, to date at least, to risk resisting an admittedly despotic regime. What travel and trade with Cuba might reveal to the Cubans in Cuba could have a favorable impression on them with regards to the benefits of a freer society, but that would require some loosening of the current Cuban regimes restrictions on its citizens and I doubt whether that is something the regime would to voluntarily or that the US government will push.
Interesting contradiction in the US use of embargoes on Russia, and Iran, for that matter, in hopes of changing their governments ways and giving up on that with regards to Cuba. The biggest difference is probably that the rest of the world did not go along with the impositions on Cuba, but it also might mean that these restrictions do not work in general and particularly when aimed at governments that have/allow less freedoms for its citizens. The Iranian regime has endured some tough economic times and Putin has never been more popular in Russia – at least for the time being.
“My reasons for being “dead set on disliking Cowtan & Way†are
-They both belonged to the Skeptical Science crowd before they put out their papers.”
thats some good math there!
-The papers were done at a time of needing to find a reason to remove the pause and the conclusion they drew fitted these aims perfectly.
The only problem is this.
1. CRU and GISS recognized long ago that they are underestimating artic warming. secretly of course in the climategate mails.
2. Independently of C&W we found the same thing.
Humans will always have motivations. The simple way to DEMONSTRATE that there motives sully their results is to
do you own estimate of the arctic and demonstrate that it is superior.
I dislike Cowtan and Way because I think Robert Way is an incredibly untrustworthy source. He repeatedly recognized points made by Steve McIntyre and others were correct in private while refusing to correct misinformation from his own Skeptical Science group in public. I have no reason to believe he would behave differently with his own work. His behavior makes me suspect he would knowingly hide problems with his own work if he thought he could get away with it. That, of course, does not mean their results are wrong. It just means I won’t give them any benefit of the doubt.
I actually started to look into their code to examine their methodology. I quit because I wasn’t interested enough to deal with how they structured their code. It was enough of a hassle to get running I decided I didn’t care enough to bother. I’ll wait until someone else examines and discusses their work in a way which gives me confidence their results are good (or bad).
On the issue of Cuban policy, I don’t think anyone can really say whether it’s good or bad. There’s so little information as to what the new policy will be. In theory, the wishful thinking I’ve heard from the White House could be great. The trick is translating that wishful thinking into effective policy – something this administration has seemingly failed at in every other case.
Brandon
I agree. I read the US will relas travel restrictions . But.. well… if someone goes to set up communication satellites, will Raul Castro still arrest them? Yeah, he returned that other guy. But does that mean he won’t arrest anyone else. If Cuba doesn’t permit people to do things like build communications systems once they arrive, the fact they aren’t violating US law isn’t very meaningful.
So, Robert is his brother’s keeper?
Who’s my brother? I haven’t been keeping track of that bastage, whoever he is…
😉 I’m bored Brandon, that was my primary motivation for that remark. Yes, there are several things wrong with my idea. Well, at least a couple I think.
I don’t think Cowtan and Way makes that huge of a difference actually. Even the difference between approaches I favor (e.g., using Merra to infill) and ones I don’t aren’t that large:
link
You can see that HadCRUT clearly underestimates the trends, but overall picture isn’t that different, unless you cherry pick the interval, as Cowtan and Way did, and then use a meaningless comparison to further exaggerate the importance of the effect. From their 2013 paper:
As I’ve said previously (originally on WUWT I believe):
What is important is the difference between the trends HadCRUT and HadCRUT + interpolating series, and then “how important it is” is graded by how large you’d expect the natural variability to be.
Using multiplicative comparisons on trends that can have either sign, as I’ve said many times now, is just totally crackers. You can get results like “infinitely times better”, when the result you’re comparing to is zero, “infinitely times worse” when your own result is zero, and so on.
I agree with Steven Mosher’s point that it’s better to do your own series, but eventually science needs to stand on its own merits.
If we find multiple problems with the analysis, methods for verification and validation, and exaggerations and errors in the way in which the significance of the work is tested, etc., then I think we have a good reason to discount this work, even if we choose to not go back and repeat what they’ve done, for them.
Mark Bofill, when Robert Way decides to publicly act as a member of a group his brother leads, then yes, he is his brother’s keeper. He’s also the keeper of every other member of his group. That’s the nature of being a part of a group. When you join a group, you get painted with the same brush as the rest of the group. If the group does something wrong, and you say you know it is wrong in private, then of course people will criticize you for not saying the same thing in public (assuming the private communication doesn’t effect change, of course).
Carrick, what you just said reminded me of something:
When Steve McIntyre began criticizing Michael Mann’s work, one of the most common responses was, “Well, why don’t you make a reconstruction of your own?” It’s interesting to compare that to what Steven Mosher says.
We lived in Miami, except for a very much mistaken move to Philadelphia of 18 months, for 14 years. During five of them, I ate lunch daily with guys who had grown up in Cuba, whose family property was seized by the Revolution in 1961, and who fled to the US early on, some via Puerto Rico.
To a man, these guys believed that Fidel wanted the embargo and the hostile relations with the US maybe more than some of the people here did. Their evidence was that every time there was a threat of a thaw, (regional not global warming), he would shoot up a boatload of refugees or do something else to make the US revisit its thoughts of accommodation.
They believed that so long as Fidel ran the place, there would never be a thaw. He needed to paint us as the bad guys, so they thought, to justify the many failures of his regime to provide a modern standard of living beyond his apparently respectable health care arrangements.
We share this country with some truly remarkable people who weren’t born here. One in particular had gotten a civil engineering degree in Havana, worked on their nuclear plant (never finished, I think) and had come to the US in a boat he had built powered by a Russian cement mixer motor with a long shaft with home-made propeller welded to it. it worked and he arrived in Key West with it and was accepted – it required feet on ground at that point. He had no English. 6 years later he had good working English, was a civil engineer at the A&E I was working at, and asked me one day to give him a list of books in English which I would recommend. I did, maybe 20. He had already read a quarter of them, including Moby Dick. I asked to see his list. It was in the hundreds.
My experience with immigrants has been very good, educated and uneducated. They are often very impressive people especially the ones who got here via challenging circumstances.
I know a few more stories similar to the above, but it seemed the best.
As an aside, anyone else familiar with Bodeguita del Medio, leave me a note.
[Edit:]Brandon,[:endEdit]
Your comment reminded me of this and I had to go back looking. here.
So he doesn’t publicly repudiate his tribe. He does privately argue the point.
Some claim the true test of a man’s character is what he does when no one is watching.
On a side note, LOL. Did I say the SkS guys weren’t villains? I’d forgotten this shenanigan. Maybe they aren’t villains, but they certainly covered themselves with dogcrap in that instance.
Brandon,
I don’t accept this. I consider myself a member of WUWT. I certainly don’t take responsibility for what each and every one of the peanut gallery over there put forward. Nor do I consider it my responsibility to repudiate it all.
Hey, to tie up a loose end-
What I thought the problem was with what I was arguing was that it (my argument) was a dodge. The thing is, it does demonstrate a lack of courage and integrity to keep one’s mouth shut about errors made by one’s tribe. Robert apparently did this. I’ve done it. I’d like to believe that doesn’t make me scum, but maybe I kid myself; perhaps my perspective would be exactly like yours if I wasn’t personally guilty of doing this.
~shrug~ I’m sure I’ll live, regardless. 🙂
Mark Bofill:
Yup. But to be fair, there is a matter of degree involved. The more you’re associated with a person, the more you are associated with them. Being in the same group makes you associated, but only to a certain degree. We’re all humans (at least for the sake of argument). In that regard, we’re all in the same group. When some people commit genocide, we are all tarred with that action.
We all look bad when huge and terrible things happen and we do nothing, but we don’t look as bad when one person robs a store. The reason is degree. There are different numbers of people involved, and there is a difference in severity. You as a human may look worse because of vandalism in my neighborhood, but only by an infinitesimal amount.
Without knowing the details, I can’t say. What I can say is Robert Way thinks global warming is an important issue. He is part of a group which advocates for action to combat it. He thought points raised by people like Steve McIntyre were important enough to merit discussion. He thought those points were important enough to call for changes to be made because of them. But he only thought that in private.
Perhaps my perspective would be exactly like yours if I was part of any tribe or group where I’d have a reason to keep my mouth shut. It’s pretty easy not to do something when you never have the opportunity to do it 😛
For those of you talking about climate sensitivity in the models, I would draw your attention to what I think is an extremely important paper recently published by Andrews et al:-
http://journals.ametsoc.org/doi/abs/10.1175/JCLI-D-14-00545.1
It confirms the non-linear relationship between net flux( heat input) and temperature gain displayed by almost all GCMs under a constant forcing, and identifies the root cause of the curvature as SW cloud radiative effect in most of the models – probably the least certain piece of physics in the GCMs. Whaddya know.
As I have shown before, this curvature in the net flux-temp relationship (aka time-dependent climate sensitivity) is one of the major contributors to high estimates of ECS from the models – along with the two other main causes correctly identified by SteveF in one of his posts above – underestimation of total forcing by aerosol offsets and overestimation of warming (temp and ocean heat gain) relative to observations.
The effect of this curvature is to “separate” the parameters applicable to the predicted runs from the parameters actually manifest in the historic runs of the GCMs, the latter implying a higher feedback and a lower climate sensitivity.
The errors arising from poor methodology from the Gregory/ Forster/Andrews school of forcings and feedbacks in recent years however have given rise to estimates of the magnitude of the total feedback in the GCMs during the historic period which are out by a factor of 2 in some cases.
Paul_K,
Since clouds can’t be modeled with the coarse grid size used in AOGCM’s, even if it was known how to do it, I presume that cloud feedback parameters are prescribed, not an emergent property.
A non-paywalled link would also be nice.
I presume the mechanism is that clouds in the models have a net positive feedback and that tropical cloud cover increases as the sea surface temperature increases.
Paul_K,
“Quantifying components of aerosol-cloud-radiation interactions in climate models”, authors including Andrews & Forster, is also of interest. From the abstract:
Paul_K,
Nice to have you back.
.
Interesting that 23 of 27 models have the same behavior on what is for sure a strictly parameterized behavior. Maybe the models are less independent of each other than some have suggested. Maybe too it is related to the consistency of model projections of mid-tropospheric warming that isn’t found in the observational data.
.
The abstract says that most of the modeled increase in sensitivity is in tropical cloud cover, which I presume means that at higher tropical temperatures the models have decreasing tropical cloud cover compared to lower temperatures. Which would be consistent with the models’ projection of less increase in rainfall than increased absolute humidity (Clausius–Clapeyron) would suggest. This behavior has already been called into question with observational data (How Much More Rain Will Global Warming Bring? Wentz et al, Science, 2007).
.
In addition, it is well known that warming between ice age and interglacial periods over the past half million years or so was far less in the tropics than at high latitudes, and that very warm geological periods (millions of years ago), with much higher CO2 levels than today, had even smaller tropics-to-poles temperature differences. If increasing climate sensitivity at higher temperatures is related to tropical cloud cover, then it would seem reasonable to expect MORE warming of the tropics when the global average temperature was higher, not less. So either cloud cover does not change with warming as the models project, or tropical clouds have a different net influence on warming than the models project.
.
Perhaps existing CERES data could shed some light (sun light?) on whether clouds in the tropics respond to temperature increases as the models suggest. I will go way out on a limb and guess that CERES data will not support the modeled behavior.
.
On a related note, I see a repeating pattern in all the model/reality comparisons, excuses, kludges, and arm waves: when the models overstate observed warming, the response of the modeling community is always to “explain” the discrepancy in a way which indicates that while current warming is modest, future warming will be much greater, even… well, catastrophic. It is sort of like the old joke about practical fusion power… a global warming catastrophe is only 20 years away… and always will be. 😉
“Maybe the models are less independent of each other than some have suggested.”
Pennell & Reichler reported that of 24 CMIP3 models, the mutual correlations were such that the effective number of independent models was about 8. I don’t think I’ve seen a similar analysis of CMIP5.
“When Steve McIntyre began criticizing Michael Mann’s work, one of the most common responses was, “Well, why don’t you make a reconstruction of your own?†It’s interesting to compare that to what Steven Mosher says.”
yes, since I’ve suggested to Steve Mc the same thing , I’d say I’ve been pretty consistent in that regard.
And yes, I was there cheering on JeffId and RomanM when they did their work.. and yes I cheered the work that O’donnel and Mc did to estimate antartica, oh ya and the work Niic Lewis did to estimate sensitivity
In the end you cant get anywhere if all you do is criticize.
It would be nice if there were a professional class for critics.
and a journal of No. but there isnt.
In the end, The contributions Nic lewis made ( by doing something) the contributions odonnel and Mc made, the contributions of troy masters.. ect.. the contributions of doing something vastly outweigh snipping at Tols heels or Mann’s heels.
Steven “In the end you cant get anywhere if all you do is criticize”.
But we do do lots of other things, living for example, working, playing sleeping, mixing with friends and family.
“do you own estimate of the arctic and demonstrate that it is superior Well, why don’t you make a reconstruction of your own?â€
Put me in your chair, give me your resources and in 3 months I could have some darn good reconstructions out there.
In particular I would do some of your models with a climate sensitivity of 1 and a climate sensitivity of zero. Care to guess which one would model the observed values best with the pause?
Or have you already done it?
If not, why not?
“thats some good math there!”
The silence from all the modelers is deafening, after all if one puts in CS zero or 1 it could not possibly work, could it?
The critic is never well received. Basically we see someone sit back and try to shoot our arguments to threads. I hate critics too. But that is science. It should be critic proof.
Not foolproof, which is what I think you mean.
Foolish arguments, like some of mine, deserve your put down though some fools are well meaning and would appreciate milder correction.
Critical arguments, addressing important points scientifically should not be dismissed because you do not like them.
“My reasons for being “dead set on disliking Cowtan & Way†are
-They both belonged to the Skeptical Science crowd before they put out their papers.â€
“thats some good math there!” Brandon said “When you join a group, you get painted with the same brush as the rest of the group” QED One and one equals two.
I did not say that estimating the Arctic by Kriging was bad, I pointed out 3 mathematically valid points against their version, the most important being your own principle which they diss, the dropping of 3 warm stations and the absolute invariance of any dissenting results in their work [the too perfect Peter principle].
Point out where those three are wrong and I will listen to you.
If Mike Mann was to make a point here and he was right I would listen to that to. That will not stop me believing that he has a bias in his approach to science and that it would make me more cautious in listening to any of his arguments.
Steven Mosher:
It’s good to know Steve McIntyre has accomplished nothing as after all, if you don’t have an answer, you’re clearly in no position to criticize a bad answer. Everyone knows it’s better to have any answer, no matter how bad, than to say, “We don’t know.”
Brandon,
You are kidding, aren’t you? I cannot see that there is anything wrong with saying “That’s not an answer” without having your own answer. Maybe you do have to be able state what an answer should include.
Steven Mosher says,
Yeah, that’s so. It’s just not important enough to me to risk / invest real time and effort. I do the blog stuff because it’s fun and interesting and free. But if I was looking for another job I’m selfish enough that I probably wouldn’t try to make a contribution to our understanding of climate, I’d most likely try to kickstart a tech company of some sort. Something like that.
But that’s correct, in the end I’m a spectator and not a contributor. I’m good with that. I don’t need a seat at the table for the science debate anyway. The policy debate is a different matter.
Steven,
So if I were to find a flaw in a mathematical proof, I couldn’t publish that unless I had a valid proof of my own? Oh, puhleeze. Of course it was difficult to falsify MBH9x because the data and the code weren’t readily available.
Also, how can you rationalize creating a temperature reconstruction from tree ring data if you don’t believe that the temperature signal can actually be detected in the data? Which is something that can’t be proved.
j ferguson, yes, I was kidding. The idea was to state the position in a serious manner so as to encourage people to realize for themselves how absurd it is rather than just tell them.
The point is when Steve McIntyre began examining Michael Mann’s work, many people tried to dismiss him by saying things like, “Well, why don’t you make a reconstruction of your own?†It was absurd. Whether or not Mann’s conclusions were unsupported by his work in no way depended upon McIntyre being able to produce alternative results of his own.
The same is true across the board. Discussing problems with results does not require one provide alternative results. It is nice if one can, but it is in no way necessary.
This is akin to a prosecutor saying the defendant must be guilty because the defense can’t tell you who the real killer is. Defense lawyers do prefer to be able to offer an alternative suspect, but being unable to doesn’t mean we should ignore the fact the defendant was on live television in another country at the time of the murder.
DeWitt:
That’s a good point.
Another way of saying it, is the problem is difficult enough of accurately extracting temperature from proxies, that McIntyre isn’t going to be able to solve it by himself.
Criticizing the failed attempts at temperature reconstruction by MBH and others, when it has been erroneously touted as an accurate reconstruct, certainly helps push the field forward.
Even if it makes McIntyre vastly unpopular with people like Mann, who are obviously indifferent to producing accurate reconstructions, or with people who don’t mind making false narratives as part of their climate change advocacy.
You can hardly require McIntyre to form a new paleoclimate research group and engage in a 10-year effort, before he is allowed to point to major errors in existing research (especially bad research that was heavily touted by the IPCC).
The claim that one must be able to do the work to criticize it is simply nuts. I don’t need to market a line of herbal remedies to call their efficacy into question. If I know math and see that someone is using it all wrong (like Steve Mc) that is a sufficient and valid point to make. If I understand tree growth (and I do) it is sufficient to point out that temperature response is not linear, nor stable over time.
Craig, here’s an extreme example:
I don’t need to manufacture my own brand of snake oil that actually works before I can criticize the claims of snake oil salesmen.
Re: Mosher Claim that you have to do the work to criticize.
The fundamental flaw in his argument is that it assumes substantial validity for the work being criticized. I don’t have to be an astrologist or perform “astrology” work to criticize astrology. In my view, the Hansenite wing of climate “science” has failed to show that their work is substantially justified. (See Freeman Dyson comments, for instance.)
In fact, I object (in a minor way that I don’t bother to raise) to being called a skeptic because the research of the warmists is not solid enough to be used as a reference point for my position.
JD
JD,
To be completely fair to Mosher, I believe his point may be valid for critics of the methods for estimating the global temperature field from relatively sparse measurements. For treemometers, not so much.
DP — I think anyone has the right to make rational criticisms. Someone who has done the work, may have more credibility, but everyone has the right to criticize. Also, since most warmists immediately jump from their temperature findings to public policy matters relating to reducing CO2 and increasing or redistributing taxes, all taxpayers have every right to criticize the Hansen wing of climate “science.”
JD
DeWitt,
I agree that there is a difference between criticizing how sparse temperature data is used to form a temperature history and criticizing nonsensical rubbish like Foster and Rahmstorf’s 6-parameter curve fit ‘proving’ the pause is nonexistent. After all, there certainly was a temperature history, even if generating an accurate representation presents challanges, so in theory at least it is possible to generate an accurate reconstruction.
.
Still, there is no obligation to offer an alternative method if you can show that an existing method is simply wrong. It is a ‘greater contribution’ if an alternative (correct!) method is offered along with a critique, but I don’t see that is required.
FWIW, Isaac Held has an interesting blog post about reconciling measured tropospheric warming and surface temperature change. Two things appear involved: an ‘optimized’ (modeled) tropical ocean temperature history and generating a fit of the measured tropospheric temperature to that ocean temperature history using a non-least-squares method that reduces the contribution of ‘outliers’. I am not sure what to make of Held’s post, but it seems a bit ‘arm-wavey’ to me.
DeWitt Payne:
I was curious if he would argue there was a difference in the two cases that would change things. He apparently doesn’t think so.
I can think of nothing more pointless than arguing this. Nothing less useful and darn few things that are less interesting. For some insane reason then, let’s proceed.
Where’d Steven say you had to do the work to criticize? That’s not what he said. More importantly, I don’t think that’s what he meant, either.
Looks like a big mess of straw to me.
‘In the end you can’t get anywhere’, and ‘the contributions of doing something vastly outweigh snipping at heels’. Well, fair enough. Criticism in and of itself doesn’t get you anywhere, and sure, often there is something more meaningful to be done. I don’t see why that’s outrageous.
I’m not going to try to make the point I thought Steven was making, as I didn’t particularly care about that in the first place. But it just never ceases to bemuse me the way it seems that Steven can share a few words and evoke unanimous disagreement, regardless of where everyone is on the climate spectrum. It’s a spectacle!
🙂
Mark Bofill:
But you’re arguing it anyway!? 😛
I figure Mosh meant to provoke more than trying to argue a particular point, but I’m not a mindreader.
Anyway, I disagree with the premise in toto. A great Sherlock Holmes quote is:
“… when you have eliminated the impossible, whatever remains, however improbable, must be the truth.”
Eliminating bad arguments, bad models or bad theories is a perfectly acceptable way to make progress. And yes it is progress.
Not everybody needs to build models, or software packages, or their own theories. If they did, science would become another tower of babel.
Somebody needs to be left to critically examine new ideas as they come into the field.
Mark Bofill:
You’ll note I didn’t offer any interpretation of Mosher’s argument at first. I just said:
Mosher responded by saying:
Mosher said he endorsed the dismissive criticisms of Steve McIntyre which were based entirely upon McIntyre not providing a reconstruction of his own. It was only at this point, after Mosher specifically said he has “suggested to Steve Mc the same” thing everyone is scoffing at, that I scoffed at him.
