By Steven Goddard
I recently had the opportunity to attend a meeting of some top weather modelers. Weather models differ from climate models in that they have to work and are verified every hour of every day around the planet. If a weather model is broken, it becomes obvious immediately. By contrast, climate modelers have the advantage that they will be long since retired when their predictions don’t come to pass.
Weather and climate models are at the core very similar, but climate models also consider additional parameters that vary over time, like atmospheric composition. Climate models iterate over very long time periods, and thus compound error. Weather modelers understand that 72 hours is about the limit which they can claim accuracy. Climate modelers on the other hand are happy to run simulations for decades (because they know that they will be retired and no one will remember what they said) and because it provides an excuse to sink money into really cool HPC (High Performance Computing) clusters.
But enough gossip. I learned a few very interesting things at this meeting.
1. Weather modelers consider the realm of climate calculation to be “months to seasons.” Not the 30 year minimum we hear quoted all the time by AGW groupies. That is why NOAA’s “Climate Prediction Center” generates their seasonal forecasts, rather than the National Weather Service.
2. The two most important boundary conditions (inputs) to seasonal forecasts are sea surface temperatures and soil moisture. No one has shown any skill at modeling either of those, so no surprise that The Met Office Seasonal forecasts were consistently wrong.
For example, just a few months ago the odds of La Niña were considered very low. Compare the December forecast with the May version. How quickly things change!
SST modeling capabilities are very limited, and as a result seasonal weather forecasts (climate) are little more than academic exercises.
Oh and by the way, Colorado will be exactly 8.72 degrees warmer in 100 years. But they can’t tell you what the temperature will be next week.
“If I don’t understand it, it must be simple.”
– Dilbert Principle
In the top picture, which boxer is weather and which one is climate? What do readers think?
kadaka:
“Now if there is this much difficulty in applying deterministic Newtonian mechanics to accurately predicting something as relatively simple as coin flipping, how can you think that vastly complex global climate models can accurately predict conditions for decades and even centuries past the starting point?”
The point you keep missing is that you can predict that about 50% of coin flips will land on heads. If you run a deterministic prediction for long enough then even if the individual flip outcome sequence is incorrect, the statistics of that system can still be correct.
Nobody is trying to predict the specific weather conditions at a city for a given day in 100 years time -they are just trying to predict the statistics of those weather conditions when averaged globally and over a decade or so. That is a very different problem.
bemused says:
July 5, 2010 at 5:16 pm
…..Nobody is trying to predict the specific weather conditions at a city for a given day in 100 years time -they are just trying to predict the statistics of those weather conditions when averaged globally and over a decade or so. That is a very different problem.
______________________________________________________
Yes and the models get a straight line. Straight lines are not seen very often in nature, curves and cycles are much more common especially with weather and climate.
North to south throubh the middle of my state
North – Raleigh NC
Middle – Fayetteville NC
South – Lumberton NC
Atlantic Multidecadal Oscillation Amazing how the temperatures follow the Atlantic ocean oscillation – a 60 year cycle or there about.
kadaka said: “Heck, with the repeated iterations even the rounding off of the numbers can grow into a significant source of inaccuracy.”
No, this concern is unfounded. The reason is that the weather attractor oscillates within strong bounds, the weather trajectory eventually returns into a very close proximity state space after several iterations. From many known attractors in hydrodynamics, the structure of phase flow on an attractor is a horrible web of homoclinical orbits, but each of this orbit belongs to the same attractor (if it is “structurally stable” one). Therefore, even if the numeric algorithm produces rounding errors and the trajectory jumps to a different orbit of the attractor, it is still on the same attractor. Therefore the system does not go anywhere (it does not “blow up”) and produces (allegedly) the same statistics.
This is in theory of course. In reality of, say, climateprediction.net, evidence is that their model could suddenly produce “outliers”, which they simply discard. They discard about 40% of runs for this reason. To me it means that their dynamical system of pseudo-weather is unphysical and fundamentally wrong, and the results of forward calculations have little to do with reality.
Al Tekhasski:
Bemused, you are confusing statistics of a relatively short-term events with individual (unique) trajectory of a long evolving object.
