Why global climate models do not give a realistic description of the local climate : Melbourne Indymedia

2:20 am

Yet, whereas the global climate models (GCMs) tend to describe the global climate statistics reasonably well, they do not provide a representative description of the local climate. Regional climate models (RCMs) do a better job at representing climate on a smaller scale, but their spatial resolution is still fairly coarse compared to how the local climate may vary spatially in regions with complex terrain. This fact is not a general flaw of climate models, but just the climate models' limitation. I will try to explain why this is below.

Regional climate characteristics
Most GCMs are able to provide a reasonable representation of regional climatic features such as ENSO, the NAO, the Hadley cell, the Trade winds and jets in the atmosphere. They also provide a realistic description of so-called teleconnection patterns, such as wave propagation in the atmosphere and the ocean. These phenomena, however, tend to have fairly large spatial scales, but when you get down to the very local scale, the GCMs are no longer appropriate.

Minimum scale
Land-sea mask for ECHAM4 There are several reasons why GCMs do not provide a representative description of the local climate (i.e. exactly where I live). For one, the grid mesh, on which they compute the physical quantities relevant for the climate, is too coarse (typically 200km) to capture the local aspects. The figure on the left shows a typical land-sea mask for a GCM.

The distance between two grid points in a GCM (or an RCM) is the minimum scale (~200km). The coarse resolution typically used in the GCMs till now has implied that the topography has been smooth compared to the real landscape and that some countries (e.g. Denmark and Italy) are not represented in the models (one exception is one Japanese GCM with an extremely high spatial resolution).

Sub-grid processes are represented by parameterisation schemes describing their aggregated effect over a larger scale. These schemes are often referred to as 'model physics' but are really based on physics-inspired statistical models describing the mean quantity in the grid box, given relevant input parameters. The parameterisation schemes are usually based on empirical data (e.g. field measurements making in-situ observations), and a typical example of a parameterisation scheme is the representation of clouds.

Surface processes
Fjords Climate models need boundary conditions describing the surface conditions (e.g. energy and moisture fluxes) in order to yield a realistic representation of the climate system. Often simple parameterisation schemes are employed to provide a reasonable description, but these do not capture the detailed variations associated with small spatial scales.

Skillful scale
Shortcomings associated with parameterisation schemes and coarse resolution explain why one gridpoint value provided by the GCMs may not be representative for the local climate. A concept called skillful scale has sometimes been employed in the literature, most of which have been linked to a study by Grotch and MacCracken (1991) who found model results to diverge as the spatial scale was reduced. Specifically, they observed that:

Although agreement of the average is a necessary condition for model validation, even when [global] averages agree perfectly, in practice, very large regional or pointwise differences can, and do, exist.

Although it is not entirely clear whether this study really touched upon skillful scale, it has since been cited by others, and used to argue that the skillful scale is about 8 gridpoints. Nevertheless, since the 1991-study, the GCMs have improved significantly, and the GCMs now are run for longer periods and with diurnal variations in the insolation.

Regionalisation
glassesblurredsharp

The figure above gives an illustration of the concept of regionalisation, or so-called downscaling. The left panel shows a typical RCM land-sea mask, giving a picture of its spatial resolution. The middle panel shows a blurred satellite image of Europe, which can illustrate how the sharp details are lost yet providing a realistic large-scale picture. The unblurred image of Europe is shown in the right panel. An analogy for the data from GCMs is looking at a blurred picture (middle above) while regional modeling (RCMs) and empirical-statistical downscaling (ESD) is putting on the glasses to improve the image sharpness (right above).

Both RCMs and GCMs give a somewhat 'blurred' picture albeit to different degrees of sharpness, and RCMs and GCMs are similar in many respects. However, GCMs are not just 'blurred' but also involve some more serious 'structural differences' such as an exaggerated Gibraltar Strait (see land-sea mask above), and the Great Lakes area, or Florida, Baja California are quite different and not just blurred (see figure below). Such structural differences are also present in RCMs (eg. fjords), but on much smaller spatial scales.