I created an opportunity for Mosher to state his position. He did. People scoffed at the position he stated. I’m not sure why you think there is a strawman involved.
Carrick,
Ole! or bravo! Touche! Or whatever is appropriate when somebody gives a response I agree with substantially that decimates my argument.
Brandon,
That’d check out except that I don’t buy it. :p I don’t buy that you believe that that’s what he was saying. Or rather, I buy that you believe that that’s what he said, but I don’t buy that you believe that’s what he meant.
Now I hope that you accept that I don’t buy that you believe that that was what he meant. Because if you don’t buy that I don’t buy that that’s what you believe about what he meant, then when and if Steven or for that matter anybody else picks this up, we’ll be four ply in the front just sorting through the case confusion, trying to sort out what they may or may not believe about your belief or disbelief about my disbelief about your belief that that’s what he meant. And at that point I’ll have to write code to keep track of what’s going on, or run an increasing risk of corrupting the heap the further we go. And I don’t want to do that. It’d ruin Christmas.
[edit: block replace of ‘meant’ for ‘was saying’]
MB “I can think of nothing more pointless than arguing this.”
This comment is most appropriately addressed to Mosher. He repeatedly makes this argument and related arguments on different blogs. Several months ago, he made this argument on Curry’s blog and suggested that no intellectual giants had criticized Hansenite “science” without actually doing it. I pointed out that Freeman Dyson had, which slowed him down a little.
On another occasion, making an argument in a similar vein, Mosher argued that insurance companies couldn’t and wouldn’t fudge their risks to increase profits (for instance, he believed they wouldn’t exaggerate risks of global warming because they had competitors.) Mosher claimed that if anyone disagreed with his position, they were obligated to form their own insurance company to take advantage of higher rates quoted by the inefficient insurance companies that were exaggerating risks. I, having seen the behavior he believed wasn’t plausible, didn’t even bother to respond to Mosher’s argument.
JD
JD,
Ah! I should have suspected.
Thanks. I was unaware that there was a prior basis for the response. Makes more sense given that detail. 🙂
[Edit: So that’s a reasonable answer. Where did Steven say that you have to do the work to criticize it? Apparently over at Climate Etc. got it now, fair enoogh]
Brandon,
Okay, in light of JD’s remark, I lose my basis for believing that you don’t believe that Steven believes that.
It all comes crashing down at this point.
…
Fine! In the immortal words of Eric Cartman, ‘Screw you guys I’m going home!’
🙂 Happy Solstice all.
Mark Bofill, leaving aside my past experiences with Steven Mosher where he made the exact same sort of argument, why would you think I don’t believe Mosher meant what he said?
Mosher often discusses his ability with and knowledge of the English language as though it is impressive. Expecting him to be able to write simple sentences seems reasonable.
But yeah, like JD Ohio, I’ve seen Mosher make this argument a number of times. I’ve had him even say it directly to me, telling me my opinions about BEST don’t matter unless I create my own temperature data set. It’s nothing new.
Cause I have a difficult time believing that’s what Steven Mosher meant. Because that’s stupid [edit, that would be a stupid thing to mean], and Steven doesn’t strike me as a stupid guy.
~shrug~
One of the main criticisms of the tree rings is whether the trend can even be extracted at all with any confidence. The raw data is a bit of a mess. It’s not a direct temperature measurement and it is confounded by numerous other parameters that are near impossible to unscramble with any confidence. The divergence problem is case in point.
The criticism is that you cannot get a useful reconstruction from this data, period. The output is overly dependent on the data processing methods. Doing another random reconstruction and proclaiming your randomly different model produced a better “truth” adds just about nothing useful.
You can process the instrumental temperature data in eleventeen different ways and you will get an increasing trend for the last 150 years with all eleventeen of them. Because the raw data has useful information in it that is easy to extract and it is near impossible to torture that data to remove that trend.
The criticism that critics must somehow extract the correct “truth” from tree rings in order to be taken seriously is not valid because the real truth is likely not able to be determined with the information provided. That is the point of the criticism. GIGO. There is no such thing as garbage in, truth out.
DeWitt, SteveF, HaroldW
(DeWitt, sorry I can’t post a non-paywalled reference.)
(HaroldW, thanks for the reference, which seems like a simple factual analysis of components. Am I missing something more important?)
From the discussion part of the Andrews et al 2014 document:
The paper highlights evolving patterns of sea-surface temperature as the main culprit responsible for the curvature in net flux. The mechanism is via the evolving change in cloud radiative feedback, and particularly the SW in net terms. A comparison of early (1-20 years) vs late (21-150 years) feedbacks shows a positive change in LW Cloud Radiative Effect (CRE) in the tropical Pacific (increasing positive feedback) which is largely offset by other negative changes in the Indian and Atlantic oceans, leaving the aggregate LW CRE flux still close to linear against global temperature in the CMIP5 suite as a whole. A similar comparison of SW CRE shows an increasingly negative feedback response in the western tropical Pacific, but this is more than swamped by increasing positive feedbacks in other parts of the tropics and extratropics (around 30 degrees N and S). Some less dominant additional positive support comes from the southern oceans plus further “help” from the N Atlantic and Arctic regions.
The authors conclude that one cannot estimate ECS from historical data. Mmm, given that the regional developments highlighted in teh CMIP5 models bear little relationship to observed warming patterns, and especially the historic SW evolution, there is of course an alternative conclusion which can be drawn from these data. Nevertheless, it is a very valuable study. I recall expressing surprise that such a study had not been done 3 years ago. They obviously read Lucia!
Paul_K,
Thanks for the additional information.
.
I tend to be less than very impressed by post-facto explanations, and this seems to me to be a clear example. As I noted earlier, there is overwhelming evidence that average warming is dominated by warming at mid to high latitudes, not in the tropics, and this is true over a huge range of climate conditions, from ‘recent’ ice ages/interglacials to long ago periods with warm temperatures all the way to the arctic circle. Increased sensitivity in the tropics over time seems contrary to this.
.
A study of model behavior like this one would be more convincing if there were evidence that models reasonably simulate reality… you know, like clear data showing that the atmosphere has responded to current GHG forcing (which is now well over 3 Watts/M^2) with the cloud changes that the models post-dict. The current forcing has been reached over 100 years (although the majority has been over the past 65 years). With that long a period to respond, the “1-20” year response is only a modest part of the total, and there should be evidence of the much longer (21-150 year) response already visible. I don’t think the current pattern of warming is consistent with the longer term response the article ‘predicts’. Of course, the modelers can (and do) low-ball GHG forcing data with arbitrary aerosol offsets, and will claim that net GHG forcing has been low enough over the past 100+ years to be consistent with the models’ high sensitivity. Intellectual corruption at its finest.
This kind of study needs to make meaningful, testable projections of behavior which can be later compared to data which does not yet exist. I won’t hold my breath, since most work related to climate models is either done post-facto, or in a way which can’t be tested within several decades. As they say, making predictions is hard…. but that is what modelers need to do to establish greater credibility. I do not think they can do it.
Paul_K,
The main problem is that models have clouds as a positive feedback. I question whether there is evidence in the real world for that assertion. A quick and dirty MODTRAN calculation shows that low clouds only reduce emission at the TOA by a relatively small amount while reflecting a much larger amount of incoming solar radiation.
Tropical atmosphere default conditions (70km looking down):
clear sky: 289.3 W/m²
cumulus cloud: 266.0 W/m²
Reflecting even 20% of the incoming 300+ W/m² more than makes up for the reduced emission. The cloud bottom temperature controls the surface temperature, not the other way around.
Paul_K (Comment #133578)
Your post and those of others here on climate model treatment of Equilibrium Climate Sensitivity (ECS) and Transient Climate Response (TCR) are timely for me. I have a layperson’s interest in the difference between climate models and individual climate model runs. I have looked at the noise, cyclical and secular trend differences and wanted to look in detail at the ECS and TCR differences and the differences in input data from which these parameters emerge.
I do not have a comprehensive understanding of the factors affecting the estimation of RCS and TCR. I have read what Nic Lewis has reported on ECS and TCR, from AR5 and a paper titled: “Evaluating adjusted forcing and model spread for historical and
future scenarios in the CMIP5 generation of climate models”, by Piers M. Forster, Timothy Andrews, Peter Good, Jonathan M. Gregory, Lawrence S. Jackson, and Mark Zelinka and published Feb 2013 in JOURNAL OF GEOPHYSICAL RESEARCH: ATMOSPHERES, VOL. 118, 1139–1150, doi:10.1002/jgrd.50174, 2013.
In my first simplistic view of estimating ECS, I used the equation: F2XCO2*deltaTemperature/(deltaF-deltaQ) . I used the RCP4.5 model runs for temperature, TOA outgoing long wave radiation,
TOA incident short wave radiation and TOA outgoing short wave radiation for the period 2006-2100. I obtained delta values by determining the secular trend (not linear) for temperature and net TOA series for the years of 2006 and 2100. I used F2XCO2 =3.71 and deltaF for RCP4.5 of 2.3 W/m2 which includes the negative aerosol forcing. I obtained ECS values that varied considerably more than those published by Forster et al (2013) in the paper noted above but with close to the same model mean. I should note that to this point all the data I used was available at KNMI.
After reading Forster et al (2013) it appears that splitting the TOA change in the energy budget into F (forcing) and feedback portions is required with the feedback portion being temperature dependent. Forster then uses pulsed 4XCO2 model simulations and linear regression to obtain F and the feedback factor to, in turn, obtain an adjusted forcing (AF) which is unique to each individual model (run?) and each RCP scenario and historical data. AR5 reports the Effective Radiative Forcing for CMIP5 climate models (Forster’s AF) from Forster and from a method referred to as Fixed SST by Hansen. The methods agree with some models but can differ in others by 20-25%.
I am currently in the process registering so that I can have access to the 4XCO2 CMIP5 model simulation data. In the meantime I was wondering whether any posters here have any comments on the Forster method or that by Hansen.
Also from AR5 Chapter 9 I see an attempt by the authors to close the door on countervailing views of the model estimates of ECS.
“The large scale climatological information available has so far been insufficient to constrain model behaviour to a range tighter than CMIP3, at least on a global scale. Sanderson and Knutti (2012) sug-gest that much of the available and commonly used large scale obser¬vations have already been used to develop and evaluate models and are therefore of limited value to further constrain climate sensitivity or TCR. The assessed literature suggests that the range of climate sensitivities and transient responses covered by CMIP3/5 cannot be narrowed significantly by constraining the models with observations of the mean climate and variability, consistent with the difficulty of constraining the cloud feedbacks from observations (see Chapter 7). Studies based on PPE and CMIP3 support the conclusion that a credi¬ble representation of the mean climate and variability is very difficult to achieve with equilibrium climate sensitivities below 2°C … High climate sensitivity values above 5°C (in some cases above 10°C) are found in the PPE based on HadAM/HadCM3. Several recent studies find that such high values cannot be excluded based on climatological con¬straints …
From Sanderson & Knutti, “On the interpretation of constrained climate model ensembles” (2012):
I thought Mosher had recently commented that models were not tuned to climate history. Have I misunderstood Mosher or S&K?
HaroldW.
It may be only me, but I cannot understand the S&K sentence you quote. Exploited? Do they mean more than simply used? Is this maybe like “exhausted?” Misused?
Re: tuning. I had a similar response from Mosher a few years back which was contained in several paragraphs explaining that verification and validation were terms of art and didn’t mean what I had loosely thought them to mean. Tuning might be another one.
I came away from this chastened but still convinced that whatever the modelrs were building up with mathematical models or representations of physical processes nonetheless were at least adjusted should they stray far from experience. For the life of me, i don’t understand why these “adjustments” don’t constitute tuning.
if I haven’t gotten anything else out of the five or six years I’ve been reading this stuff, I’ve at least become nervous about my ability to communicate clearly and economically.
Nuts.
HaroldW (Comment #133582)
HaroldW that paper you linked on first glance sounds like it discusses some basic issues about modeling and observed comparisons that are interest to me. Thanks.
Does not this more complete excerpt below answer the question you posed? Bolded part if my HTML coding works.
” We show some ‘truth plus error’ like properties exist for historical and present day climate simulations in the CMIP archive, and that they can be explained by the ensemble design and tuning to observations, although both models and tuning are imperfect. For future projections, structural differences in model response arise which are independent of the present day state and thus the ‘indistinguishable’ interpretation is increasingly favored. Our inability to define performance metrics that identify ‘good’ and ‘bad’ models can be explained by the models having largely exploited the available observations. The remaining model error is largely structural and the observations are often uninformative to further reduce model biases or reduce the range of projections covered by the ensemble. The discussion here is motivated by the use of multi model ensembles in climate projections, but the arguments are generic to any situation where multiple different models constrained by observations are used to describe the same system.”
I think the tuning reference here would not necessarily require a tuning knob and parameter, but rather a reference to observed data to evaluate the model and rebuild or remodel if required.
j+ferguson,
I’ve also had that same feeling about my ability to communicate clearly.
My reading of “exploited the available observations” is that the modelers use those observations to tune the model parameters. Hence, all models — “good” and “bad” — match those observations reasonably well. To use the wording of a recent CA post, we need “out-of-sample” observations, and historical observations are “in-sample” and therefore aren’t useful in evaluating the validity of models.
[Edit: cross-posted with Kenneth’s above, but I think we’re saying the same thing.]
I don’t know about the strict definition of tuning wrt models, but it’s well known that a necessary condition for a model is that it hindcasts at least the last 100 years of the instrumental temperature anomaly with reasonable precision. Absolute temperature can be off by ± 2 degrees.
Christie and Spencer say the same sort of thing about the MSU derived temperatures, i.e. they’re not tuned to the surface data. But you can bet that if they didn’t match radiosonde data or other data to a reasonable approximation, adjustments would be made to the program. In fact, they have been. The biggest being the adjustment for satellite orbit drift.
I suspect that they mean that the monthly temperatures are not adjusted individually. But that doesn’t mean that differences between satellite and other forms of temperature measurement aren’t investigated and dealt with if a cause can be found. IMO, that is tuning.
Kenneth,
Here is a non-paywalled version of Forster et al:-
http://www.atmos.washington.edu/~mzelinka/Forster_etal_subm.pdf
It is a terrible paper IMO, taking us on a completely circular logic tour to finish with pseudo-parameter estimates which are meaningless (the AF values) or completely erroneous (the feedback values).
HaroldW, Maybe we’re in the out-of-sample right now.
One of the more pithy observations at CA in the last couple of weeks was, IIRC, if the grasp of the physics of some of the ‘systems’ modeled was so tentative, how could be the models be so good for the historic period?
Sniff anything?
Paul K (#133588)
The published version of that paper is available here.
PaulK, thanks for the comments on that paper – which I found linked below:
http://onlinelibrary.wiley.com/doi/10.1002/jgrd.50174/abstract
I am of the opinion that that paper gives the official or one of the official calculations of the CMIP5 model ECS and TCR values. What are the alternative methods? And which is the best/better in your judgment?
DeWitt+Payne (Comment #133587)
“I don’t know about the strict definition of tuning wrt models, but it’s well known that a necessary condition for a model is that it hindcasts at least the last 100 years of the instrumental temperature anomaly with reasonable precision.”
?
It seems to me that a model with past data inserted for reference is not hindcasting at all.
If you set it to reproduce the past it can only spit out the anomaly programmed into it.
” Absolute temperature can be off by ± 2 degrees”.
Do you mean the real temperature which is the anomaly is off the mean by up to 2 degrees. what sort of running mean would be used. Is there a yearly mean incorporating all of the data as it comes in and recasting the mean or is it restricted to a certain period eg 1979-2000.
angech,
There are a number of conceptual errors in your post above. The first is that you clearly don’t understand the concept of anomaly wrt temperature. The anomaly is calculated from a thirty year baseline average, like 1961-1990 inclusive. If you’re looking at monthly data, each month is averaged seperately. For example, January. The model global average temperature for the month of January for each year from 1961-1990 are averaged. That average is then subtracted from the model temperature for January of every year. That’s the anomaly. If there’s an absolute temperature difference between model and observations, that is lost when calculating the anomaly.
Observed temperatures are not ‘inserted’ into the models. Well, they used to be when models were forced with prescribed sea surface temperatures. But that is no longer the case. The model calculated anomalies for 1901-2000, say, are compared to the measured anomalies. If the disagreement is too large, the model parameters used are obviously no good and are adjusted until there is agreement. But, of course, it isn’t that simple. In fact, the parameters and parameterizations have evolved with the models such that hindcasting is almost always successful.
Kenneth+Fritsch (Comment #133585)
†We show some ‘truth plus error’ like properties exist for historical and present day climate simulations in the CMIP archive, and that they can be explained by the ensemble design and tuning to observations, although both models and tuning are imperfect.”
I would not expect any model to deviate in historical climate simulations or to be able to do a true hindcast over entered data time scales. All the past data is built in and the program should not be able to deviate from its information.
Tuning can only be done on “new ‘ information, that is data entered after the initial programming.
This is where new temperature adjustments to records with changing data sets can be entered.
The old data sets which the programs were set up with should remain theoretically inviolate. This may explain some of the early dips that now appear when the programs inbuilt climate sensitivity should project an ongoing upward temperature response. The early temperatures should not have been able to be adjusted downwards when adjustments to past temperatures have been made.
On the other hand provision may have been made to alter the past data and the absolute mean in the original programs in some of the models. This would constitute automatic tuning and would not need any assessment of current conditions, just inputing the new base records.
Of course this should show up as a completely different model year by year. Does anybody have one of these records, Steven or Zeke or Kenneth and could anyone publish a snapshot of one of these models for say 2000, 2005, 2010 to show the differences.
Alekhine said, I think if such a difference exists, search as hard as you can and you will find it [note the ‘if”].
“cannot be excluded”….meh.
Climate sensitivity of 10C “cannot be excluded”? Perhaps this statement should end with something like…”but this is extremely unlikely”?
Perhaps we would be interested also in what the other side of the range also cannot be excluded.
It is exactly this type of statement that is quote mined by activists and sometimes the more cynical side of me believes they are placed there for just that purpose.
“consistent with”….double meh.
“I thought Mosher had recently commented that models were not tuned to climate history. Have I misunderstood Mosher or S&K?”
It might help if I explain precisely what I mean by tuned.
Tuning is a systematic process where you vary uncertain input parameters to minimize the error between model outputs(M`) and real world observations (M).
For example.
You have a model that is
M` = f(a,b,c,d,e,f,g,h,i,j..) Where M` is the metric of interest ( there may be other outputs ) and {a,b,c..} are uncertain/free parameters that you can vary.
each parameter has a ‘range’ of what are considered to be possible values and you can construct your model to vary these parameters in a systematic fashion to effectively find those settings or range of settings which minimize the difference between the model output M` and the observations M.
A simple example from engineering. We had to create a high fidelity model of the F-5 flight control system with very little documentation ( we had left Northrop and could not get the documents) Well, we had system diagrams and we had flight test traces. The system diagrams told us what basic building blocks there were in the system ( the physics in a metaphorical sense) But what was missing were key specs on certain devices ( acceleromters for example ) So you coded up the model and threw in numbers for all the missing device specs. Then you matched it to the flight test.. And after a few weeks of this you could say that the accelerometer HAD to provide a measurement within x% of the true ‘G’s’ in order to match the flight test. This was proof that the original component delivered these specs even though no one had access to the actual spec sheets. Actually for some folks it was better proof than a spec sheet would have been. Given the flight test data and the block diagram one could interatively solve for the device performance specs.
This worked because
A) we had reliable data.
B) we had a full representation of the physics.
C) tuning is basically just rearranging the math and givens.
or to put it more precisely, this works subject to your assumptions on 1 and 2.
Given you have certain test data, given you have a full physical understanding, you can use modelling to resolve uncertainty in your input parameters. Inverse problem style. Instead of predicting the output once takes the output as given and predicts the inputs. durr.
Now the cool thing here is that the model actually output more than one metric. It output M1,M2,M3, M…. and so you could
tune on M1, and then check on M2,M3,M4 etc. Of course if M2,M3,M4 can be derived from M1, that doesnt tell you much, but if they are in different physical measurement units that improves your confidence that you’ve done things right.
So what do you know when tuning fails.
Suppose you had a model that predict global ice and your one tunable parameter was “initial ice cover”. and you varied this from 0% to 100%. and suppose that none of these settings led to anything close to the right answer. Well, that would be good evidence that your physics was wrong or incomplete. Even failed modelling experiments give you information. yes Virginia models do give you observations. In this case the observation is that your physics is wrong. I always laugh when Sceptics say that models dont produce evidence or data. And then they use the FAILINGS of GCMS as evidence that they dont capture all the physics. err, well we will leave that lapse for another day.
Moving on
Suppose you can reproduce M or have a small error between M` and M. What’s that show you?
It shows you that Given acceptance of your observations M, you can assert a complete relevant physical understanding of the system that is within the uncertainty of your input data. That is there is possible world where your physics is true. Or rather, given the uncertainty of the input data your physical understanding can’t be rejected. It is a good explanation, subject as always to improvement or replacement.
Back to GCMs. There are a couple of interesting points to make.
First is that GCMs have huge metric spaces. We could look at temperature, land ocean contrast, spatial variability, SST, MAT, SAT, wind, clouds, precipitation, sea surface salinity, etc. Tuning to one of them and getting the others correct actually gives us confidence that the physics is correct. But If we tune to temperature and get the land ocean contrast wrong, Then WTF .. danger will robinson.