No, I’m not confused at all. The thing that is trying to be predicted is not the 100,000 year+ range of possible states that the climate system may ever evolve into. People are trying to predict 100 years into the future i.e. in the context of the 100,000 year turnover period you quoted it is still “a relative short term event” (any reference where that figure came from by the way?).
“you have unspecified strange attractor of weather that, according to known facts from nonlinear dynamics, would certainly have long tails of slow-evolving fluctuations.”
-I am genuinely interested in this so if you have any good pointers where I can read more on this specific topic (in relation to the atmosphere) then I’d really appreciate it.
“Even to get a single turn around climate attractor one need to get full data for entire atmosphere (and all boundary conditions as well) for at least 100,000 years. Only then you can have some basis for climate prediction.”
…and yet, if you give the system an external hammer blow by say, turning off the sun, you can still predict that the climate will get colder. It is not black and white, it is possible to make predictions about certain aspects of the climate with differing amounts of uncertainty.
With weather forecasting, you could say “my model is not perfect, I can never know all of the initial conditions” so there is no basis for weather prediction. But, weather forecasts of synoptic scale systems are pretty accurate up to 3 days ahead. Lives are saved every year as ships know in advance how to avoid gales. A certain amount of pragmatism is not a bad thing.
Gail Combs:
“Yes and the models get a straight line. Straight lines are not seen very often in nature, curves and cycles are much more common especially with weather and climate.”
No, the models don’t produce straight lines -in fact they are very wiggly. People often get confused about this as often the time series shown in publications are smoothed with running means or are the average of an ensemble of different model simulations.
bemused:
“The thing that is trying to be predicted is not the 100,000 year+ range of possible states that the climate system may ever evolve into. People are trying to predict 100 years into the future i.e. in the context of the 100,000 year turnover period you quoted it is still “a relative short term event” (any reference where that figure came from by the way?).”
The 100,000 years came form the period of ice ages. I am under an opinion that if the attractor in a climate model does not reproduce this cycle, the entire topology of attractor cannot be trusted, and the 100-years long trajectory should not be trusted either. After all, you would still need a multitude of SAME experimental conditions to validate this model to build an ensemble. If we have only one sequential trajectory but the return map of the overall attractor needs 100,000 years, one cannot conclude that 100-year sequential segments originate from nearly the same state, so no statistics can be derived.
[ and I probably should use “conflating” instead of “confusing”. Damn English!]
“I am genuinely interested in this so if you have any good pointers where I can read more on this specific topic (in relation to the atmosphere) then I’d really appreciate it.” [regarding slow tails]
Unfortunately, I cannot provide any references with regard to applied atmospheric studies; this observation came from highly academic studies of simplified hydrodynamic of Couette-Taylor flow between rotating cylinders of about 30 years ago. The general idea is that if a spatially distributed system has nearly individual localized attractors that are weakly coupled due to sheer physical distance between them, this weak interaction leads to very slow type of “wobblings”, “walkings”, etc. , and weaker interaction simply yields slower motions. I believe today this kind of models are called “lattices”.
“But, weather forecasts of synoptic scale systems are pretty accurate up to 3 days ahead. Lives are saved every year as ships know in advance how to avoid gales. A certain amount of pragmatism is not a bad thing.”
I was under impression that they can do 5 or even 7 days… I agree that climate models do not need precise weather “forecast”, but I still think that climate models must reproduce typical weather elements and their individual shapes and dynamics pretty accurately, otherwise statistics of goofy events will result in goofy outcome.
“The 100,000 years came form the period of ice ages. I am under an opinion that if the attractor in a climate model does not reproduce this cycle, the entire topology of attractor cannot be trusted”
Well, my understanding of the latest thinking (and I stand to be corrected) is that the ice ages occur primarily from external forcing of the system by Milankovitch cycles. i.e. external forcing of the system changes the shape of the attractor. I don’t think they’re caused by internal variability of the system (within a stationary attractor).
I guess it all depends on how you define the ‘system’ and what you consider to be external to it. If you don’t give a model the external forcing then it won’t reproduce variability on those timescales.