There are a few low resolution GCMs where one could do the kind of tuning mentioned above. Example would be FAMOUS.
http://download.springer.com/static/pdf/373/art%253A10.1007%252Fs00382-010-0934-8.pdf?auth66=1419283244_a63f6ba109325c92b4917bf0320d95d8&ext=.pdf
However, Tuning a GCM in this manner in intractable. In short,
you can’t afford to vary all the parameters to actually tune
the model to output parameters. The best you can do is “hand tuning” Someone on the team found a range of settings that get’s you close. dont touch those knobs fool. touch this knob or that knob, but leave the others alone.
One approach to improving this situation is to create an emulation of the GCM. James has some work on this
http://classic.rsta.royalsocietypublishing.org/content/365/1857/2077.full
So.
1. Do the teams vary all the parameters in a systematic fashion to minimize the difference between the output metric and the observations. No. They don’t Tune.the models are not TRAINED.
2. Are there knobs they twist in a limited fashion to give “sane” results? Yes.
3. Do all teams twist the same knobs and look at the same metrics? No. Some, for example, may only do a sanity check that they get 1850-1890 “correct”.
If they ACTUALLY tuned and actually matched observations that would be a GOOD thing as it would demonstrate a good physical understanding within the limits of input uncertainty/structural uncertainty. It would show a consistent physics capable of explaining things within known uncertainties.
One doesnt want less tuning, one wants mo better tuning.
“I’ve also had that same feeling about my ability to communicate clearly.”
HaroldW, if, by that comment, you mean one’s ability to clearly understand another person’s communication I am with you there. When I read some technical paper’s I go into it with my own ideas of what the paper will be attempting to resolve and in some cases how. That anticipation is a major hindrance to reading and understanding a paper and is why I sometimes have to reread the paper with a more open mind.
The Sanderson and Knutti paper you linked is a mild example of my problem. The paper talks about structural errors in the models as a bias, I think, and other more random errors that the authors appear to attribute to the model not being constrained sufficiently (tuned) to the observed reality. They talk about the centeredness of the model ensemble mean and the errors around that mean. What I found lacking or I misread or is due to “my lack of communication abilities” was the stochastic part of the model results arising from white/red noise, or “weather” noise as Lucia sometimes refers to it, that is associated with a chaotic system. How would that noise be separated from the tuning noise? When the discussion turns to the inability to compare models means to observed temperature trends over 15 year periods I have never heard anyone refer to the problem being related to tuning noise or a reasonable facsimile.
Also the authors make no bones whatever about implying that some parameters of climate models are tuned to the historical record and probably should be, and further that tuning has it limits when it comes to structural errors in the model, i.e. there are errors that cannot – or perhaps should not – be tuned out of the models.
Does anybody know if Koonin’s questions about model rescaling in the APS review ever got answered?
It starts here on page 257.
http://www.aps.org/policy/statements/upload/climate-seminar-transcript.pdf
KOONIN: “So, to me, it looks like they set a calibration against the historical data and then they wiped out that calibration in doing the centennial projections resulting in probably a 25, 30 percent overprediction of the 2100 warmings.”
As near as I can tell the models scaled forcings for hindcasting to tune their models, but when they set the models in forward prediction mode they reset these scalings. For example the GHG forcing may have been scaled as 0.7 during hindcasting, but was set to 1.0 for forward prediction. This almost seems too fantastical an error to be true(?), but it wasn’t answered in the APS review.
DeWitt+Payne (Comment #133594) angech,
“There are a number of conceptual errors in your post above. The first is that you clearly don’t understand the concept of anomaly wrt temperature.”
Thanks ,I am trying to improve and have made conceptual errors in the past, some of them are a bit repetitive but I am willing to learn. I am still lost.You said.
“The model global average temperature [model GAT] for the month of January for each year from 1961-1990 are averaged.”
a GAT is incorporated into the model, fine. Surely all future calculations are done off the base of this average, which is the summation of all the anomalies to this date.
“If there’s an absolute temperature difference between model and observations, that is lost when calculating the anomaly.”
Lost on this comment. surely the difference is the anomaly?
You said †Absolute temperature can be off by ± 2 degreesâ€.
So is the absolute temperature you are referring to the anomaly temperature differing from the mean.
Or the mean model temperature from which the anomaly has moved 2 degrees.
Is the absolute temperature the anomaly or the mean?
Steven+Mosher (Comment #133597)
I agree that modeling in any endeavor is a good thing and even a better thing when all that goes into a model is transparent and its further assumed that the model is not necessarily a finished product and without limitations. There is bit too much in these discussions that it is all physics on one side and it is all tuning on the other. I think what is needed is to concentrate on the specifics and details of the matter. As you say when you have lots of parameters in model one could expect that over tuning one part of the model will under tune another part as long as the model in the case of climate is primarily based on the applicable physics is not being willy-nilly fitted to the observables – which I think most of us would agree it is not.
I would question a couple of issues here though. If the incorrect aerosol forcing were used in order to compensate for the climate model emergent parameter of Equilibrium Climate Sensitivity in the historical period, how would this problem be manifested in other model parameters? Also I would never use an example of modeling a system where the “truth” can be tested out-of-sample in a reasonable time period as in your reverse engineering of a flight control system example without pointing to the very real differences with testing a climate model out-of-sample.
I remember being appalled when I first discovered that Excel had a function that allowed you click and drag single points on a graph while the program automatically adjusted the data in the speadsheet for you.
Kenneth
“As you say when you have lots of parameters in model one could expect that over tuning one part of the model will under tune another part as long as the model in the case of climate is primarily based on the applicable physics is not being willy-nilly fitted to the observables – which I think most of us would agree it is not.”
based on WHAT evidence do you say this.
In 2007 Gavin pointed me to ModelE code. I went through it.
have you?
How many tunable parameters does the typical model have,
can you name them.
Skeptics need to be more skeptical of their skepticism
This reference does a good job of defining model tuning and explaining how climate models are tuned:
Tuning the climate of a global model,
though it appears to sidestep some of the issues with tuning of inputs.
Kenneth,
“I am of the opinion that that paper gives the official or one of the official calculations of the CMIP5 model ECS and TCR values. What are the alternative methods? And which is the best/better in your judgment?”
I have not seen any good papers which quantify feedbacks at all. You need to distinguish between on the one hand estimation of ECS and TCR in GCMs, which are well-defined and can be accurately estimated directly from the GCM runs, and on the other hand the analysis of the feedbacks which are effective in the GCMs. It is the latter problem where the nonsense starts.
I hesitate to accuse people of wilful stupidity, when in reality, it may just be the common-or-garden kind, but the feedback analyses published to date are generally all very misleading in my view – they grossly underestimate the total feedback value applicable in the GCMs in the historic period.
The problem actually starts with a high school maths error. If one starts with an equation of the form:
Net flux = F -lambda*T where lambda is a constant,
then a plot of Net flux vs Temp for a constant forcing run should yield a straight line of gradient (-lambda). When the actual GCM results show something which is not a straight line, it means that lambda cannot be a constant. When lambda is not a constant, the value of lambda is not given by the gradient of the plot. It is given by the secant gradient of the plot i.e.
lambda(t, T;F) = (F – Net Flux(t))/T(t)
This is pretty basic calculus.
The value of F should be given by the stratospheric adjusted forcing value (the old “Fa” values), a point I will return to.
Instead of doing this, the Forster approach takes the late-time gradient for the calculation of the feedback. It is easy to show that this value normally lies below (and sometimes well below) the possible minimum value that lambda can achieve according to the run data.
So the total feedback term estimated by Forster is generally too low. A compensation is required and it is obtained by (a) calculating an adjusted forcing value (AF) by construction; this AF value cannot be physically related to the forcings and feedbacks which were actually applicable in the historic runs and (b) filling in the missing gain in an F vs T plot with the dreaded kappa estimate. Meh.
Why should Fa values be used IMO? Because, they can be related reasonably objectively to the RTE calculations before any tropospheric adjustments in the model. The use of the new ERF forcing values already eliminates some of the so-called fast feedbacks by including modeled tropospheric adjustment. The AF values are closer to the ERF values than to the Fa values, but generally lower, and inspection of Forster’s data for the quadrupling runs shows that the feedbacks so eliminated can take up to 30 years and span several degrees of temperature change. This high feedback period for the so-called fast feedbacks dominate the historic runs and is much much more important than the rather meaningless late-time gradient of net flux vs temp, although it would take me a while to explain why.
So all in all, I think that the approaches in recent years have grossly understated the feedbacks applicable in the GCMs during the historic period.
Tom Scharf,
I think that the question was answered. There was no rescaling of the centennial scale projections. However, the conversation was not about tuning hindcasts. At the last minute, the chapter 11 authors added in a quick fix to try to regain some credibility in the short term temperature projections. This was an ad hoc adjustment done offline (after the circulated second order draft). Koonin’s question was why, having rescaled the decadal predictions, they did not also rescale the centennial scale projections.
Mosher,
It would help if someone who actually has the programming and math skills to make sense of the Model E code would actually help those of us who don’t instead of making drive by cryptic comments that lead nowhere.
DeWitt Payne (Comment #133609)
Your comment is much the same as I was going to post in reply to Mosher. To be frank I do not even understand what Mosher’s point was in answer to what he excerpted from mine. My comment was intended to note that a model based primarily on the physics of the system being modeled (as I judge climate models to be) would be difficult in most cases to arbitrarily tune by a simple adjustment of a parameter as it could well affect other parts of the system adversely and put those parts further from the observed reality. I am curious whether using aerosol forcing could fit a model better to the historical period without throwing another part of the model output out of kilter. I thought that was in line with what Mosher said in the excerpt below but I could have misunderstood because I am not sure what is the significance of different physical measurement units:
“Now the cool thing here is that the model actually output more than one metric. It output M1,M2,M3, M…. and so you could
tune on M1, and then check on M2,M3,M4 etc. Of course if M2,M3,M4 can be derived from M1, that doesnt tell you much, but if they are in different physical measurement units that improves your confidence that you’ve done things right.”
Anyway, Mosher, if you can use or spin my post to fit your view of wrongheadedness by skeptics in the world of climate science be my guest. Heavens know that it exists on all sides of the AGW issue. If you write another book just make sure you spell my name correctly.
Paul_K (Comment #133607)
Paul, thanks for the comments. I have had some of the same reservations about the Forster paper that you talk about in your post here – although I need to reread it to fully understand the details. The CMIP5 ECS and TCR values from his paper are the same as those published in chapter 9 of the AR5. I assumed that Forster used his AFs to derive those values, but I could be wrong.
Do you have any alternative values that are calculated and/or published for the CMIP5 models? Also do you have a source for the 4XCO2 CMIP5 model run data? It was not available from KNMI and I went to the link below to obtain it but have not downloaded it yet.
http://cera-www.dkrz.de/WDCC/ui/EntryList.jsp?acronym=ETHc2
Carrick (Comment #133606)
Carrick that link is broken for me.
DeWitt:
Well anyway, it’s not like you’re going to learn very much about the the quality the approximations or implicit tuning just by looking at the Fortran 90 source. And it’s not like you need Gavin to help you find the source, since a link to the tar ball is posted on their website in any case.
So just a strange comment from Mosher.
Kenneth, that’s odd. I just tested it again and it works.
Anyway, try this link.
Paul_K, Thanks for the info. I assumed there had to be an answer to that made sense.
Models predictions are evidence?
By the clinical definition of the term “the available body of facts or information indicating whether a belief or proposition is true or valid” maybe one could make a case it is technically evidence as in “information”. The push back is due to some who wish to use model predictions as “fact”.
It would seem that before models should be used as material evidence, they must first be validated / have a good track record.
Of course it depends on what you are using them for.
Are GCM’s evidence that warming will likely continue as GHG’s continue to increase at BAU? Useful information Yes. Fact no (in the sense it is a prediction).
Are models evidence that sea level rise will exceed 1M by 2100? If the highest of the range of model outputs of the worst case emissions scenario is used, this is information. When one shortcuts this to “sea levels will rise 1M by 2100” the models are actually evidence this won’t happen. Are models currently trending toward worst case? No, they are 15% low to median currently. How likely is RCP8.5 to be a reality? Pretty low, especially if you believe China’s commitments. So are models evidence of a 1M rise by 2100. I would say information=no and fact=no. But it cannot be excluded, ha ha. This can be remedied by simply including the low end in statements, “sea levels will rise 8 to 39 inches by 2100”. When is the last time you saw that kind of phrasing in the media? The least offending version of this deceptive practice is the ubiquitous “up to” phrase. Sea levels will rise up to 1M by 2100. Accurate? I guess. An honest effort to convey the information to the user? Not so much. It is phrased for maximum impact as many AGW claims are nowadays.
Are hurricane season models evidence of the upcoming hurricane season? Not really given their very poor track record, not much better than random.
Are hurricane tracking models evidence of where a hurricane will be in 48 hours? Very strong evidence given their very good track record.
Are weather models evidence of if it will rain tomorrow in location X? Yes, but not so much in Florida in the summer where tropical patterns are pretty random. Normal prediction is 50% chance of scattered thunderstorms tomorrow every day all summer, we don’t even watch the weather prediction, no useful information.
It should be obvious that comparing model outputs to observations is evidence of how valid the model is.
Carrick (Comment #133614)
That link works. Thanks.
Carrick,
The first link went to ‘forbidden’.
The second is ok.
Wow.
Figure 2 of the model tuning paper is a good poster for uncertainty.
Figure 4 is the most interesting to me. The idea that models have an unforced equilibrium TOA radiative imbalance larger than, in some cases much larger than, the current estimated TOA imbalance, i.e. they don’t conserve energy, should be profoundly disturbing.
Hi again Kenneth,
“The CMIP5 ECS and TCR values from his paper are the same as those published in chapter 9 of the AR5. I assumed that Forster used his AFs to derive those values, but I could be wrong.”
Can I re-emphasise that the problem is not with estimation of ECS and TCR in the GCMs. (In other words, the values may bear no relationship at all to reality, but they can be estimated unambiguously for each GCM.) ECS is normally estimated from a “late-temperature extrapolation” of a net flux vs temperature plot to a zero net flux value, the plot data for which comes from a long-running step doubling or step quadrupling of CO2. TCR is simply measured by running a 1% p.a. CO2 case several times from a timepoint in the pre-industrial control runs, and taking an average of the resulting temperatures at the point of doubling CO2.
The ECS can be calculated without Forster’s AF value, but in practice the relationship should be very close to ECS = AF/alpha, where -alpha is equal to the (negative) late-temperature gradient on the net flux vs temperature plot. Forster then assumes that this gradient reflects the feedback in the model (it doesn’t).
In theory at least all of the CMIP5 runs are available here:-
http://cmip-pcmdi.llnl.gov/cmip5/data_portal.html
I have not personally tried to crack this nut. Various friends have reported problems with protocols and compatibility issues. Also, you will need to post-process very large amounts of data, and my PC still has a hand-crank on the side. Troy Masters may be willing to help you if you ask him nicely.
Thanks again, Paul. The light is getting brighter.
2 random thoughts:
If you wanted to “tune” a model, a neural net might be a lot better than a person.
The issue for models is not to cover the parameter space, but to find it’s limits and the critical paths for trajectories.
Paul, I should add that my limitation, at least for now, is me. My computer has not yet let me down in doing some rather large simulations or in working with large data bases. I have the tools now to do with good confidence the ECS and TCR calculations from the CMIP5 model data bases and will be downloading the data today. My main interest is in getting a good feel for what the variations in these parameters will be from model run to model run.
Mosher said “Do the teams vary all the parameters in a systematic fashion to minimize the difference between the output metric and the observations. No. They don’t Tune.the models are not TRAINED.”
Since the CGM modelers do not perform a systematic approach – your definition of tuning – what would you call what they do to “minimize the difference between the output metric and the observations.”?
Mosher
Mosher, going through the code would not tell you whether tuning one parameter untunes another. As for “lots”– well, there are “lots”. Or not. Depends on how many one considers “lots”. Some people call three tweakable parameters “lots”; some people don’t consider it “lots” until you hit 20. Looking at Model E code isn’t going to resolve this. (And if it has fewer than 3 knobs, something is wrong. Seriously.)
The fact that this is not a Navier-Stokes solver and the model can’t resolve scales that exist in the physical world means the model must parameterize ‘diffusion’ of all sorts (momentum, mass, heat and so on. That’s already 3 knobs). IF they include interactions between oceans and atmosphere, they must parameterize that transfer across phases. They must parameterize cloud physics.
This is true even if the decision to parameterize results in no tunable constant because the choice was “neglect that and set it to zero”. Setting something that is non-zero to zero is in some ways a “knob”. You just turned the “knob” to zero.
On the issue of whether they “tune” or not: that also depends on precisely what one means by “tune”. Of course they can’t tune in the way one tunes better understood systems. One thing about aircraft: They are engineered. By definition that means that designers are largely working in physical ranges where we understand the things that matter rather well. If we did not, we generally couldn’t design the device. So: yeah, if you call the exercise of fitting parameters to behavior of a designed device ‘tuning’ people creating AOGCM’s don’t do that. But that’s not necessarily what everyone means by the word “tuning”. Some just mean “fiddling with the knobs to do as well as one can.” And modelers do do that — to at least some extent. (There isn’t even anything wrong with doing that. It’s mostly right in fact. But it is done!)
Anyway: engineers did not design the planet. Ab-initio computations of the climate are not possible. Tweaking of parameters is done.
Lucia:
Put another way the solution space of possible Earth-like planets is infinite, but the Earth is just one of them.
So it’s going to take a fair number of parameters to specify *this* planet, leaving aside the issues with the approximated physics (which require ad hoc parameter choices to make them work).
Further, the aerosol forcings (and other inputs) are historically poorly constrained. This together with the selection of the solar forcing provides an indirect tuning mechanism.
“engineers did not design the planet.” Magratheans disagree! 😉
Back in the real world, I’m struggling to understand Mosher’s distinction between “tuned” and “trained”. I understand that there is a huge space of observations, and a large number of adjustable parameters. Further, the sensitivity of various metrics to the parameters is not derivable in advance. (Presumably model runs can establish this, but perhaps at an unacceptable cost of computational time.) Well, OK, let’s agree that it isn’t a simple minimum-error problem. Sanderson&Knutti acknowledge that tuning isn’t perfect.
The fact remains that we have an ensemble of models whose sensitivities vary by a factor of 2 or so, whose average temperatures vary by +/-2 K or so, and yet whose anomalies are claimed to match 20th century temperatures to a remarkable accuracy. To me, this suggests that history has been a factor in the adjustment of those models. Perhaps not directly, in the sense of distance-minimization. Perhaps indirectly, in the sense of rejecting certain regions of parameter-space (and/or algorithm-space) which don’t match the historical record well.
It’s not that I object to tuning as a principle. Given that the models are computationally unable to numerically integrate the underlying equations, some simplification/parameterization is necessary, and those parameters are not likely directly computable from physical principles. In which case, using observations to set parameter values makes sense. What bothers me is that the GCMs are claimed to be able to predict ahead 100 years (and more!) despite the fact that this involves extrapolations to regimes for which there are no observations to constrain the algorithms/parameters.
Regarding tuning ModelE,
I looked at the FAQ. I read that there is a parameters database module anyway for loading all the parameters quickly on a restart.
Mildly interesting Not quite interesting enough to spur me to investigate in more detail. But I’d start by looking at this module probably if I wanted to get to the bottom of the parameters ModelE uses.
Note – assuming that ‘parameters’ and ‘tunable parameters’ are the same thing might be a valid idea, might be a preposterous idea, don’t know and am unlikely to go find out.
[edit: the discussion of primitive metadata in the faq is interesting too and makes me think that these guys ought to consider bringing in professional software engineers.]
I think it behooves me to praise these guys for putting in a good bit of extremely readable documentation. No way to know if it covers everything you’d need to know or not, but this for example. I don’t always see an effort in software to help a reader understand that much. Brief? Sure, but still helpful.
Mark Bofill
They aren’t necessarily the same. “Viscosity of air” is a parameter– and if it corresponds to the physical viscosity of air known from measurements, no one would call that “tuned”. (Or maybe Mosher would? But if so, few others call that “tuned”.)
But turbulent viscosity or any other sort of diffusive parameter which represents something like “approximation for sub-grid behavior” is a “tunable parameter”.
That said: if the tuning was somehow known and matched what was done in other fields, one might have greater confidence in the ‘parameterization’. If it’s value, for etc is unique to climate science or to an individual model– well, that’s fully, truly “a knob”. I don’t know how turning that knob to better match an answer is characterized by Mosher. Lots of people consider that to fall in the broad range of what is called “tuning”. But Mosher seems to define “tuning” as something rather more narrow– which is fine. It may well be useful to define it narrowly when doing control theory or other types of work. But I don’t think his more restrictive definition of the word is shared by “everyone”.
If that definition of “tuning” was shared, “everyone” would need to come up with yet another word that meant “fiddling with the knobs to get better agreement, though possibly not in a highly systematic way.” I think many people call that “tuning” too.
Hi Lucia,
Merry Christmas. I had expected that Mosher would have a very specific way in which the word “tuning” could be applied to the development and management of models.
I had suggested “adjustment” instead. I don’t suppose his usage was intended to defuse suggestions that the models are adjusted or modified to match experience. But surely they are. How else could one have any idea whether these things are representative of anything?
Lucia,
I think it helpful to draw a distinction between assumptions and parameters. Tuning in models is mostly adjusting uncertain parameters to match history (more-or-less) under the influence of a specified set of uncertain assumptions. I suspect it is mainly via the assumptions (snooping the diagnosed sensitivity, of course) that modelers bias the models toward ‘desired/expected’ behavior. A couple of glaring examples are historical ocean heat uptake and historical aerosol off-sets. I do not know what the current CMIP5 estimates are for current heat uptake, but in the past the models tended on average to calculate more heat uptake than can be documented by measurements (especially ARGO). I am pretty sure the modelers (on average) assume current aerosol offsets which are considerably higher than AR5’s best estimate (~1.4 watts/M^2 versus ~0.9 watt/M^2 in AR5). If the objective is to reasonably match historical warming (and this seems a minimum requirement for model credibility), then “tuning” of uncertain parameters based on overstated values for ocean heat uptake and overstated aerosol offsets will inevitably lead to higher than correct model estimates of sensitivity, both transient and equilibrium. More importantly, that tuning will lead to overstated projections of future warming…. which is what seems to be going on right now…. as well as incorrect patterns for rainfall, rainfall frequency and intensity, statistical characteristics of variability (regional and global), and many other measurable parameters.
.
In any normal field, these kinds of discrepancies would lead to revisions in the model tuning based on best measured values for things like aerosol effects and ocean heat uptake, but as Richard Lindzen long ago observed, when the models and the measured data conflict, the field focuses on finding reasons why the measured data are wrong, rather than on why the models are wrong. If the modelers have any sense at all, they will accept measured data and tune the models to better fit that data, even if it means significantly lower diagnosed sensitivity. I’m not counting on that happening. More likely, the long list of proffered explanations/excuses (often in conflict with each other!) for why the data are wrong and the models right will continue to grow. Anything (Anything!) except lower estimates of sensitivity.
.
I find it all a bit sad really.
.
On that cheerful note… Happy holidays to you and yours.
j ferguson,
Merry Xmas to you too!
Well… yes. To some extent, turning the knobs to match observations is “the scientific method”. It’s only when people proclaim things are somehow “right” or “near right” or that we should have confidence in projection/predictions before the ability to project/predict has been demonstrated that “tuning” is a bad thing.
In other “normal” fields, discrepancies between models and predictions mean models don’t predict whether a design will improve of degrade performance. Failed predictions can be detected relatively quickly– and are often obvious. So of course models are revised.
Mind you– modelers do tend to over sell models in all fields. On the other hand, they need to sell them to people who use them, and those people can tell if the models were correct enough to be useful. The customers will risk some money– and even use models that aren’t entirely “correct”– judiciously.
One of the problems with climate science is that the “customers” are…. who? Funding agencies? Their metric for “good enough” is often “modeler published a research paper that other modelers thought good.” And anyway, the time scales for testing are soooo long that people can’t detect that a vintage 2014 climate model was “right” or “wrong” for decades. Plus, from the POV of a funding agency, the only ‘opinions’ on whether the models were “right or wrong” that matter or those of climate modelers.
In contrast, customers for CFD applied to aeronautics (or other fields)– those were manufacturers. They didn’t really on modelers, academics, governent agencies or “international panels” assuring them things were “useful” based on moderlers assessments. The manufacturers decided how/when to use models, what to believe, how to interpret tests and so on.
Are CFD models useful? Heck yeah. Were they useful even long ago when their use was more limited? Heck yeah. But modelers were forced to moderate claims about what really worked because people could and did check. And many manufacturers weren’t going to spend money on a CFD package and rely on it merely on the say so of the person who wrote the package!
Thanks Lucia. Merry Christmas!
That sounds pretty reasonable to me.
Further on that note, we might consider what to call this:
So it sounds like people change the code around until it ‘works’, and sometimes we decide after the fact what ‘worked’. I don’t know what to make of this. ~shrug~ It seems like messing around with the code could count as a knob, depending on what we were messing around with.
CFD? ? Fluid Dynamics?
John,
Computational Fluid Dynamics, I think.
Yes. CFD= computational fluid dynamics.
“Mosher, going through the code would not tell you whether tuning one parameter untunes another. As for “lotsâ€â€“ well, there are “lotsâ€. Or not. Depends on how many one considers “lotsâ€. Some people call three tweakable parameters “lotsâ€; some people don’t consider it “lots†until you hit 20. Looking at Model E code isn’t going to resolve this. (And if it has fewer than 3 knobs, something is wrong. Seriously.)”
In one tuning study I looked at there were 32 tunable parameters
And yes, looking at the code can tell you about the tuning process.
It’s what’s NOT there that counts
“Steven,
the contributions of doing something vastly outweigh snipping at Tols heels or Mann’s heels.
So if I were to find a flaw in a mathematical proof, I couldn’t publish that unless I had a valid proof of my own? Oh, puhleeze.
#############
Of course you could publish it. And That has nothing to do with my point. My point would be publishing the correct theory outweighs snipping at peoples heels. Its about balance
” Of course it was difficult to falsify MBH9x because the data and the code weren’t readily available.”
No code. no data. no claims worth refuting.
“Also, how can you rationalize creating a temperature reconstruction from tree ring data if you don’t believe that the temperature signal can actually be detected in the data? Which is something that can’t be proved.”
easy. do that math and show the answer. easy. No signal will give you floor to ceiling error. or rather show that there is no signal. easy peasy.
Mosher,
Sure, in the best of all possible worlds. But, Pollyanna, we don’t live in that world.
And unpublishable.
Not to mention, that sounds very similar to what McIntyre actually did, which you refer to as snipping at the heels.
Steve Mosher
No. it can tell you what might happen during tuning. It doesn’t tell you what was done.
What’s not there does tell you what can’t be done– but that’s not a heck of a lot of info!
What parameters are provided might be a moot point if it is common practice to modify the code to effectively change parameters. I have no idea what the code modifications referred to previously are typically about in practice. I’d think though that if they were simply to fix defects, crashes, core dumps etc., that by now the model would be at the point where stable versions exist and this wouldn’t be necessary.
Steve Mosher’s example of model tuning, while interesting is I hope not what GCM modelers do . Flight control models can be very specific to the aircraft from which the flight test data was taken. There is no reason whatsoever to suppose the underlying “physics” is necessarily right. If you take another aircraft, things could be a lot different, e.g., aero-elastics could be a much bigger factor and that could mess up the force and moment curve types, i.e., they could be a lot less linear.
The situation for GCM’s I would argue is far more similar to tuning a turbulence model. You are hoping the model will perform “out of sample”, i.e., be able to predict flows that were not in the database you used to tune the model. In reality, model developers, and there are really only a very few in the world who really do stuff anyone cares about, use intuition and expert judgment a lot and mathematical data fitting is not used at the highest level to set constants, etc. Data fitting is used at low levels to parameterize boundary layer profiles, etc.
Mosher’s example is interesting, but somewhat different than what I hope GCM modelers do. It’s unlikely in my view, that GCM modelers are as advanced as turbulence modelers because the data for clouds etc. is just so much worse. I’ve mentioned it several places, but can’t remember if i mentioned it here. The paper by Leschzinger and Drikakis in Aeronautical Jounal in 2002 is very good on turbulence modeling.
Lucia, CFD is tremendously useful as a quantitative tool for designing the outer mold lines of an aircraft. For separated flows it can be useful in a qualitative way. For certifying an aircraft, its currently not accepted by the FAA. That could change but I’m not sure the public is ready to base public safety on simulations quite yet.
From my experience in working on large problems, at some point, the inputs outweigh the code in complexity. Model E isn’t even a very large code, as project sized codes go.
In this case, where you have multiple choices for inputs, even understanding the complexities of a particular input (e.g., aerosols) can be a lifetime project.
I know for people like Gavin Schmidt, that he has to consult with specialists within his own building when questions about aerosols come up. That’s not a slam on him, it’s more of a statement of where the real complexity of the problem arises.
In other words, there’s quite a bit more to modeling climate than solving Navier Stokes, and studying how a particular model solver works isn’t necessarily going to give you that much insight into how to accurately model climate.
Actually Carrick, Navier-Stokes has something similar. Turbulence models are designed by specialists and CFDers just use those models. When questioned about that they say “I trust Dr. Famous Modeler.” The same is true of unwinding schemes. Most CFDers don’t know much and just use something from the literature. I do believe this complexity of the model issue that you point out is becoming more of a problem. It probably makes it more difficult to construct simpler models that encapsulate insights into the physics if no one really understands the interactions. People get specialized and become specialists in one type of sub grid model, specialists in writing and maintaining large codes, numerical algorithm people (even though my information is that GCM groups are systematically driving those people out, probably because of too many inconvenient questions), model runners, analyzers of output, etc. A lot of people to coordinate. It’s a job that is very difficult to do well.
David
Agreed. On of points of ‘engineered devices’ is also that people are working in regions where devices ‘work’. So: Predicting lift on airfoils at low angle of attacks was do-able long ago.
Predicting what happens at high angle of attack? With a separation bubble? Different. But the fact is: ‘stall’ is not a good thing– and pretty much that’s the region where the plane doesn’t work. So, designers didn’t need stupendously good tools to predict that behavior. (Not that people aren’t interested!)
In contrast: climate is what it is. Whatever is important is whatever is important. That’s the points I’m trying to make about the earth’s climate not being an engineered device.
Modeling is used to guide and accelerate design.
But not being used to certify is sort of my point. It won’t be used to certify until lots of people other than modelers agree it ‘works’ sufficiently well for that.
Yes, Lucia, I agree. However, one minor quibble. Stall is actually important for airplanes. With flaps deployed you need to know the maximum lift and that condition is right on the verge of stall. If your code stalls early as most do, it is important. I believe flight tests sometimes go up to the buffet boundary which is also a massively separated situation.
David
Yes. We agree. I actually avoided the issue of “predicting when stall occurs” in the previous comment. That’s much more difficult that predicting lift when stall has not occurred. It is also very important– and so has been an active area of research for… oh… a century? (Since Prandtl at least.) Obviously, people have been trying– and succeeding in improving the ability. But that problem isn’t easy.
My point is merely that: To a large extent, engineering is does focus on ‘the possible’ and the main operating conditions are in regions where things are fairly well understood. And our codes do work in many regions that are important to be able to design things that “will fly”. But people don’t claim the fact that the codes predict those things well implies they predict other things well. They don’t: it’s perfectly possible for a code to be very good at doing one thing (lift on airfoil before stall) and failing at others (when stall occurs, flow around the airfoil after stall when separation occurs.)
Heck we even understand the physics for why stall occurs. And yet, it’s still difficult to predict as well as people would like. We know why separation occurs– still hard to predict flow after separation.
Back to the issue of “main” conditions: Test flights are not limited to main operating conditions and yes, the operate in regions where codes cannot work well. After all: we do want to know how the plane functions. We want to know that even if we can’t predict it using any code. So: Test.
Where the codes do not work– everyone knows and admits that testing trumps code. Heck, testing trumps code even where or if one thinks code does or should work. It’s just that testing has shown code does work for certain things.
And getting back to SteveF’s point about “normal fields” and code: In a “normal” field– like aeronautics– if codes fail to predict, people admit the failure. You don’t just have modelers assuring other modelers (and the public) that the codes are “useful” and so — somehow– we can rely (in some way) on the “predictions”. In most fields there is a large group of users (e.g. test pilots, manufacturers etc) who are entirely outside the ‘modeling’ effort and whose opinions are totally inpendent of ‘modeling’ and those opinions count. In fact– they count more.
Everyone in any engineering field respects modeling. But in other fields other information balances modeling. In climate? Not so much. And less and less over time.
If nothing else: there isn’t much data and what they have comes in s_l_o_w_l_y. Often it’s ambiguous– and the effort to interpret it itself involves many tenuous modeling assumptions and is rife with the potential for cherry picking. (Tree ring paleo for example.)
In aeros: Data comes in fast enough to make sure modeling doesn’t get totally unhinged from reality.
Carrick
Yep. I only raised the NS connection because in 133605 Mosher was challenging whether people had looked at model E and the tenor of his comment seemed to suggest that codes did not involve ‘lots’ of parameters. (At least that’s the tone I read into
In fact later, he admits there are what I would call “lots” of parameters at least some models
Or maybe Mosher merely means in some study about something else (aero?) something had 32 parameters. In which case, Mosher is being too cryptic for me to follow. Because: yes, some studies on something somewhere in the world will have more tunable parameters than in climate models. To which one would say: So?
In fact: climate models do have “lots” of tunable parameters– and they have “lots” even if someone like Ken did not dig into ModelE to count and determine the specific number. These models must have lots because at a minimum they have the number of tunable parameters involved in Navier-Stokes. (And that’s true even if climate isn’t ‘all about’ or even ‘mostly about’ NS) Above that, they must have tunable parameters for anything and everything sub grid which is a lot of stuff (clouds, transport across ocean/atm, heat,mass transfer between dirt and atm, ice physics etc.)
And besides that: If Mosher thinks Ken is wrong because there are not lots of parameters, Mosher should tell us the number he figured out and say why that number is not “lots”. Merely challenging Ken (or others ) with “have you counted?” Uhmmm… if there are lots, how does not knowing the precise number make the number “not lots”? (Answer: it doesn’t. There are still “lots”.)
And on top of that: model runs require assumptions about forcings (aerosols, even solar) and testing against data is difficult because among other things we don’t have much data. (Detailed precipitation every where in the world in 1820? Nope.)
This expresses what bothers me about the current use of climate models extremely well. I agree and this cuts deep towards the heart of what I consider the problem to be with climate modeling in general.
Mark–
My view is that
(a) We know AGW is real from very simple models which are generally supported by data we have. Specifically: temperatures have risen detectably since the time people first predicted the effect of CO2 and other gases.
(b) We can be pretty darn sure that AGW is not ‘miniscule’ and the rise over a century is probably not mostly natural variation. It’s probably the CO2 and other gases.
(c) Beyond that: we don’t know.
(d) AOGCM’s appear to be over-predicting and the issue is not “weather noise”. Whether that’s sensitivity in the models, choice of historic forcings or what have you…. not entirely sure. But the collection of models used by Climate Scientists in AR4 and AR5 have been over-predicting warming. It’s generally wiser to assume that models (or modeling efforts) that have been over-predicting will continue to over predict. Because whatever factor was resulting in the overprediction– it’s there contained in the models themselves of in the mass of assumptions involved in “the modeling effort”.
(e) Of course, it’s possible, the factor causing “over prediction” is something like “unknown volcanoes” etc. But I’m dubious. And I tend to see the general group dynamics to be trying to “save the reputation of our babies– the models”. But…. we’ll see.
BTW: wrt (e) modelers always want to sayve the reputation of their babies ‘the models’– that happens in all fields. But in other fields, people have more data and can do more controlled experiments. So eventually, people need to fix a model. Then modelers get excited about the new fix– and things go on. But climate is dominated by modeling because there is hardly any data.
->The real consensus. I buy this too.
-> The mainstream view. I agree it’s probably so, but it’s not certainly so, and I’ll argue against this when I feel energetic and foolishly overconfident. 🙂 If I had to wager, yes, I’d put my money on this side of the line.
-> The activist bathwater has cooled to an unappealing lukewarm temperature at this point.
dirty no good gosh darn denier territory! :p
Regarding the models, modelers, saving their babies, and the scarcity of data, do we see these sort of issues in economic modeling? At first blush I imagine the problems in those fields would be similar with respect to the data.
Science of Doom had a post on tuning climate models focused on the Mauritsen, et.al. paper back in July.
Yeah, at a quick glance some of the issues seem strangely familiar.(edit, removed link)
Anyway, why did I ask in the first place? I’m curious because I thought it might give insight towards the impacts of political / policy motivations on modeling. Do the projections of economic models reflect the political philosophies of the modelers? Do I invite the tinfoil hat inquisition by even asking this?
(my answers in order: 1)explanation follows in paragraph, 2) I don’t know, 3) yes, I invite the tinfoil hat inquisition by demonstrating my counterfactual conspiratorial ideation and should be sent for immediate reeducation at processing camp 51)
Mark Bofill,
I don’t see a lot of similarity between economic modeling and climate modeling, save for that they both tend to be very wrong when confronted with reality. 😉 Climate models, at least in theory, are based on pretty well known physical processes. The problems are in the details of sub-grid scale parameterized behaviors (clouds, rain, etc.) and some seemingly arbitrary underlying assumptions. On the other hand, economic models don’t have any connection to physical processes, and so are at bottom purely empirical… mostly past correlations used to predict future events (though I am sure economic modelers would disagree!)… which is generally not a good strategy for predicting the future.
.
Of course both are strongly influenced by the political thinking of the modelers; Milton Friedman and Paul Krugman (Nobel Prize winners both) working together on an economic model is hilarious to even contemplate.
SteveF, yes, I see. In climate modeling people (presumably) agree in the overwhelming majority of the cases about the basic phyiscs. In economic modeling, maybe the validity of the individual ‘building blocks’, whatever they are, the basic principles, are more questionable. So maybe it’s an invalid comparison.
Thanks.
Interestingly, they say more or less what Kenneth said. For example Mauritsen writes:
They also use the word “tuning” the way people here have been using it– not in the more restricted way Mosher uses the word.
Lucia,
I found the paper and SoD’s discussion interesting. I wish the paper had addressed the fundamental issue of underlying assumptions like the size of aerosol effects in much greater depth. They do say that aerosol effects are uncertain and important, and acknowledge that there is something of an inverse correlation between assumed aerosol effects and model sensitivity, which implies tuning to an assumed aerosol effect. But that is as far as it goes. I would expect that modeling groups faced with this fundamental uncertainty would explore a wide range of assumed aerosol effects and ‘tune’ for at least a few different aerosol levels. I have not seen any published work along those lines.
SteveF
Oddly, the existence of the IPCC process may distort the activities of modeling groups away from what ‘makes sense’ if one wanted to really ‘do science’. The difficulty is that– to some extent– all modeling groups are going to “want” to contribute to chapters they know will be written, and they will want– to the extent possible– to contribute runs that are used in projections. This will tend to be a point in their favor w hen presenting to funding agencies. Otherwise, the question of “why weren’t any of your runs used in X” is going to be ‘out there’.
What that means is that some amount of time (x%) is going to be spend on “things that will be in IPCC” or “will be used in group intercomparison” and/or anything that is a “world-wide” comparison activity. Since budgets will always be finite, this means that exploring wide ranges of anything else will be more difficult to justify– not because they aren’t important but because they are not ‘visible’.
In contrast, in normal fields, a company will eventually explore something because it does affect manufacturing, design, development etc. Trickle up to funding agencies is imperfect, nevertheless an aerodynamicist|mechanical engineer|chemical engineer|biologist will need to be aware of which problems manufacturers have trouble dealing with and so on. And manufacturers want things to sell to consumers be they individual customers, or — in the case of weapons– the government. Consumers want things “that work”.
But in climate science? The linkage between “consumer”, “manufacturer” and “climate modeling”? Either not there or weird in the sense that the “consumer” might be something like “Greenpeace” or “Heartland” or some group who is looking for a nail to hammer. Or it might be someone who can’t tell if what they got “worked” (e.g. can the intelligence community tell whether the climate projections and associated predictions of warming climate leading to terrorism “work”? Nope. In contrast, a test pilot and others can tell if an airplane flies, how well and so on.)
Lucia,
You are probably right about the distorting influence of the IPCC AR process and the pressure to be included in that process. Seems to me that the climate modeling field doesn’t ask enough important questions, because they are too interested in satisfying the desires of their funding governments.
Here is one simple example: In the discussion thread on the tuning post, SoD was kind enough to dig up an AR5 graphic showing a (weak) correlation between model diagnosed absolute surface temperature and diagnosed sensitivity to GHG forcing. In his most recent post at RealClimate, Gavin showed a similar graphic of projected rate of warming from today through 2070 (IIRC) versus model diagnosed absolute surface temperature. Once again, a cooler modeled absolute surface temperature correlates with faster projected warming. Considering the clear discrepancy between the high sensitivity estimated from glacial (when it was much colder than today) to interglacial transitions and the empirical estimates of sensitivity from the last century or so (consistently low sensitivity), there seems to me a few obvious questions: Does climate sensitivity to GHG forcing actually fall with rising average surface temperature? If so, why? Does the much drier atmosphere of a glacial period increase sensitivity to forcing by changes in cloud behavior?
.
Those are the kinds of questions which I think the modeling groups should be asking, since the answers have the possibility of improving the models and making them much more useful. Spending time (and lots of money) on ‘production runs of projected doom’ doesn’t help improve the models…. and publishing those projections only makes real improvements in the models less likely…. since the modelers then have too much commitment to the published projections to easily walk them back.
Lucia has an important point about who the customers for GCM’s are when used for climate. It’s a little clearer when they are used for weather forecasts. Forecast skill is very important to the customers of weather forecasts. In principle, one might argue that the customers for climate would be the same as for weather, but in practice, climate changes are too far in the future to be worried about by private enterprises.
Carrick,
Thank you for that thread to Models and tuning, very informative.
The gist of it was that they made 3 other models with tuning by altering various inputs arbitrarily.