“I was under impression that they can do 5 or even 7 days”
Well, yes. These vague statements are all rather meaningless really unless we define exactly what it is we are trying to predict. Small scale features like thunderstorms may be predictable a few hours in advance, synoptic scale features may be predictable up to 3-5 days (or more in favorable conditions), planetary scale waves are often predictable out to 10 days or longer. In the tropics, seasonal forecasts out to 3 months for certain variables have been demonstrated to have skill.
“but I still think that climate models must reproduce typical weather elements and their individual shapes and dynamics pretty accurately, otherwise statistics of goofy events will result in goofy outcome.”
Absolutely. But you may be surprised by how well the latest GCMs do reproduce atmospheric modes of variability. High resolution models produce very realistic convective storms and tropical cyclones. Synoptic systems and planetary waves have more or less been nailed for some time. Things like El-Nino Southern Oscillation are pretty well represented these days. The Quasi-Bienniel Oscillation in the stratosphere is beginning to appear in the latest models. Many models struggle to produce a fully fledged Madden-Julian Oscillation in the tropics, but there are promising signs from the latest global weather forecast models. The models are becoming less ‘goofy’ as time goes on. It is worth noting that the models spontaneously produce these modes of variability on their own -they are not programmed to do so.
@Gail Combs says:
July 5, 2010 at 5:51 pm
http://digitaldiatribes.files.wordpress.com/2009/09/amoraw200908.png
That is no use for a barbecue summer forecast then, 1975/76 have dissapeared. I do see a sub-cycle of around 7yrs though.
Bemused, I appreciate your posts. It is well known that ENSO parameters correlate with land temperatures. The cause and effect mechanism is fairly well understood as the hydrological cycle. Thus the theory of land temperatures being significantly affected by oceanic temperature is established.
The major underlying assumption of AGW models has to do with added heat from increased greenhouse gas concentration effects. Between land and water, only water has the potential to absorb this heat (land is such a poor candidate for this part of the theory that it can be dismissed). The modeled hypothesis is that the stored heat is then released along with natural heat under warm ENSO parameters, thus adding to the increased land and SST temperatures of warm ENSO parameters. It is also modeled that cool ENSO parameters will not be as cool due to these same greenhouse gasses.
An alternate modeled hypothesis has to do with AGW air temperature changes creating human-induced changes in the natural bimodal atmospheric circulation patterns. This model is less well accepted because air and land do not store heat.
Also, of great importance, the set of models predicting increased warming assume that CO2-AGW is added heat, not overwhelmed heat.
But what if the models have underestimated natural ENSO parameters? And what if the models have not accurately modeled all ENSO parameters, including clouds? Is it possible that these weaknesses in the models produce enough measurement error that any CO2 warming would fall within that error band? Is it possible that what we are seeing as a warming trend, is within the error band of natural variability and weak modeling of this variability?
Pamela Gray:
“But what if the models have underestimated natural ENSO parameters? And what if the models have not accurately modeled all ENSO parameters, including clouds? Is it possible that these weaknesses in the models produce enough measurement error that any CO2 warming would fall within that error band? Is it possible that what we are seeing as a warming trend, is within the error band of natural variability and weak modeling of this variability?”
I’m not an ENSO expert, so don’t want to comment too deeply on the specifics of it, but in short no, I don’t think the warming trend is within the bounds of natural ENSO variability, or caused by poor modeling of ENSO.
Look at the GISS observation data here:
http://data.giss.nasa.gov/gistemp/graphs/Fig.E.lrg.gif
-you can see how ENSO effects global temperatures (you can match up individual peaks and troughs). However, the warming trend is larger than the ENSO variability.
Another good read is here:
ftp://ftp.gfdl.noaa.gov/pub/gav/PAPERS/VW_09_ENSOCCreview.pdf
-look at figure 3. It shows the ENSO cycle in a model where no CO2 increase is applied (i.e. the control run). It shows the model produces a stable ENSO over 2000 years (showing variability on annual, decadal and centennial timescales, but no drift).
This article was extremely interesting, especially since I was searching for thoughts on this subject last week.
I have been coming to this blog for a couple of days now and i’m very impressed with the content!
thanks & regards