” asked how our model might have differed had a slightly different
path been followed. In so doing we created a small
number of alternative worlds”
In none of them did they consider altering the climate sensitivity to CO2 which makes all the models run hot.
Mosher and Zeke and Stokes could prove this by asking for an output of any of the models with a sensitivity of 1 and see how much closer to actual observations it would be but it would be anathema to them.
Without doing this all the models put extra heat into the system that is not real and then have to make excuses that there is a disparity in heat loss with the models compared to observations, the “missing heat” by model theory.
The sad thing is that all the components they use have rational reasons for their bounds or levels with scientific input into why such levels should be used.
When they do not work and they arbitrarily put in other values they are throwing science out the window.
How can one say “In World 2 we lower the planetary albedo by about 1 percent by doubling the conversion rate,” no scientific reason, simple human guesswork.
Say it had worked, The planetary albedo is still the real albedo, the massaging may work but only by luck not by science and hence cannot be guaranteed on repeat or anywhere else.
David Young
There is no particular overlap. Tour companies, shipping companies etc have a critical need for short and medium term weather reports. Farmers would like reports with a range that lets them make decisions about growing crops.
In some sense, all would like very long term predictions– but not really any more than anyone else. And anyway.. they can’t check the accuracy of forecasts.
angech,
“In none of them did they consider altering the climate sensitivity to CO2 which makes all the models run hot.”
.
You keep saying this, but as several people have pointed out, it is just mistaken. Sensitivity to GHG forcing is an emergent property, it is not something that is “adjusted” by modelers. The emergent properties of the models are influenced by a host of modeling choices, which interact with each other and all influence the model’s behaviors, including diagnosed sensitivity. Sensitivity is NOT something which can be adjusted with a control knob. There are modeling choices which can influence sensitivity, but no direct controls.
SteveF (Comment #133675)
” angech,
* In none of them did they consider altering the climate sensitivity to CO2 which makes all the models run hot*. Sensitivity to GHG forcing is an emergent property, it is not something that is “adjusted†by modelers. ”
I beg to differ. I agree Climate Sensitivity to GHG forcing can be an emergent property. Steve Mosher has explained that it can be derived from TOA. When you look at different years different CS “emerges” for those years.
That does not stop it being used in models. As in, doubling of CO2 causes x increase in temperature plus positive feedback factors y minus aerosols z is incorporated into all these models directly and indirectly.
To say this does not happen is to denigrate all climate models. They have to put in temperature response to CO2 increase and positive and negative feedbacks in the formula which is the Climate Sensitivity.
I may be deliberately misreading or I may be unintentionally misunderstanding these comments in Carrick’s link but most of them show or imply Climate Sensitivity incorporated into models.
*One of them even says 1) …… modelers somehow changed their climate sensitivities*.
“Controlling the Global Mean Surface Temperature
. There are at least three reasons why abiding to this ideal need not be successful 1. Climate models may not exactly conserve energy. 2. The climate sensitivity of the model to the various forcings may not match the real climate system, and the forcings themselves may be erroneous.
We speculate that the fact that the bulk of models exhibit positive TOA radiation imbalances, and at the same time are cold-biased, is due to them having been tuned without account for energy leakage.
3.4. Climate Sensitivity
] Equilibrium climate sensitivity to a doubling of CO2 is a standard measure of models’ sensitivity to external forcing. We have performed simulations of the parallel worlds coupled to a 50 meter deep mixed- layer ocean with pre-industrial levels of CO2 and with doubled CO2 concentration
World 2 and 3 both have adjusted CO2 -forcings which are 10–20 percent higher than the standard ECHAM6 model
the higher sensitivity of World 3 relative to the standard model is mainly a consequence of a stronger adjustments to CO2 -forcing
Models that have a high climate sensitivity tend to have a weak total anthropogenic forcing, and vice-versa. Rational explanations are that 1) either modelers somehow changed their climate sensitivities, 2) deliberately chose suitable forcings,
More than a century ago it was first realized that increasing the
atmospheric CO2 concentration leads to surface warming [Arrhenius, 1896], and today the underlying physics and feedback mechanisms are reasonably understood (while quantitative uncertainty in climate sensitivity is still large). Coupled climate models are just one of the tools applied in gaining this understanding [Oreskes et al., 1994; Bony et al., 2012].”
As said I may be deliberately misreading or I may be unintentionally misunderstanding these comments. The first in fact may be about climate sensitivity derived from the model rather than incorporated in the model though I would argue that the parameters used are fitted with assumptions that CO2 doubling will cause a more generous response in them [positive feedback] than one gets from the straight Arrhenius theory.
The other comments disagree with your statement
“2. The climate sensitivity of the model to the various forcings may not match the real climate system
the higher sensitivity of World 3 relative to the standard model is mainly a consequence of a stronger adjustments to CO2 -forcing
1) either modelers somehow changed their climate sensitivities, 2) deliberately chose suitable forcings
Coupled climate models are just one of the tools applied in gaining this understanding with increasing the
atmospheric CO2 concentration leads to surface warming the underlying physics and feedback mechanisms are reasonably understood.”
Angech,
Yet you agree.
Looking for a good argument with someone? 🙂 I’m like that sometimes.
[Edit: I’m not merely being obnoxious or trying to be funny exactly, I’m pointing out that you and Steve have no disagreement in particular. I wonder if a phrase or two is being or has been misconstrued]
angech,
Climate sensitivity is an emergent property in models. Modelers do change it– by changing other things which cause a new sensitivity to “emerge”. There is no direct climate sensitivity knob in the model.
Kenneth Fritsch,
Were you aware that there is a 2,000+ year GCM control run from NOAA’s GFDL CM2.1 model? Wittenberg 2009 looks at the ENSO like behavior from this run. Science of Doom just put up a post on this. One conclusion is that if ENSO does behave like the model, even 150 years of observation are nowhere near sufficient to characterize the behavior statistics. I’d love to see what the PDO and AMO indices calculated from this data look like.
Gavin has a new paper stating clearly that there are knobs and that they tune (or fiddle or fit or adjust, whatever) quite a bit.
http://pubs.giss.nasa.gov/docs/early/Schmidt_Sherwood_1.pdf
Thanks Craig.
I can’t decide if I’m being B.S.’d here or not.
What do you folk make of this?
[Edit: I know, sorry; this is off the topic of parameters and tuning and such. But I am genuinely curious about this. That and I feel my comments are somehow incomplete unless I edit them. 🙂 ]
After kicking it around awhile, I think it’s sort of a dodge.
If the climate models are skillful enough for some useful purpose, then that’s all I ask really. Can they help me make useful and accurate predictions about something I care about in reality. If they are skillful enough, they’ve got value. It doesn’t make them ‘right’, or … it doesn’t tie back to the falsification issue all that well.
At the end of the day, if it works, and to the extent it works, it’s right. To the extent it doesn’t work it’s wrong. None of this justification or apologetic changes this. You can call it science and bless it with philosophical approval oil, but what difference does that make if you still get the wrong answer?
From Schmidt et al., “Configuration and assessment of the GISS ModelE2 contributions to the CMIP5 archive”, JAMES (2014):
Just to note that the tuning space is not limited to numerical coefficients in equations, but to choice of algorithms as well.
“Closely enough”? Skeptic’s complaint is that the models have insufficient skill at predicting the metric modelers have claimed is the most robust one. The metric highlighted in the IPCC reports and so on. This isn’t a nit-picky, I’m looking at cloud-cover in zimbabwe and find it doesn’t match well issue!
A needed long term prediction: Will it be possible to ship thru the Arctic between Asia and Europe and to the Russian north.
Given that special purpose ships would probably be needed with icebreaker escorts, and that supplying the Russian north is an important issue to Russia, this is a simple, practical example of where a GCM could contribute.
Eli+Rabett (Comment #133686)
“this is a simple, practical example of where a GCM could contribute.”
I’ll bite. Why a GCM when you just need a local Arctic model?
Lots of which already exist.
Why icebreakers when you predict there will be no ice in the long term?
Lucia (Comment #133678)
angech,
Climate sensitivity is an emergent property in models. Modelers do change it– by changing other things which cause a new sensitivity to “emergeâ€. There is no direct climate sensitivity knob in the model.
Perhaps I have been reading too many Skeptical science posts.
In this one “How sensitive is our climate?
Examining Past Temperature Projections”
it says.
In 1988, NASA climate scientist Dr James Hansen produced a groundbreaking study in which he produced a global climate model that calculated future warming based on three different CO2 emissions scenarios labeled A, B, and C (Hansen 1988). Now, after more than 20 years, we are able to review Hansen’s projections.
**Hansen’s model assumed a rather high climate sensitivity of 4.2°C for a doubling of CO2.**
His Scenario B has been the closest to reality, with the actual total radiative forcing being about 10% higher than in this emissions scenario. The warming trend predicted in this scenario from 1988 to 2010 was about 0.26°C per decade whereas the measured temperature increase over that period was approximately 0.18°C per decade, or about 40% lower than Scenario B.
Perhaps the newer models do not assume climate sensitivity, just the older ones?
The way I read it other things or parameters are changed as you said which in effect mandate that the model will disgorge a high temperature response in response to CO2 increase. The property we use to describe this response is the climate sensitivity which emerges from this biased input.
Describing it solely as only an emergent property from the model is fine. You are right, lots of others do it.
I thought I could describe this property “climate sensitivity “in the same way as James Hansen.
That I could say climate sensitivity was programmed into these models and was too high.
So what can I call the collection of forcing parameters in response to increasing CO2 that is inputted into models to produce the emergent CS phenomenon?
“There is no direct climate sensitivity knob in the model.”
If not, does not having 3 or 30 smaller knobs that indirectly do exactly the same thing refute this.
if x,y,z = a. Can I not say that the emergent property is too high because it was inputted too highly.
angech
No. You are making the mistake of thinking SkS’s language is precise. Hansen’s model did not “assume” a climate sensitivity. It’s climate sensitivity was an emergent feature of his model.
Hansen didn’t write the SkS post.
The forcing parameters don’t produce the emergent Climate Sensitivity. Other parameters do. The combination of forcing and climate sensitivity do affect the temperature response.
You can say you think the emergent property is too high. You can say it is too high because other parameter choices made it too high. (You would need to justify this– but it’s a claim that can be made.)
But climate sensitivity is not “inputed”. It’s not in any sort of “input” file. It’s an “output”. Other things are input.
You can’t call an “output” and “input”– and you can’t call it that even if the output is caused by the input. It’s stull an “output”.
Eli–
Theoretically, customers could exist. But those potential “customers” can’t know if a prediction is works until the future being predicted arrives. And, more over, those are not the “customers” funding the development of models. So, those ‘customers’ aren’t customers– because they don’t actually “buy” the product.
In contrast, customers for weather predictions can tell if predictions worked last month– a few months ago and so on. The reason they can tell is the predictions were available before the weather arrived, they could see how well they did. And they can use them for future planning now.
Weather forecasts have honest to goodness paying customers– and not just government funding agencies.
Lucia and SteveF thanks for the explanation that climate sensitivity is an output only, looks like I am out on a limb.
Will repost on this only if I find a solid support somewhere.
Seems to me the main ‘customers’ for climate model projections are groups, politicians, and individuals who want to justify the large cost (political, societal, and economic) of immediate drastic reductions in fossil fuel use. Any climate model which projects less than very rapid warming is obviously ‘not fit for purpose’. The often heard comment that reducing fossil fuel use is “the right thing to do anyway” pretty fairly sums up what these customers are looking for. The customer is always right, as they say.
Well, Lucia does these comparisons here of course.And start dates do matter because of variability. But the last I checked, for the MSU era, observations were all less than Scenario C:
http://climatewatcher.webs.com/SatelliteEraTemperatures.png
And Scenario C was the one in which CO2 emissions went to zero in 2000. This means that doing nothing ( BAU ) was more effective than doing everything ( ceasing emissions entirely ) was modeled to be. Any trend can have eventual consequences over a long enough duration, but it’s clear global warming has been exaggerated.
Mark: Gavin’s concept of Popperian falsificationism is absurd. Predicting antimatter is a strong prediction and finding it a good confirmation. For more complex systems we need to look for degrees of agreement. In this regard, the modelers have been way too forgiving of their models and claim even the vaguest similarity of model and output is a wonderful thing, like the proud parents that put the kiddies art on the fridge and proclaim Picasso lives there. To claim (as Gavin does) that the models should only be compared to a naïve baseline is also absurd. Such a comparison only shows the models are better than nothing, not that they are useful. If my model predicts correctly 50.01 percent of the time compared to a random expectation, this does not mean it is useful or statistically different from 0. Getting vaguely good agreement may be encouraging and indicate that you are on the right track scientifically, but it does not mean you should get to make gas $10/gallon or shut down all coal plants or restrict the number of children people have or stop poor countries from getting electricity (all things being proposed or done already).
Thanks Craig.
I find it hard to get a good handle on the scope when I look at arguments like this, because I have a very particular / specific personal view regarding science and why science is valuable. In my world, the only darn reason I care about science at all is that science explains and thus allows comprehension and prediction of parts of the world around us. It lets me make things that work and lets me figure out why other things don’t work. It follows naturally that I don’t find much merit in Gavin’s argument, because his argument largely seems to relate to how we label these poorly performing models and whether or not we give them time and interest for reasons other than how well or poorly they perform.
So the value I find in science is that of somebody who cares about certain practical applications of it. But this is probably not the only legitimate perspective out there. What angles am I not approaching this from exist from which Gavin’s argument might have merit, I guess is what I’m wondering about.
If I was a theoretical scientist, would my perspective be different? Etc. If there are further cases besides that one anyway. Is there a perspective from which this makes good sense?
Lucia,
Yes. I thought this was a pretty weak argument too. I didn’t start wondering what was wrong with the projections because of the modeling of rainfall in Fiji; it was a little more blatant than that. Maybe AR4 SPM’s first paragraph [edit: first several sentences] had something to do with it. I don’t feel like it’s nit picking in any meaningful sense of the phrase to call that wrong.
Lucia, what do actuaries do?
They get fired when they’re wrong.
Eli,
If you think that question is supposed to make a point, just try to make it. The rule for rhetorical questions at this site is: You may only ask them if you volunteer your own answer in the same comment where you ask the Q.
If it’s just some strange pointless off the wall question…. well, I advice a lighter hand when pouring fermented carrot juice.
Re: lucia (Comment #133689)
Dr. Liljegren,
I have a question related to your comment. I accept your statement that climate sensitivity is an output of the Climate Models. However, do the modelers not include as part of the Climate Models a calculation that a specific rise in atmospheric CO2 concentration will cause a proportional rise in the average atmospheric temperature, ceteris paribus?
Will J. Richardson, excuse me for butting in… The short answer is “yes”.
The models include well-understood radiative physics that accounts for the reduction in the radiative heat loss associated with an increase in atmospheric CO2 concentration. While this results in an increase in atmospheric temperature, all things being equal, the actual change in atmospheric temperature is more complicated, because all things are not equal.
Estimates put the amount of warming from this effect at around 1.1 °C/doubling of CO2, but it doesn’t take a full blow climate model to compute this effect (as long as you stipulate ceteris paribus).
Secondly, if you increase the temperature, the absolute moisture in the lower atmosphere increase. This also reduces radiative heat loss and also tends to warm the atmosphere (all other things remaining the same).
Again you don’t need a full blown model to estimate the sensitivity associated with this. If memory serves me, they get somewhere around 2.2-2.5 °C/doubling of CO2.
To get the “actual” sensitivity requires a full blown climate model, which involves a number of other feedbacks including cloud feedbacks and the biosphere response, most of which are modeled very imperfectly, if at all.
Hope this helps.
Carriock,
I suppose this is a bit like the blind man’s attempt to describe the elephant he’s encountered, but I have some questions.
Do models run with reference to a randomizer (random number driven function)?
I think many of us have finally realized that there is not a list of variables or coefficients at the head of the code which can be changed to tune. It is much more than “whistle a few bars.” Tuning must include modifications to the actual construction of the model – to the mathematical functions and the balance among them.
As you can see, I’m wrestling with how they must work, not how they do work, which may be more productive for me, than trying to wade through code which included math well beyond anything I could absorb.
I’m especially interested in the randomness component if there is one, how it is incorporated in the code (what is randomized?) and how the magnitude of its effect is gauged, well, tuned?
John Ferguson,
The models are, as far as I know, completely deterministic. That is, from a specified initial condition, a model will generate one specific trajectory which is ‘temporally chaotic’, meaning that it evolves in patterns which are related to each other, but which never exactly repeat. If you start over from the identical initial conditions, then the trajectory will be exactly repeated. If the initial conditions are modified even a tiny amount, then a completely different and unique trajectory develops; similar in pattern, but never exactly the same. So when a model is going to be subjected to increased GHG forcing, the modelers start with some set of ‘control’ conditions, then allow the model to ‘spin up’ to some ‘equilibrium-like’ state over many ‘model years’ before increasing the GHG forcing. Initial conditions could be randomly generated prior to ‘spin up’ I suppose.
Whether adding some random noise to the calculations, to simulate chaotic variability in sub-grid scale processes which are only “parameterized” would make a more realistic model is an interesting question.
Carrick,
I think there is some uncertainty in the water vapor amplification, but my recollection is that the best estimate is a bit under a factor of two, yielding a net sensitivity near 1.9 to 2 per doubling.
SteveF,
MODTRAN Tropical atmosphere, constant RH, 17km (tropopause) looking down, clear sky, all other settings are default. Increasing CO2 from 280 to 560 ppmv requires a surface temperature increase of 1.92K to restore upward flux at the tropopause to the 280ppmv value. That does not include lapse rate feedback, which would be negative. IIRC, every other atmosphere require lower surface temperature increases for CO2 doubling than for the Tropical atmosphere. Clouds, at least those included in MODTRAN, don’t make much difference either.
DeWitt,
Thanks. Do you have any idea of the net sensitivity range for cooler surface temperatures?
Dewitt,
I fooled around with the Modtran calculator. For constant relative humidity and clear skys in the subarctic, doubling the CO2 leads to a surface temperature increase of under a degree on average. I did not explore what all the different sky conditions do (different kinds of clouds, rain, etc) to the calculated sensitivity, but it looks like the tropical response you noted (1.9K) represents a global maximum.
.
The (warming) devil is in the (amplification) details.
SteveF, about a year ago, I posted this at Climate Etc:
http://judithcurry.com/2013/12/26/seasonal-radiative-response/
The global average temperature (GAT) varies quite a bit from January to July every year and when the GAT rises 2.5C it corresponds to the putative 3.7 W/m^2 increased outgoing longwave radiation (OLR). This includes all skies and a strong water vapor covariance.
“Real Climate
Absolute temperatures and relative anomalies
— gavin @ 23 December 2014 As mentioned above, the main problem is that the global mean temperature is very much an emergent property. That means that it is a function of almost all the different aspects of the model (radiation, fluxes, ocean physics, clouds etc.).”
So there is no actual absolute temperature along with no actual climate sensitivity because they are both only emergent properties of models.?
Wikipedia says
” For coupled atmosphere-ocean global climate models (e.g. CMIP5) the climate sensitivity is an emergent property: it is not a model parameter, but rather a result of a combination of model physics and parameters. By contrast, simpler energy-balance models may have climate sensitivity as an explicit parameter.”
A climate model models the Radiative forcing due to doubled CO2.
“Climate sensitivity has a component directly due to radiative forcing by CO2, and a further contribution arising from climate feedbacks, both positive and negative. “Without any feedbacks, a doubling of CO2 (which amounts to a forcing of 3.7 W/m2) would result in 1 °C global warming, which is easy to calculate and is undisputed. The remaining uncertainty is due entirely to feedbacks in the system, namely, the water vapor feedback, the ice-albedo feedback, the cloud feedback, and the lapse rate feedback”;[12] addition of these feedbacks leads to a value of the sensitivity to CO2 doubling of approximately 3 °C ± 1.5 °C, which corresponds to a value of λ of 0.8 K/(W/m2).”
So the implied Climate sensitivity which is fed into the models is 0.8 K/(W/m2) or a rise in temp of 3 degrees centigrade due to the parameters chosen.
Those parameters fed in are too high giving an emergent CS which is also too high.
SteveF, look at Ramanathan 1981, “The Role of Ocean-Atmosphere Interactions in the C02 Climate Problem”.
In particular see Figure 4. This gives an increase of 2.2°C/doubling of CO2.
I realize this is more complex than just holding relative humidity constant (as is typically assumed). But it is probably more accurate than MODTRAN type calculations too.
I’ve seen it. I don’t have anything nice to say about it. I don’t even like the color choices.
Carrick,
The problem is that paper introduces moist convection and it’s influence on tropospheric profiles without noting the huge influence of clouds on both solar and upward IR fluxes. It introduces some “feedbacks”, while ignoring others. The 2.2C value doesn’t seem to me to be at all a fair value for the no-feedbacks radiative flux. The article also notes in its conclusions that land warming “may” (weasel words already in 1980!) not precede ocean warming due to the strong ocean influence described in the paper. It didn’t work out that way, which is not much of a surprise considering the selective use of amplifying factors in the paper.
.
I did find the article informative though. It is pretty clear that battle lines were already drawn between modeled and empirical estimates of warming, with some of the usual suspects on both sides. The case for alarm was already being advanced using modeled warming. When I read an early paper like this one, and then compare the “likely range” for climate sensitivity in the Charney report with that given in AR5, I think one can fairly say that there has been little real progress on the most critical question related to GHG driven warming over 30+ years and hundreds of billions of public dollars spent. Empirical estimates of sensitivity today remain far below modeled estimates. Modelers continue to offer excuses/explanations for why the empirical estimates must be wrong and the (much higher) model estimates right. I think the paper you linked to helps explain why: even 35 years ago the modeled results were being hyped to support reductions in fossil fuel use. It is the damaging influence of green politics on science, beginning to end.
Carrick,
I’m less than impressed with the argument in the paper that any increased formation of low level clouds from higher humidity and the consequent increase in albedo is only a local effect and can be ignored globally.
I’m going to move the comments about the “yellow/purple” blogs to the new thread. 🙂
Venturing further off-topic, if such is possible on an open thread, you might want to check out the video mentioned at BH. Basically, SkS regurgitation.
Ugh. Emergent property.
I think I get it. There is no “input forcing” that is named “climate sensitivity to carbon dioxide set any number you wish here and the model will produce the preferred policy supporting output”.
Yes. Literally.
I also get that modelers very likely comprehend perfectly well how to indirectly adjust the internal parameters to move this indirect emergent property up and down, and to pretend they have no knowledge of this and this is “simply physics” that cannot be altered less Einstein’s theories would be voided is what I am calling BS on. Nobody has run a model that produced this emergent property that resulted in a low climate sensitivity? Impossible? They just aren’t smart enough to comprehend how to do this? One isn’t interested in doing these “edge” cases to see how well they might compare to future observations? Our detached climate modelers don’t comprehend the implications of producing an emergent low climate sensitivity model that better matched recent observations? Two decades of lower than expected observations says absolutely nothing about the probabilities of the high end of sensitivities in the PDF? Are we to assume that two decades of higher than expected observations would not result in expanding the long tail?
So adjusting the models this way will likely result in less than perfect hindcasting. Perhaps the assumed historical forcings in use are also less than perfect and produce a set of supposedly independent models that are much more homogeneous than advertised due to this restriction. Possibly the historical forcings have forced modelers hands to higher emergent properties?
So yes there is no input called sensitivity in the models. We are also treated to magazine covers with an earth with a huge dial on it named carbon. Ugh.
Tom+Scharf (Comment #133752) thanks for putting it so eloquently. The battle of semantics beats the battle of physics still.
Tom Scharf,
The modelers are human, and subject to the same human foibles as the rest of us. Do they keep one eye on their model’s diagnosed sensitivity when making ‘optimizations’? Almost certainly. Could they tone down climate sensitivity and projections of future warming by adjusting some assumptions and parameterizations? Of course they could. Will they? Only kicking and screaming. These are folks with a huge vested interest in maintaining the threat of extreme warming; their resistance to changing the paradigm is perfectly understandable.
.
But there has been some progress. A few formerly higher sensitivity models have in fact been toned down (the GISS family of models, for example). The GISS suite now covers 2.4C to 2.8C per doubling, a car cry from James Hansen’s 4C per doubling, and not so very far from the empirical estimates by Nic Lewis and others. With continued relatively slow warming, pressure will grow for the modeler’s to revise their assumptions and parameteriztions and bring the models more in line with reality. The process will be slow and halting of course… but reality can not be forever ignored. It is inevitable there will be some idiotic policies implemented (California raising taxes on fuel, ever more subsidized solar panels in cold cloudy places), but these will make no real difference in total emissions, and in twenty years time the divergence between models and reality will pretty much demand that the models be broadly revised…. or risk being ignored.
SteveF, my main reason for pointing to the paper was that there are models that include considerably more physics than e.g. MODTRAN which arrive at higher sensitivity numbers for direct CO2 + water vapor feedback (however, because Ramanathan does have a limited treatment of the adjustment of the vertical atmosphere to changes in forcing, it is of course a bit more than just that, though it is the change in water vapor feedback processes than dominate the response in his model).
I think Ramanathan is genuinely attempting to arrive at what he thinks is the truth. I am pretty skeptical of the notion that he’s part of some alarmist movement, in 1981 especially, intending to generate some massive increase in funding for alarmist-based climate science, by publishing a result suggesting a lukewarm 2.2°C/doubling of CO2 sensitivity.
As the paper states, this is only a 1-d 17-layer model (2 ocean layers + 15 atmospheric layers). Clearly it is not trying to capture all of the physics of a full GCM here. Nor is there actually a serious attempt to correctly model cloud physics—that clearly is well beyond the scope of a 17-layer 1-d model. But neither are e.g. the two layer models that give rise to estimates of 1.1-1.2°C/doubling of CO2. And, clearly, neither is MODTRAN. 

Anyway, it’s not worth overanalyzing this paper, nor being too skeptical of it. There aren’t many fudge factors, and while I might, with the weight of all of the papers published since then to support me, perform the analysis differently than Ramanathan did, I do think this was a good effort at understanding the physics behind climate change, even if we can’t expect the numbers that come out of his simple model to be particularly accurate
I believe SteveF is exactly right. One would have to question the intelligence of the modelers if they didn’t know broadly the effects of their parameters on things like sensitivity, and there are many, many parameters. Turbulence modelers for example, do know how their parameters affect thing like boundary layer thickness and response to adverse pressure gradients. They often have various versions with different values for these “emergent” properties.
Carrick,
Fair enough, there is a wide range of complexities which can be used to calculate atmospheric heat transfer, and the more complex, the more the results depend on uncertain parameterizations. My technical objection to the paper is that it invokes moist convective transfer, but appears to ignore the potentially large influence of the clouds that moist convection inevitably produces. My overall impression of the paper is that it was written at least in part to counter lower published estimates (and these are specifically referenced by the paper). Only a few more decades of data will (empirically) narrow the plausible sensitivity range and warming projections; the models, as currently implemented, just can’t do it.
SteveF:
I think there is value in analyzing different sub-processes that are associated with climate feedback, even if you can’t consider all of them: As you know, the influence of clouds is not likely to be accurately modeled using a 1-d model.
I think we could view this paper (and I do) in the same light as estimates of the 1.1°C/doubling of CO2 from direct CO2 forcing. This paper provides insight into the magnitude of important climate feedbacks, though not all of them, and I personally have no problem with separating out the water vapor effects that don’t include a change of phase (other than surface vaporization, which is relatively easy to model) from those that do and are difficult to accurately model.
The 2.2°C/doubling of CO2 number I would regard as the number we might get, were there no feedbacks from clouds. To me, this suggests the amount of heavy lifting that is needed to shift the model from this nominal 2.2 number away to lukewarm value like 1.8 or high sensitivity like 4 °C/doubling of CO2. It’s not a big stretch to suggest that with feedback we could get 1.8°C/doubling with all feedbacks accurately modeled, but it seems a bit dubious to suggest we could go from 2.2 to 4°C/doubling just by adding cloud feedback.
So if anything, this result is a much bigger problem for very high and very low sensitivity models than it is for modestly more, or modestly less, sensitivity models.
I’m not sure this was a motivation of the paper—there is after all nothing wrong with critically examining the as then existing estimates of climate sensitivity—but the papers that had very low climate sensitivities are almost certainly wrong and the criticisms leveled at them by Ramanathan appear separately valid.
Anyway, I see not including sectors that contain large modeling uncertainty like cloud feedback a strength of this paper, not a weakness, as long as you don’t try to interpret the results as an estimate of total sensitivity. I don’t think Ramanathan actually viewed this paper as an attempt at the total ECS, but including language that cautioned against interpreting his results that way would have been helpful.
I think the model is also useful for studying the inverse problem (that is inferring ECS or TCR from data). Were I to attempt the inversion process, I could see starting with a model such as this (though some effort would need to be made to include a parameterization of clouds), which clearly does a better job with the physics than simplistic energy balance models.
Given the discussion of tuning and how models are built I figured I would point out a talk given by Steve Easterbrook on some of the “software engineering” aspects of climate model development (http://research.microsoft.com/apps/video/default.aspx?id=115276). During the first part of the talk Steve gives a pretty hard sell supporting the “consensus” view that climate models are very trustworthy. I didn’t find it very convincing but it does have some interesting background on the process. There is also a, in my opinion more interesting, series of QA towards the second half where he mentions a couple of things that caught my attention.
The first is the description of a bug which was causing problems with soil moisture levels which was managed for a while by “adding some extra tuning to the model” so in fact “tuning” to get the right answer. Next is that when this bug was finally fixed is was done by looking at 5-6 other models and finding that all but one of them had the same bug, indicating that there is much less model independence that is often implied. Finally, there is an interesting side discussion on how the models are so interconnected that swapping in new modules (ocean modeling) is basically impossible, they are written using mainly Fortran77 (no surprise), and have huge code growth/churn which is a huge red flag for a model that is probably broken in many ways but through continuous tweaking is rebalanced to produce the “correct” result.
I have continued my effort to look behind the curtains of the science wizards of climate modeling as I have with paleoclimatologists temperature reconstructions and primarily as a matter of feeding my curiosity about the details of the science records that are not published. As a layperson I suspect that level of science that goes into the climate models is at a higher level than that practiced by most paleoclimatologists. The details are, however to me, a matter of sometimes being lost in the blackbox, which in the case of paleoclimatology is avoidable, while for climate modelers not so much.
With the CMIP5 climate models I have found differences in the noise levels and types of noise that would appear to this layperson to be a reason for suggesting that at least some of the models have been dead wrong on portraying the climate. That might not be a general condemnation if the model skills could be separated into a noise part and the part that is probably more important if the models are to taken seriously in estimating future temperatures and climates and that is the part that would relate to a models calculated and emergent equilibrium climate sensitivity (ECS) and transient climate response (TCR). These quantities have taken on added attention since the advent of published papers on methods of estimating ECS and TCR using observed data, and, of course, the inevitable comparison with models. To that end I have been attempting to determine how the CMIP5 climate model ECS values published in the IPCC AR5 review were calculated. After some discussion with PaulK here at the Blackboard and doing some of my own research, I found the method used for the published ECS in the AR5 review.
The link to the Gregory paper describing the data used and calculations of ECS from an abrupt 4XCO2 burst model experiment is given below and followed by a link to the AR5 chapter 9 page 817 reference to it and with ECS calculated values.
http://www.readcube.com/articles/10.1029/2003GL018747?
https://www.ipcc.ch/pdf/assessment-report/ar5/wg1/WG1AR5_Chapter09_FINAL.pdf
The method for ECS calculation involves regressing the changing surface temperature against the net top of atmosphere (TOA) radiation deficiency over at least 140 years after the 4XCO2 burst and extrapolating the total change in surface temperature at which the TOA radiation is a net zero. The data for doing this calculation are given at the following link. I used the monthly values for the models for rlut, rsdt and rsut which are TOA outgoing long wave radiation, TOA incident short wave radiation and TOA outgoing short wave radiation, respectively. The net value used in the regression was obtained by subtracting the sum of rlut and rsut from rsdt . Also required to adjust the 4XCO2 surface temperature series (tas) was the Pre-Industrial Control (PIC) for each CMIP5 model. That data was taken from KNMI. All data were converted to annual for doing the regressions.
http://cera-www.dkrz.de/WDCC/ui/EntryList.jsp?acronym=ETHc2
The downloading of these data was a bit laborious and I have to date finished calculations for 19 model runs which are listed in the linked table below. The table shows the results of my “normalizing” the temperature data 3 different ways. First I merely subtracted the first year’s temperature from the remaining years to obtain a measure of temperature change, secondly I used the first years of the PIC corresponding to the years of the model derived 4XCO2 temperatures and subtracted those PIC temperatures from the 4XCO2 temperatures and finally third I did a subtraction as in the second method except I used the corresponding final years the PIC for the subtraction. The table also includes the published AR5 ECS values for the model where it was available. As can be seen from the table the correlations are high for all three subtraction methods with the third giving the highest values. That third method is probably the one closest to that recommended by the authors but I was unsure on reading the paper. Of interest here is that the ECS calculated values are sensitive to the values being subtracted and further it needs to be noted here that the PIC series can have randomness and/or drift that lends to uncertainty in the ECS results. My calculations agree reasonably well in most cases with the AR5 published ones although the match is far from perfect and are much higher for 2 model runs. I have three links below that show 4 typical plots for the ECS regressions and the PIC series for the models used.
I plan to complete the CMIP5 ECS calculations for all model/model runs with data available and determine why my calculations are not in better agreement with those published in AR5.
Link to table ECS regressions:
http://imagizer.imageshack.us/v2/1600x1200q90/661/EmIV0V.png
Link to ECS regression plots
http://imagizer.imageshack.us/v2/1600x1200q90/912/VzPchg.png
Links to PIC series:
http://imagizer.imageshack.us/v2/1600x1200q90/905/Y4lakv.png
http://imagizer.imageshack.us/v2/1600x1200q90/540/Ul5cS2.png
Kenneth, I’d suggest trying a robust estimator in place of linear regression, like total-least squares or Deming regression.
My guess is there is noise in the x-axis as well as the y-axis, for which OLS is known to produce biased estimates.
Carrick (Comment #134121)
Carrick, that point was noted in the link to the Gregory paper and the authors tended to discount the magnitude of the temperature series noise and its effect on OLS, but since I have the data in good form I agree that total least squares should be also used for calculating ECS.
Carrick, I used the gx.rma function in R library(rgr) to do the total least squares regression (TLS) on TOA deficiency versus 4XCO2 temperatures for the 19 CMIP5 model runs I worked with in my previous post. I did the TLS regression using all the available data and again using the first 30 years of data. I was concerned with the clustering of the data after thirty years because the TOA deficiency and temperature change only very slowly after 30 years from the initial abrupt 4XCO2 burst. I also redid the RLS calculation (which will differ a little from my previous table results because I used different PIC values) using all the data and the first 30 years of data. The linked table shows the results for all 19 model runs with the AR5 published ECS.
The results show that in general the TLS regression produces higher ECS values than RLS and values further from the AR5 published ECS values than those derived from the RLS regression. All these results show, in my view at least, that the ECS values produced are dependent on method used for calculation and rather more squishy than the published ECS values might otherwise appear. I need to link a table with the CIs for the intercept from the regressions that produces the ECS values and further determine all the factors leading to uncertainty in the results.
I also know that there are more data available for calculating model ECS values than those published in AR5 and am wondering if this has to do with the data not being available in a timely manner of if there is some other reason. I should also request from the IPCC the details of their calculations, but would not hold my breath for a reply based on my previous requests.
http://imagizer.imageshack.us/v2/1600x1200q90/633/auXZkg.png
Kenneth, interesting work. When you have errors in variables, it is a trickier problem to solve robustly, than when you don’t.
But in general, if you don’t account for the errors in variables, you expect an attentuation bias in your estimate of the regression parameters. That is your estimated value divided by the actual value is always less than 1. I think this is why you see TLS to be higher than RLS.
Since I view TLS to be more reliable for this case, for me, the take home is that the models are doing even worse compared to data than the AR5 writeup suggests. This assumes your results stand up to scrutiny of course.
“This assumes your results stand up to scrutiny of course.”
Something which would not be a concern or less of one (as it always is for me without confirmation) if detailed accounts of how the AR5 published ECS values were calculated were available. I have to assume that the AR published reference to the Gregory paper and its description of the method is correct. The AR5 chapter 9 and the Gregory paper refer to 2 other methods of estimating the model ECS which are a slab method where there is no exchange with the oceans and equilibrium occurs quickly in a matter of decades after the 4XCO2 burst and the other where the model is run for a long time after the 4XCO2 burst and the ECS calculated from the forcing and deficit in the TOA radiation. The first method is criticized for not being the coupled model that needs to be evaluated and the second for not necessarily accounting for all the feedbacks. The ideal method and very expensive one I have been told would be – given no model drift or at least a capability of accounting for it – running the model for thousands of years after the 4XCO2 burst and tracking the surface temperature. Unfortunately after doing some literature searching I was not able to find ECS values derived from the CMIP5 model runs with these alternative methods.
I think the TLS CI range for the intercept (ECS) will be larger than that for RLS and need to take a look at it.
I found a more current paper by Andrews and Gregory on the OLS regression used for the CMIP5 published ECS values and have linked to it below. The paper shows some plots of the regressions which have allowed on comparing with my plots (using OLS and not TLS) to determine that the differences in the derived ECS values between their results and mine are from offsets in the values of the TOA radiation deficits used. The offset varies from model to model. This will allow me to more readily determine the source of the difference.
Interesting that this paper’s authors recommend concentrating on the TCR over the ECS for which they have provided the official method of calculation. I suspect their reservations with ECS has to do with potential linearity failures beyond the 150 years used in the abrupt 4XCO2 experiment and ECS calculations. Also the TCR value is a better predictor of what the temperature will be in 100 years going forward whereas for ECS to be fully realized takes thousands of years. That surely makes TCR more appropriate for policy for the next 100 years.
Carrick, my thought that the confidence intervals for ECS would be larger for the TLS regression over OLS regression was not correct. TLS gave on average for the 19 model runs I have been using a spread of 0.14 and OLS gave 0.27. Also the differences I see between my OLS regression and that same regression used by Andrews for the CMIP5 model ECS values should have no bearing on the difference I see between OLS and TLS regression, i.e. that difference should also be seen using the Andrew and Gregory data. Thanks, Carrick, for suggesting that I look at TLS. I wonder why Gregory and Andrews did not do the same. They talked albeit briefly about the noise in the both series being regressed and the arbitrary decision to use one or the other as the dependent variable.
http://onlinelibrary.wiley.com/doi/10.1029/2012GL051607/abstract
Kenneth, in my opinion, you can’t obtain a meaningful estimate of ECS with noisy data, if the response time of the climate model is 2000 years and you only have 100 years of data.
Isaac Held on his blog talks about transient versus equilibrium sensitivity . It should be fully comprehensible to you, now you’ve dipped your toes in the water. I certainly agree by the way that TCR is more relevant to climate policy than ECS.
It is without dispute an error to use OLS when you are regressing noisy series against each others, without applying a bias correction for the noisy in the independent variable.
As I said above, OLS will tend to bias the estimate of TCS and ECS low, and if the climate models are running too hot relative to data, will make the climate models look like they are in better agreement with the data than they really are.
Regarding tuning, this paper by Golaz et al. is interesting. From the abstract:
The December UAH global anomaly was +0.32 degrees, down 0.01 degrees from November.
Carrick, I think I may have to take a pause in my ECS calculations of the CMIP5 models and determine better why my results appear to be different than those obtained as part of the official record. Here I am talking about using OLS regression as was used in the official published versions. I have yet to find a publication (or review like AR5) that has a comprehensive listing of the ECS values from all the CMIP5 model/model runs that have data available for making the calculation. In fact I see publications of the distribution of CMIP5 model ECS values where the models used is far from comprehensive.
I went back and checked my data inputs and calculations for the 2 model results for which I found much higher ECS values than those published in AR5 Chapter 9 , i.e. BNU-ESM and CNRM-CM5, and obtained the same result as I had previously calculated. As I said before the differences in my results are that for a given model all the TOA radiation deficits are offset by the same amount across the plot of radiation deficit versus temperature and almost always upward on the plot, but by different amounts for each model compared. There are not differences in slope or even in the distribution of data points around the regression line. This is a somewhat frustrating development for me at this point in my analysis. I have used calculated TOA differences from the 4XCO2 CMIP5 experiment by subtracting the sum of the TOA outgoing long wave radiation and the TOA outgoing short wave from the TOA incident short wave radiation. If I were leaving out some unaccounted radiation at this point I do not see how it would be the same for every data point (year) over the entire series – as would be required to explain the differences in my plots and those published in the Andrew and Gregory paper that I linked in a previous post.
Carrick, I hope your links from Held’s blog might shed some light this offset difference I am observing.
Carrick, I contacted Tim Andrews by email telling him about my failure to calculate the same ECS values by his OLS regression method and asking for advice. He replied with 24 hours and told me that I needed to subtract out the TOA Net radiation of the Pre-Industrial Control (PIC).
When I did this my ECS values were much in line with the published ones, and where there were differences, the slopes matched very well, indicating that any differences I am seeing in ECS values were due to differences in the TOA PIC net radiation used.
The upshot of this is that some models maintain an excess or deficit over the entire time span in TOA radiation budget after a few hundred years of model spin-up under pre-industrial climate conditions . The amount of radiation out of budget varies from model to model and is reason I was previously getting large differences with some models from the published values. I find not only these differences surprising but also the the fact that all models do not trend towards a zero TOA radiation budget under PIC conditions.
I ask Andrews about using total least squares instead of ordinary least squares, given the error in temperature and radiation variables, but have not had that query answered yet. I also asked him about the publication of the ECS values for more models where data are available for calculation.
I am wondering now whether the full range of uncertainty is acknowledged in calculating the ECS values given the need to adjust the 4XCO2 temperatures and TOA radiations using the PIC values and recognizing the variations in those values.
Kenneth Fritsch, thanks. Interesting. Will you being posting your newer results?
So now we know that many of the current crop of GCM’s not only don’t get the GMST correct, they also don’t conserve energy.
In reference to a comment of mine about the lack of regional skill of GCM’s (getting the equator warmer than the poles does not demonstrate regional skill), Science of Doom had this to say:
Indeed.
Carrick (Comment #134406)
Carrick, I plan to gather all the data for the CMIP5 models required to calculate ECS values using OLS and TLS. And by the way the differences I showed previously between the 2 regressions methods will remain since all the subtraction of PIC TOA Net does is move the entire distribution up or down by a constant amount.
DeWitt Payne (Comment #134407)
Interesting conversation at that blog and about things I have contemplated about climate models. I should do a little more reading over there.
Kenneth Fritsch (Comment #134404) January 10th, 2015
I hope you might give the results of the ECS you calculated and could give an estimate of the ECS of the models if the TOA did not have any TOA radiation deficits. I.e. a zero TOA radiation budget.
Your higher initial results suggest a built in high ECS of the models or parameters set to show a higher emergent ECS. [Sorry, one of my hangups at the moment]
” I find surprising the the fact that all models do not trend towards a zero TOA radiation budget under PIC conditions.”
Since TOA radiation estimation is due to ECS estimation is the fact that these parameters are set high the cause of the TOA radiation budget not being able to trend to zero?
You said “given the need to adjust the TOA radiations using the PIC [by subtracting out the TOA Net radiation of the Pre-Industrial Control (PIC)”.
How can subtracting out “the TOA Net radiation of the Pre-Industrial Control (PIC)” give a downwards effect on the models unless it was negative in the first place as less radiation should mean less warming in the models in the first place?
Please feel free to point out the nonsense in these questions or ignore them if nonsensical but if there are parts that make sense could you clarify those points?
angech (Comment #134412)
“I hope you might give the results of the ECS you calculated and could give an estimate of the ECS of the models if the TOA did not have any TOA radiation deficits. I.e. a zero TOA radiation budget.”
I have just recently become interested in the details of the calculation of ECS for the CMIP5 models and thus my replies are from the perspective.
The regression of the 4XCO2 net TOA radiation and temperature (both adjusted by subtraction of the PIC net TOA radiation and temperatures) and then extrapolation of the regression straight line to zero net TOA radiation is supposed to do just what you evidently require. The equilibrium takes a very long time to occur given the heat turn over of the oceans. If that equilibrium occurred near simultaneous then the calculation of the ECS would be simple(r). For the AR4 IPCC review the ECS was calculated from ocean slab models were the ocean could not change its heat content. There equilibrium was assumed, I believe, after 20 years.
The Transient Climate Response (TCR) is what I believe you would obtain if you assumed no TOA imbalance or made provisions for it in your calculations. That value is normally published alone or with the ECS.
The CMIP5 models have a residual TOA radiation imbalance (some much larger than others) that as I observe it tends to remain rather constant over the spin up time period. That residual effect evidently carries over into the 4XCO2 experiment and must be subtracted out. In other words the extrapolation in the regression needs to be to that residual TOA radiation imbalance and not zero. Subtracting puts you at zero.
Why this occurs is the question I have put to Tim Andrews and hope to hear how he views this effect.
I can only edit my comments once and thus I need to add the embolden words to the sentence below.
The CMIP5 models have a residual TOA radiation imbalance in the PIC (some much larger than others) that as I observe it tends to remain rather constant over the spin up time period.
For those interested I have linked the 200 year time span plots of 36 CMIP5 model pre-industrial controls TOA radiation budgets below. The data were extracted from the KNMI Climate Explorer for the yearly mean global values for rlut, rsdt and rsut which are TOA outgoing long wave radiation, TOA incident short wave radiation and TOA short wave outgoing, respectively. The net radiation is calculated from the sum of rlut and rsut subtracted from rsdt.
http://imagizer.imageshack.us/v2/1600x1200q90/911/bLTwei.png
http://imagizer.imageshack.us/v2/1600x1200q90/908/hGFpiS.png
http://imagizer.imageshack.us/v2/1600x1200q90/909/QekGqw.png
I received a reply from Tim Andrews on my queries:
He said he had not looked at using total least squares regression for calculating ECS and had not looked at other CMIP5 model data available for calculating ECS values. He noted that it would not be unexpected for the models to have TOA radiation deficit/surplus after a few hundred years of spin up since the ocean effects are in millennia time scale. I think at this time I will merely thank him for his replies.
I think there at least 30 CMIP5 models with data that would permit calculating ECS values which is about a third more than I have seen published. I plan to download all these data and calculate ECS values using both OLS and TLS regression. What is of interest to me is that most of the plots of pi control TOA for the CMIP5 models, shown in the preceding post, do not appear to be even trending slightly towards zero. Also some models start and end near a balanced TOA energy balance while others can plateau for 200 years a values significantly removed from zero.
Kenneth Fritsch –
I’m pretty sure I’ve read the answer somewhere before, but I can’t recall what the answer was (or where I’d read it). For CMIP5 runs, if the model can not reach equilibrium (viz, no TOA imbalance) before the start of the forcing scenario, what actions (if any) are taken to mitigate the effects of the non-equilibrium initial condition?
Kenneth,
This is a strange new meaning of equilibrium: i.e. a return to a pre-existing TOA imbalance after a perturbation. So we’re supposed to believe that the lack of reaction to the spin up TOA imbalance doesn’t affect how the model reacts to ghg perturbation? To paraphrase the activists, it’s worse than we thought.
I recently finished downloading the CMIP5 data (21 giga bytes and 33 model/model runs) required to calculate the ECS by regression per the methods used for the published ECS values in the IPCC AR5 review (see links below from my previous posts and Chapter 9 page 818 for details). I used the ordinary least square regression (OLS) from the AR5 methods and total least squares regression (TLS) that was suggested to me by Carrick owing to the error in the both variables being regressed. The results of these regressions and the published AR5 ECS values are listed in the linked table at the bottom of the post.
The link to the Gregory paper describing the data used and calculations of ECS from an abrupt 4XCO2 burst model experiment is given below and followed by a link to the AR5 chapter 9 page 818 reference to it and with ECS calculated values.
http://www.readcube.com/articles/10.1029/2003GL018747?
https://www.ipcc.ch/pdf/assessment-report/ar5/wg1/WG1AR5_Chapter09_FINAL.pdf
The method for ECS calculation involves regressing the changing surface temperature against the changing net top of atmosphere (TOA) radiation deficiency over at least 140 years after the 4XCO2 burst and extrapolating the total change in surface temperature at which the TOA radiation is a net zero. The data for doing this calculation are given at the following links. I used the monthly values for the models for rlut, rsdt and rsut which are TOA outgoing long wave radiation, TOA incident short wave radiation and TOA outgoing short wave radiation, respectively. The net value used in the regression was obtained by subtracting the sum of rlut and rsut from rsdt . Also required to adjust the 4XCO2 surface temperature and TOA radiation series were these same values from the Pre-Industrial Control (PIC) for each CMIP5 model. Most of that data were taken from KNMI. All data were converted to annual for doing the regressions.
Abrupt 4XCO2 data from here:
http://cera-www.dkrz.de/WDCC/ui/EntryList.jsp?acronym=ETHc2
piControl data from here:
http://cera-www.dkrz.de/WDCC/ui/EntryList.jsp?acronym=ETHpc
Where there are differences between the ECS published values and those that I calculated (for OLS) it can be seen that the slope values match very well. That match indicates to me that these differences are the result of the differences in either the pi Control (Pre-Industrial Control spin up for the models) surface temperature or the pi Control Net TOA radiation values used to adjust the 4XCO2 surface temperature and Net TOA radiation (by subtraction). The authors of the paper describing the regression method talk about detrending the pi control temperature and radiation series which is something I did not do nor did I see why it should be done. I could, however, go back and do it simply to prove to myself that what I think caused the differences is correct.
It should be noted that the TLS regression results show on average a higher value for ECS than that produced by the AR5 method using OLS. While I wanted to get these results posted here as I promised Carrick, I have found a real interest in the differences from CMIP5 model to model in the pi Control temperature and Net TOA radiation series over a few hundreds of years of spin up. The mean value of these series and the trends vary considerably in my view. I have linked some of the pi control Net TOA radiation series in a post above already and I now am in the process of putting the plots of several CMIP5 models with the pi Control and 4XCO2 temperature and Net TOA radiation series and trends all together in one place. Intuitively I think more of the model pi Control series should be closer to a net zero TOA radiation and a flat trend and a temperature that is not trending significantly away from the mean.
Regression Results:
http://imagizer.imageshack.us/v2/1600x1200q90/540/jhixta.png
Thanks Kenneth—that’s a lot of work!
For the others, there’s a discussion of what piControl is, in this open-access article:
HaroldW (Comment #134482)
So far the only mitigation revealed to me is to subtract the pi Control values.
I get a lot of vague feelings when I read and analyze climate model data which to this point in my learning experience I owe to not seeing the entire picture. In this case I have a vague feeling that the Net TOA radiation from the pi Control runs which is significantly different than zero and can be trending is not so much a failure to reach equilibrium as it is some residual in the modeling process whereby the model tends toward the residual value and not zero. The trending might well be the model drifting – even though I have read that the CMIP5 models are much less drifting than the CMIP3 ones and that further (and I have to refresh my memory here) the CMIP5 drift is attributed to some effect different than that thought to cause the CMIP3 drift.
I have linked below 11 links with each link having 3 CMIP5 models plots that show the time series of the pi Control and 4XCO2 surface temperatures (tas) and TOA Net Radiation (TNR). For the pi control tas and TNR I have also shown linear trend lines in red and for the 4XCO2 tas and TNR I show red lines representing a spline smooth. To avoid confusion from the start, please remember that it is change in TNR and change in tas that is regressed to obtain an estimate of ECS while here I am making plots of these 2 variables (used for regression) against time in years. I have also linked a table that contains the results of the trends of the pi control tas and TNR and the means of these same series – all with confidence intervals (CIs) used to determine whether at the 95% limit the trends and means are different than zero. The CIs for the trends were adjusted for autocorrelation by assuming an ar1 correlation and using the rough and dirty factor of (1+ar1)/(1-ar1).
I show these results to note the differences from model to model and to be used at some later date to indicate which models better represent the real world. Generally it becomes obvious looking at the 4XCO2 TNR plots that the extrapolated time to reach zero varies considerably from model to model. There are also significant variations by model in the mean TNR and mean tas for the pi controls. Significant trends for pi control tas and TNR series were found for 15 and 5 (out of 33 models), respectively.
I would appreciate any comments here about which models appear to be unreal based on these results and why you make that judgment and further without assuming that the models left as more real are capable of representing the real world climate.
I found the paper: http://iopscience.iop.org/1748-9326/9/3/034016/pdf/1748-9326_9_3_034016.pdf dealing with same pi control properties as I have discussed in this post and previous ones. What caught my attention was the authors attempts to relate TNR to tas and TNR to ocean heat content (OHC). I would think that if a models TNR could be directly and highly correlated with the models OHC than I could dismiss my previous speculation about those models, where the TNR has evidently leveled off at some higher level, having a spurious TNR zero point. Unfortunately what I see in the paper is the authors using overlapping 10 year trends of Earth system energy content (which can be directly related to TNR) versus the overlapping 10 year trends of tas and versus the overlapping 10 year trends OHC. Why did not the authors make a direct correlation in place of using 10 year trends and further what affect would the very high autocorrelations caused by using overlapping trends have on the correlation results? Anyway the paper found that the correlation of the overlapping trends for TNR versus tas averages around 0.25 for the 30 of the CMIP5 models with a low of 0.08 and high of 0.63, while the same correlations of TNR versus OHC average around 0.95.
It is obvious that the place to look when determining how real the TNR values of the CMIP5 models for the pi control are over extended time periods is OHC since it has shown from observations, I believe, to account for 95% or more of the TNR variation. I am not at all sure that is what the authors of this paper have accomplished.
Plots of pi control and 4XCO2 tas and TNR series:
http://imagizer.imageshack.us/v2/1600x1200q90/537/16XtJY.png
http://imagizer.imageshack.us/v2/1600x1200q90/901/nveUVn.png
http://imagizer.imageshack.us/v2/1600x1200q90/661/O2MO3H.png
http://imagizer.imageshack.us/v2/1600x1200q90/661/T84uWf.png
http://imagizer.imageshack.us/v2/1600x1200q90/540/PKnGZv.png
http://imagizer.imageshack.us/v2/1600x1200q90/661/uFXX3P.png
http://imagizer.imageshack.us/v2/1600x1200q90/908/SlT37K.png
http://imagizer.imageshack.us/v2/1600x1200q90/903/pO221f.png
http://imagizer.imageshack.us/v2/1600x1200q90/537/zeD8Az.png
http://imagizer.imageshack.us/v2/1600x1200q90/661/JJMBac.png
http://imagizer.imageshack.us/v2/1600x1200q90/540/Bfusi8.png
Table of pi control series trends and means:
http://imagizer.imageshack.us/v2/1600x1200q90/540/REu6QB.png
I did some simulations in an effort to determine what kind of auto correlation would occur using overlapping 10 year trends as was done in the paper linked above in my previous post. That auto correlation was around ar1=0.964 which means after adjusting the CIs for auto correlation for the correlation reported in this paper would put the lower limit at zero correlation. The authors of that paper did not talk about what this overlapping does to auto correlation so I suspect that are not aware of this problem. I have also assumed that a correlation just like a trend needs to be adjusted for auto correlation – although I have not seen it done.
I have also completed my downloading of the piControl average sea potential temperature (thetaoga) in an effort to determine whether the CMIP5 models piControl series account for TOA net radiation by a rise or fall in the temperature of the global oceans.
Kenneth,
How anyone can look at those plots and take seriously model projections 100 years in the future is beyond me. One doesn’t know whether to laugh or cry.
DeWitt Payne (Comment #134869)
DeWitt, unfortunately the people who are in positions to act on the results of these models do not see these details and climate scientists and modelers, in my view, in general tend to shy away details as in the other example I am familiar with being the details of the proxies used for temperature reconstructions. That is crying shame in my book.
DeWitt:
One of the more curious things that has come out of this is the use of ordinary least squares to estimate ECS from the models. As we all know, this biases ECS low.
It appears the main effect of this is, by producing “low ball” numbers for model ECS, to make the models look like they are in better agreement with the data (which do not support high ECS values) than they really are.
is there a paper in this…on the other hand, where would it get published?
I have plotted from the CMIP5 models where data were available on a global mean and annual basis, the changes from year 1 for the piControl potential sea water temperature series versus the accumulated Top of the Atmosphere (TOA) Net radiation from the piControl TOA net radiation series. The plotted results are shown in the links below and include the linear regression line in red. I also linked a table that summarizes the results for each CMIP5 model analyzed along with an explanation of the calculations made.
The data for my analysis were taken from this link:
http://cera-www.dkrz.de/WDCC/ui/EntryList.jsp?acronym=ETHpc
My interest in this analysis was the result of reading the paper linked here:
http://iopscience.iop.org/1748-9326/9/3/034016/pdf/1748-9326_9_3_034016.pdf
In this paper the authors have used essentially the same data I used and were relating overlapping 10 year trends in the piControl sea water temperature with overlapping 10 year trends in the piControl TOA Net radiation accumulation after removing the overall trends from both series using a Butterworth filter. I question the wisdom of using overlapping trends since it leads to series with very high auto correlation (ar1) values that on adjusting the resulting correlations (or regression trends) for this dependence puts the correlation (trends) at zero. I would also question the authors using residuals from the filtering process if they were concerned with the overall relationship of ocean heat content (OHC) and TOA budget as I was. Perhaps for their purposes of a limited look at 10 year trends this might be more appropriate.
In my analysis I wanted to determine whether the CMIP5 models were consistent in the accumulated TOA Net radiation and the change in the sea water temperature (OHC), i.e. there should be a straight line relation close to a theoretical trend slope that I calculated as 2.63E-10 W/m2/degree K. I saw no need for detrending or working with residuals as I would think any drift in a model should show in both OHC and TOA Net.
Looking at the plots and table it is rather obvious that there are large difference between models with some models having a poor straight line relationship, some models showing a negative slope and some other models showing trend coefficients far from the theoretical value. Based on these results, I would say that of the 26 CMIP5 models analyzed in this manner only CESM1-CAM5 and GFLD-CM3 and perhaps IPSL-CM5A-MR met expectations of how the model would be expected to perform. The most troubling are those models that have a very straight and noiseless regression line but it is negative. Those models appear to be relating changes of accumulated energy to OHC in some deterministic manner but in reverse of expectations.
http://imagizer.imageshack.us/v2/1600x1200q90/631/2DVaS6.png
http://imagizer.imageshack.us/v2/1600x1200q90/673/TVGFGX.png
http://imagizer.imageshack.us/v2/1600x1200q90/661/ILN33H.png
I need to add here a link for the plot I left out in my previous post.
http://imagizer.imageshack.us/v2/1600x1200q90/913/kfz4Tj.png
Carrick (Comment #134872)
In the reply I received from Tim Andrews he said he had not looked at using total least squares regression for calculating ECS. I was surprised by his reply since the authors of the paper he coauthored talked about the errors in both variables and which to use as the dependent/independent variable.
Kenneth Fritsch –
The relationship you mention — 2.63E-10 W/m2/degree K — doesn’t make any sense. Units problem?
HaroldW (Comment #134897)
Look at the linked table for the conversion of both W/m2 (TOA Net Radiation) to joules for a year period of time and sea water temperature (K) change to joules of OHC. I regressed T versus W/m2 and that is why I noted (poorly) W/m2/K.
Oh, I see the problem now. I wrote down the incorrect value in my post – it should be 2.63E-03 which is the value in the table. Harold, thanks for finding my error.
Kenneth Fritsch –
I see now. I hadn’t actually opened the link to the table — I thought it was more examples of T-vs.-accumulated-TOA graphs.
So fixing the units, the relationship you expected was something like
dT ~ 2.63E-3 (K / (Wm-2 * yr))*integral (TOA), or
dT/dt ~ 2.63E-3 (K/yr / Wm-2)*TOA ?
If so, the current imbalance of ~0.6 Wm-2 (if I remember correctly) corresponds to 0.0016 K/yr.
Very odd that some of the graphs display a negative correlation. You’d think any sort of energy-conserving model would have a positive correlation. I have to think that the reported TOA imbalance (or the OHC, I suppose) is incorrect.
But that’s the problem. From what I’ve gathered, most models don’t conserve energy.
DeWitt,
“From what I’ve gathered, most models don’t conserve energy.”
.
I agree that it is surprising, but I don’t think that simple things like conserving energy and mass in the models matters nearly as much to the modelers as conserving frighteningly high climate sensitivity estimates in the face of contrary empirical data. The band must keep blaring the horns of alarm, or the public may cut off the stupid and wasteful funding excesses. (25+ models, and most of them comically wrong… Yikes!).
Carrick, here is what the original 2005 paper on estimating ECS by ordinary least squares had to say about the relevant parts of the authors’ analysis. Gregory was coauthor of both papers but no other coauthors of these 2 papers were common to both. Someone who coauthored the first paper obviously understood the issue of using total least squares (or a reviewer) but apparently waved it away.
“The regression slope for N against delta T gives a = 0.99 ± 0.07 W m_2 K_1. The uncertainty from the regression uses the RMS deviation in N (the dependent variable) from the fitted line to obtain an estimate of the uncertainty of the points. There are two possible problems with this.
First, there is interannual correlation of variability so the points are not independent. This leads to an underestimate of uncertainty but not to a bias. Second, the choice of dependent variable is arbitrary. N and delta T both have random noise, but regression assumes there is no uncertainty in delta T.
This tends to flatten the slope and underestimate its uncertainty. To gauge the size of the effect, we regress delta T against N, obtaining a slope of _0.94 ± 0.06 K W_1 m2, whose reciprocal is 1.06 ± 0.07 W m_2 K_1. The effect is not serious. However, the product of the two slopes equals the square of the correlation coefficient, so the difference is greater when the points have more scatter. Correction for this could be made by a more elaborate procedure based on statistical properties of variability in the control experiment. Ordinary regression is adequate when statistical uncertainty is low.”
http://onlinelibrary.wiley.com/doi/10.1029/2003GL018747/full
“Very odd that some of the graphs display a negative correlation. You’d think any sort of energy-conserving model would have a positive correlation. I have to think that the reported TOA imbalance (or the OHC, I suppose) is incorrect.”
Harold, from observation it is judged that most of the TOA imbalance goes into the sea water (95% or more) but from what I read not all models put that much of the imbalance into the sea waters. I doubt that would explain the differences seen here but it is something to keep in mind.
I do feel better about my calculations in finding that at least a few models had OHC close to the theoretical. I also checked some of the models with negative slopes and indeed these models showed increasing accumulated TOA and decreasing sea water temperatures.
Off the top of my head I would think if the net TOA was running at a constant excess with these models with negative slopes by subtracting that flat line down by a constant amount to something negative would now show a decreasing accumulation with a decreasing sea water temperature and a positive and linear slope would result – but again that is not what the data showed.
“If so, the current imbalance of ~0.6 Wm-2 (if I remember correctly) corresponds to 0.0016 K/yr.”
Harold, that would be on average for all the global sea water. I used that model output because the averaging was already done for me. Otherwise if I used the ocean depth levels and latitude and longitude (which is also available as model output) the amount of memory required goes from in the range of 100 k bytes per model to giga bytes per model. If you are thinking in terms of practical temperature measurements for the observed ocean in degrees per year it would be much higher since most of heat is absorbed in a small fraction of the total volume.
Kenneth Fritsch (#134915) –
Agree that .0016 K/yr is a global average. That was more of a thinking-out-loud comment about whether the factor was sensible. The surface will warm more quickly than the depths. Somewhat arbitrarily dividing the ocean into the top ~1000m & below, the top layer has ~20% of the volume. If the bottom is changing negligibly, then the top layer would be warming at .0016 K/yr / 20% = 0.008 K/yr or 0.8 K/century. Or 1.6 K/century at the surface, decreasing linearly to 0 at 1000m depth and below.
And (#134912) – Yes, not all of the TOA imbalance should go into the ocean. But most of it should, or so it is widely believed. A positive correlation is expected, with the slope of 2.6E-3 (K/yr / Wm-2) an upper bound. If the energy isn’t going into the ocean, then where is it going? — ice melt? atmosphere (including latent heat as water vapor increases)? land masses? None of those are thought to be large enough sinks. Perhaps it hasn’t been allowed enough time to reach an equilibrium? I’m at a loss.
Harold,
The evidence would seem to imply the co-called “missing” energy is leaving the system entirely.
@ SteveF,
You said,
“I agree that it is surprising, but I don’t think that simple things like conserving energy and mass in the models matters nearly as much to the modelers as conserving frighteningly high climate sensitivity estimates in the face of contrary empirical data.”
+1
This paper (draft) linked here talks about the most important feature of model tuning being the TOA energy budget and that that tuning is most often accomplished using cloud parameters.
http://www.mpimet.mpg.de/fileadmin/staff/klockedaniel/Mauritsen_tuning_6.pdf
I continue to attempt to make sense of what I found in the piControl discrepancy between OHC and TOA energy budget in most CMIP5 models that I analyzed. I was told, by the lead author (Timothy Andrews) of the paper describing the regression method used to estimate ECS values from an abrupt 4XCO2 experiment for AR5, that the CMIP5 piControl surface temperature and TOA radiation budget series must be used to adjust the CMIP5 model 4XCO2 surface temperature and TOA radiation budget series (by subtraction).
On first look this requirement might be thought to be related to the need for the model to spin up by way of the piControl, and that if that spin up has not produced a stable climate system going into the 4XCO2 experiment, the piControl series would be needed to adjust the 4XCO2 series. I have shown that not making the adjustment for TOA for the 4XCO2 series (using the piControl series) produces generally higher and more varied ECS values.
What I see on further observation of these piControl series is that for most models the series appear reasonably stable and particularly with regard to the large relative amount of changes in the 4XCO2 series. The adjustment in this viewed in this way appears to me to be one of subtracting out some residual amount of TOA net radiation and temperature that vary very little across the series time period. This development would put the need for adjustment in a very different light.
The model piControl and 4XCO2 series are different than the model series that are used to determine historical and scenario climates and perhaps are tuned differently. I suppose I now should look at the OHC from the global sea water temperature change versus the changes in the accumulated TOA energy budget for the RCP4.5 scenario CMIP5 model runs or perhaps historical model runs.
Kenneth –
The paper you cited is available in published form here.
I think that is also the paper that Carrick linked as a good rendition of how modelers tune their models and we discussed it. I just wanted to point to the observation that modelers have to tune their models to get the TOA energy budget correct. That is why I am now looking that budget in the CMIP5 RCP4.5 models versus OHC.
“Whether adding some random noise to the calculations, to simulate chaotic variability in sub-grid scale processes which are only “parameterized†would make a more realistic model is an interesting question.”
palmer has some stuff on this I can prolly find
Steve Mosher,
“palmer has some stuff on this I can prolly find”
.
I hope you can, because I think it is a very interesting question. My guess is that adding random noise would tend to make the model less “classically chaotic” in patterns and more random… but that is only a guess.
ok, I will go get it
http://www.newton.ac.uk/seminar/20100823100011001
around minute 25 he starts citing some literature
one of the examples he mentions
http://www.newton.ac.uk/files/seminar/20100823164517451-152520.pdf
Steve Mosher,
Thanks. Both the links were interesting. I was particularly struck by Plamer’s underlying justification for enlisting more mathematicians in the modeling effort…. ‘for the good of humanity’. Could I sit with Palmer and have a beer (mine cold, his ambient temperature), I would try to suggest that most everyone wants to advance ‘the good of humanity’, not just climate scientists. People everywhere tend to overestimate the importance of what they are doing, and that includes climate scientists. We would all be a lot better off if there were a great deal more humility distributed around the world.
.
One thing I very much liked was Palmer’s recognition that a multitude of climate models makes no economic or practical sense, since the only justification for a multitude is evaluating “variability”. I would argue that the ensemble does not even do that very well.
Kenneth Fritsch, thanks for the comments. Regarding this quoted material:
As long as you restrict yourself to “variability” this is probably true. But there are error sources that scale with trend (e.g., incomplete spatial coverage), and of course you have to include all of the error sources in your error budget, not just the ones that make everything look copacetic.
The known spatial coverage error results in a negative bias when the trend in the data is positive and a positive bias when the trend in the data is negative.
I think your study shows that it is serious.
What about software coding errors in the models? Do these kinds of errors form a possible source of ‘noise’ which could affect the simulation of chaotic variability in sub-grid scale processes?
Moreover, are software coding errors inherently deterministic in their basic nature? Or does it depend upon the specific type of coding error that was made?
Beta Blocker,
I cannot speak to the larger issues at all, but
No they are not. At least sometimes, a language specification lays out rules and if the rules are broken the results are undefined. This means there’s no predicting the result as far as the spec is concerned. It’s implementation dependent. I can recall at least one case from my personal experience where the behavior of buggy code changed depending on what platform it was being run / what compiler suite was used, and particularly how parameters and local variables in functions were allocated on the stack. Drove my semi-technical manager berserk, because she didn’t get that the program’s behavior was only guaranteed to be the same across platforms if none of the rules in the language spec were broken; I.E., if all behaviors were defined.
Mark Bofill,
I agree that coding errors can yield unpredictable results, certainly between platforms, and maybe even on the same platform, depending on how much the OS “constrains” an application (eg. does the OS let you read or write outside the bounds of a defined array?). My guess is that the (mostly) Fortran software used for models is largely unconstrained, and so can blow up easily and in unpredictable ways if there are any serious errors. These types of errors would unexpectedly terminate the (weeks to months!) long runs of the models, and cost a lot of money, so these types of errors probably are not in the code. If there are significant errors, they are more likely ones that impact the accuracy of calculations, rather than errors which influence the flow of the program, since the latter would probably crater spectacularly. The lack of a perfect energy balance could indicate an ‘accuracy bug’.
SteveF,
That sounds pretty reasonable to me. In glancing through the MOD_E faq though I got the impression that it wasn’t all that unusual for the code to ‘crater’, but one would think and hope that this wouldn’t happen often, for the reasons you gave. Dunno.
Does anyone know if any of the global circulation models are subject to formalized software quality assurance programs of some kind, ones which might include software requirements specifications, code reviews, software configuration management practices, system documentation requirements, and testing standards for production systems?
A fundamental difficulty in doing end-to-end software V&V for a general circulation model would be that it is not possible to determine ahead of time in precise objective terms what a “correct” output is. The correct output is whatever the simulation software produces assuming that the climate modeler’s intentions have been successfully translated into the model’s simulation code.
Mark Bofill,
“I got the impression that it wasn’t all that unusual for the code to ‘crater’, but one would think and hope that this wouldn’t happen often, for the reasons you gave.”
.
If their code is that buggy, then it ought to give them pause. From a practical standpoint (and since supercomputer time is not cheap!), maybe they regularly store the values of all variables so that they can recover and restart at a known “time” should things go badly.
.
Beta Blocker,
I very much doubt there is the kind of V&V that you are talking about done with the GCMs (they are huge programs that have evolved over decades, after all). Heck, I very much doubt there is that kind of V&V for most any program except a few “life-or-death” programs like might be used in combat or for flying an airplane by wire. Of course, the simpler the the program the easier formal validation becomes; certainly no product from Microsoft (at least no product I have ever used!) is rigorously verified and validated…. they often have horrific bugs. But I still use Microsoft products. To paraphrase George Box: “Most all significant programs have errors, but many are still useful.”
The more complex the program, the more difficult it is to find and fix bugs. In fact, when you fix a bug, there’s a significant probability you’ll create a new bug. Apple’s iOS, for example, seems to break Bluetooth communication to external devices like GPS and OBDII dongles with every other version.
DeWitt Payne (Comment #134940),
And some iOS versions regularly lose the ability even to read their OWN internal GPS sensors. I have twice had to install a new iOS version to ‘reactivate’ an internal GPS unit that suddenly refused to work with an earlier version. Yup, very complicated programs are extremely difficult to debug; even if they were ever at one point bug free, every subsequent change, improvement, or modification runs a serious risk of introducing new bugs.
“A fundamental difficulty in doing end-to-end software V&V for a general circulation model would be that it is not possible to determine ahead of time in precise objective terms what a “correct†output is.”
There is. GCM’s solve partial differential equations. And the right check to perform is whether the output actually does satisfy the equations. This is more often done by checking the conserved quantities like energy and momentum. But you can, if you want, substitute back in the equation directly. Of course, there are also uniqueness issues, but they are fairly generic.
The task of verifying is much simpler than solving, so the process should pick up bugs in the solve process.
I think Nick, you are dramatically oversimplifying. In any numerical simulation of a partial differential equation, the equation is never exactly satisfied. The truncation error is the amount by which the numerical solution differs from the exact solution. Even for well defined problems such as air flow over a wing, this error is typically on the order of 1% – 10%. This is a lower bound for the global error measured for example in the L2 norm.
A second problem of course is that all scales are important in a turbulent flow. The unresolved scales must be “modeled” typically by using a nonlinear artificial viscosity. The error in these models can be very large and is strongly problem dependent. They are typically tuned using a small set of flows for which there is data. In GCM’s, things like clouds don’t have adequate data for this process to be very meaningful.
Your formulation is technically of virtually no value in determining what the actual error is. Consistency is the technical term used if the numerical scheme has a truncation error that goes to zero as the grid size goes to zero. Stability means the numerical scheme is bounded. For elliptic systems, stability + consistency implies convergence to the exact solution. However, for nonlinear non-elliptic systems, this theory is of little value. You design things so they are consistent and stable and hope for the best. But even for elliptic problems, the errors are typically much larger than laymen suppose. There is a vast and rigorous mathematical literature for the elliptic case. You can look at Leszek Demkowicz of Ivo Babuska who were pioneers in developing this theory.
For those interested, Babuska has devoted his golden years (he is in his 80’s) to verification and validation. His papers are an excellent start.
David,
My formulation will tell you whether the solution process does what it was supposed to do. That is responsive to what was requested – an a priori definition of correctness, which will test whether the software is defective. For this purpose, you should test whether the solution satisfies the equations as discretised (in the same way). And yes, that would be a problem with say adaptive meshes, but I don’t think GCMs have that.
Well, Nick, its more complicated than that. The discrete equations are so complex with very complex unwinding stabilization required, that the only way to verify that the output satisfies these equations is to write an equally complex code to verify this which of course is equally likely to have bugs. What you can do is use very simple situations such as a flat plate boundary layer where analytical solutions are known to weed out gross bugs, but these tests are wholly inadequate really. We have a CFD code that has O(100) man years of effort in it and we still find bugs all the time. Usually they have a minor effect but not in all cases. Our sub grid model has been fixed for 20 years so its much easier for us than for atmospheric modeling. Every time a sub grid model changes, you need to go through a new verification process.
“unwinding”
Upwinding?
But do GCM’s use it? It isn’t much mentioned. Yes, there is a whole lot of stuff like diffusion correction. But most of this is meant to stabilise the solution process, not alter the solution.
I think verification code would be simpler. But anyway, if you get agreement, then there is a good chance that neither code has bugs. Or, less ambitiously, if you get reasonable agreement, then the bugs are small (in effect).
Kenneth Fritsch –
Gupta et al., Climate Drift in the CMIP5 models (2013) may be of interest. From the abstract, “metrics that include the deep ocean … where drift can dominate in forced simulations.”
David Young,
“A second problem of course is that all scales are important in a turbulent flow.”
.
Which is something that has always puzzled me. The mathematical description is of an continuum. But if we know that ‘all scales’ are important, then it seems clear that at very small scales, the assumption of a continuum is incorrect… there is random local fluctuation in pressure and motion of molecules at the few microns and smaller scale (Brownian motion and all that). Does this not mean that there can never be an exact solution, and that there is an irreducible level of ‘randomness’ in turbulence which can never be completely modeled?
Yes Nick, I meant upwinding. The alternate is artificial dissipation which is an old and inferior method. You need something to make the scheme stable. Perhaps the hyper viscosity Bowning has pointed out is so large nothing else is needed. That would of course be a very bad way to stabilize your scheme.
Technically, “all scales” means all scales down to the viscous limit: This is known in the literature as the “Kolmogorov spatial (or length) microscale”, which is on the order of 1-mm for the atmospheric boundary layer.
SteveF, This is not my area, but there are the Boltzman equations which I think are supposed to describe the molecular level stochastic motions. This is an area where a lot is going on. I need to research it when I have time. Generally, the Navier-Stokes equations are thought to model things correctly enough if we average the molecular level interactions.
SteveF,
Well, it isn’t my area either, but I’m retired now, so I can say anything 🙂
Navier in fact got his version of the N-S equations by your query in reverse. He was an enthusiast for Dalton’s new theory, and he worked out elasticity and fluid equations using its principles. I think his logic (re fluid) is now deprecated, but it led to the right answer.
The Euler equations give the conservation of momentum, together with linear compressibility (or none). That doesn’t really raise continuum issues. N-S added a constant (or continuum) diffusivity of momentum. That is where turbulence is usually modelled. Eddies increase the diffusivity, not only of momentum, but of heat and just about anything else being advected. The diffusivity originates from molecular motions (molecular diffusivity); turbulence enhances it.
That is why Navier got it right. Once you think on a molecular scale, you can’t help getting diffusivity.
SteveF,
ps L F Richardson anticipated your query:
Big whorls have little whorls,
that feed off their velocity.
Little whorls have lesser whorls,
And so on to viscosity.
Viscosity being the diffusivity of momentum.
Carrick,
The viscous limit doesn’t mean absolutely zero energy in motion at scales below the limit. It means the energy decays very rapidly at those scales. So, more “for all practical purposes” it’s gone.
How that interacts the possibility that slight perturbations at smaller scales could have a non zero effect–and given chaos possibly affect a trajectory? Dunno.
Lucia, it’s exponentially damped in wavenumber isn’t it?
Even in the inertial sub-range (Kolmogorov spectrum) you expect a variation in $latex 1/k^{5/3}$ for temperature and velocity components ($latex 1/k^{7/3}$ for pressure), so that’s already a spatial filter of sorts.
This is more your and David Young’s expertise, but it’s my understand that sub-grid resolution is handled by assuming a Kolmogorov roll-off, correct?
Carrick
When the method of dealing with sub-grid does anything directly, I think that’s usually the case.
My only point is that whenever you are using a sub-grid approximation to deal with what it does, these “chaos type” question people ask can’t really be answered.
The “chaos type” is: In real life could a very small perturbation– say a molecular “blip” affect which trajectory you take? And if yes, could that “grow”.
In subgrid modeling, all you care about is total “averaged” stuff in some sense.
I’m not sure how to clearly explain the distinction: but you could have various trajectories all of which still have energy roll off exponentially in the subgrid, but the exact trajectory of how something moves is slightly different. If the trajectory deviations grow, things might change. So… the “butterfly flap” in 2010 might shift the “hurricane” landing position in 2014. But meanwhile, the “energy” at all the scales is still small both in the cases where the ‘butterfly flapped” and where it did not flap.
I think the modeling or even energy dynamics don’t really tell us the full answer. (I could be wrong.)
Below, I have copied a couple of emails that I sent to Geert Oldenborgh of KNMI concerning a difference I have between KNMI and DKRZ databases for the radiation variables rsdt, rtsut and rlut for the CMIP5 RCP4.5 models. I want to resolve these differences before posting any more results here on the relationship of global sea water temperature changes versus changes in the accumulation of Net TOA radiation. I will await a reply from Oldenborgh who has been prompts in past relies and then perhaps correspond with DKRZ. I would also appreciate any suggestions from posters here.
e-mails
I have been using data from your website for the CMIP5 radiation variables of rsut, rlut and rdst in order to obtain the TOA net radiation and in combination with data from the DKRZ database site at the link http://cera-www.dkrz.de/WDCC/ui/EntryList.jsp?acronym=ETHr4 . I do this because there are some CMIP5 data at DKRZ that is not available at Climate Explorer. I use KNMI because it requires considerably less computer power on my part to download data that is in global mean form and does not need the downloading of nc gridded files, that are required for the DKRZ data, and then obtaining global means.
I have some puzzling differences between what I download from KNMI for the radiation variables noted above and what I obtain (on a sampling basis for some 12 model runs) from DKRZ in gridded form and converted to global means using my code in R. I have checked my R code by calculating the GISS 1200 km global mean anomaly from a gridded nc file from KNMI against what I obtain from that series downloaded directly from KNMI. Any differences (for the years 1957-2014) between the two are in the rounding error range.
To be more clear on the differences I see between KNMI and DKRZ results, I will tell you that I see differences in the Net TOA (by Net=rsdt-(rsut+rlut)), for all the model runs I have compared, where the difference varies by a near constant amount for every year in the series. If can be very close to zero for each year in the series for some models up to 0.75 watts/m2 for every year for some others.
I have racked my brain looking for a logical way these differences can occur by way of a calculation error on my part and have not found any. Do you have any information that might bear on my puzzlement here? Could there have been a constant amount of correction to these variables in one or the other data bases? Any suggestions you might have will be greatly appreciated.
Kenneth Fritsch
On further analysis of the DKRZ and KNMI data bases, I have found the same yearly constant differences in the rsdt and rlut variables that are different for different models and different for rsdt and rlut but in all cases for an individual model the same for every year in the series. For all models compared the constant yearly difference for rsdt is always larger than the constant yearly differences for rlut. The difference between DKRZ and KNMI found for rsut for a given model is within rounding error.
Since I obtain very close results for rsut for both KNMI and DKRZ, I am assuming that my handling of the data is not the problem and the data differences for rsdt and rlut are real.
I could further analyze this problem by comparing directly a download of KNMI and DKRZ gridded rsdt, rlut and rsut data in nc form for a given model and then convert both to global means using my R code. In fact I tried this but the download from KNMI appeared to either stall or was taking a very long time.
Kenneth Fritsch
Hey guys,
I’ve done some personal research on the climate of East Tennessee and would like some feedback. I downloaded the GHCN Daily data for 7 sites (2 official and 5 COOP) then wrote the code to convert and analyze the data. I used a 60 year period (1951 -2010) for the baseline mean and standard deviations. Then I ran simple linear regressions on the anomalies over 30 year periods to look for trends. On average r^2 was rarely more than .25 but that would be expected given the large variability on small scale local climate.
Missing data was a big problem (especially COOP’s) and I tried to account for it by using a counter. Then I calculated quality percentages for trend and mean periods to help determine reliability. Some of the data goes back to 1889 but information about a site (when and where it was moved, etc) only goes back to 1948 as far as I know. Knoxville and Tri-Cities (TYS, TRI) are the official sites and therefore the most reliable.
Overall from 1981 to 2010, East Tennessee has warmed but not to the degree it cooled from 1951 to 1980. Thunderstorm frequency is up in the last 30 years which makes sense in a warmer environment. Rainfall is generally up but snow is uniformly down. Part of this is due to the switch to ASOS in the 90’s when -9999 started showing up for snowfall and snow depth but the COOP sites all show a downward trend with better data quality than the official sites.
The most intriguing part was that the lower elevation sites in general had relatively weak max temperatures increases from 81 – 2010 but very strong min increases (~ 5 C /Century). These sites were at or near the bottom of local river basins. Tri-Cities, the second highest observation site sitting on a high valley plateau, was just the opposite. Relatively strong max temperature increase coupled with a slight decrease in minimum temperatures. Tri-Cities is a great place for radiational cooling.
I assume this is related to the height and depth of the nighttime inversion but would be interested on any comments from others.
It was much easier to post links for the full result tables as images. Also, I’ve posted a link to an image of the geographic region with each site labeled.
Thanks.
Tom R
1981 – 2010
http://jan.imghost.us/GzKQ.png
1951 – 1980
http://jan.imghost.us/GzIX.png
1921 – 1950
http://jan.imghost.us/GzI3.png
1891 – 1920
http://jan.imghost.us/GzHU.png
Geographical
http://jan.imghost.us/GzGe.jpg
Inversions
been playing around with it
http://onlinelibrary.wiley.com/doi/10.1029/2008JD009879/pdf
Also
Another approach to understand inversions/cold air pooling is
by computing the TWI
explained here
http://www.readcube.com/articles/10.1002/2013JD020803?
picture here
http://www.ecoweb.info/sites/default/files/styles/slidshow_big/public/images/img_638-1677_TWI_DarEsSalaam.gif
Thanks for the reply.
I’m having a hard time wrapping my brain around the mechanism for the differences in warming/cooling rates in relation to elevation.
Just a few hundred feet.
Tom R
UAH anomaly for January, 2015 is 0.35, if anyone still cares.