Model resolution, (Source: Strand, NCAR)(Source: Strand, NCAR)

Yet the images shown here for present climate models do not really show features down to kilometer scales that may influence the local climate where I live, such as valleys, lakes, mountains and fjords, even for RCMs (the lower right panel shows an optimistic projection for improved spatial resolution in GCMs for the near future). The climate in the fjords of Norway (can be be illustrated by the snowcover) is very different from the climate on the mountains separating them. In principle, ESD can be applied to any spatial scale, whereas the RCMs are limited by computer resources and the availability of boundary data.

What is the skilful scale now?
My question is whether the concept of a skillful scale based on old GCMs still apply for the state-of-the-art models. The IPCC AR4 doesn't say much about skilful scale, but merely states that

Atmosphere-Ocean General Circulation Models cannot provide information at scales finer than their computational grid (typically of the order of 200 km) and processes at the unresolved scales are important. Providing information at finer scales can be achieved through using high resolution in dynamical models or empirical statistical downscaling.

The third assessment report (TAR) merely states that 'The difficulty of simulating regional climate change is therefore evident'. The IPCC assessment report 4 (chapter 11) and the regionalisation therein will be discussed in a forthcoming post.
5 blog reactions

21 Comments »

1.

Any comments on this?

James C. McWilliams
Irreducible imprecision in atmospheric and oceanic simulations
PNAS | May 22, 2007 | vol. 104 | no. 21 | 8709-8713
http://www.pnas.org/cgi/content/full/104/21/8709

It emphasises that the problem is not just one of decreasing grid size.

What about the specific proposition that:

"No fundamentally reliable reduction of the size of the AOS dynamical system (i.e., a statistical mechanics analogous to the transition between molecular kinetics and fluid dynamics) is yet envisioned."

Is this a problem of principle, like sensitive dependence on initial conditions, or something that enough years of hard work might crack?

Comment by Joe — 27 May 2007 @ 6:14 am
2.

Thanks Rasmus,

I did not fully grasp the concept of skilful scale. Are you saying that GCMs do not reliably represent climate to a resolution of 1 grid point and that to achieve accurate representation you need to average over about 8 grid points? Hence the "skilful scale" is 8 grid points.

[Response:This is basically the point, yes. But there has not been much discussion about what the skilful scale has been lately, so I'm not sure if it is still true. -rasmus]

Do climate scientists expect any surprises as resolution increases? Were there "surprises" between the 1980s GISS models and the latest models? I note that in our region (Australia) your minimum scale map leaves out Bass Strait (between Australia and Tasmania) and Torres Strait (between Australia and PNG). These are significant water ways for local climate and ocean currents. They are also about 150 km wide - close to 200 km - so why would they be omitted?

[Response:One Japanese model does have a very high spatial resolution, but I don't think there are any particular surprises. Perhaps an improved resolution may provide a better rpresentation of the MJO and the monsoon system as well as cyclones. The very high resolution model makes very realistic pictures of the cloud and storm systems, and the guys presenting the results are fond of showing animations which look very much like satellite pictures. Quite impressive. -rasmus]

Do GCMs capture coarse topographic features, eg the Tibetan Plateau?

[Response:apparently not well enough. -rasmus]

Comment by Bruce Tabor — 27 May 2007 @ 6:17 am
3.

I agree that for quantitative studies GCMs cannot be used at a regional or single gridpoint scale. However, the results can be considered valid at a qualititave level.

[Response:I think this is true too, but downscaling should in general add value to the simulations. -rasmus]

I have used a VERY coarse resolution GCM (8 degrees by 10 degrees) to help plan a vacation, and it worked quite well. The data provided by the GCM was better than any guide book or even the CIA World Factbook. In addition, the GCM provided more than just temp + precip, including other useful variables like cloud cover (tanning), soil moisture and ground wetness (camping) and wind speed and direction (windsurfing).

See my EdGCM writeup on my vacation planning for more details: http://edgcm.columbia.edu/outreach/showcase/cambodia.html

Comment by mankoff — 27 May 2007 @ 7:06 am
4.

It would be helpful to provide a useful definition of what is meant by "local".

Comment by Charles Raguse — 27 May 2007 @ 8:37 am
5.

No, no, no. Why doesn't the AUTHOR OF THE ARTICLE give a useful definition of LOCAL climate!

Comment by Charles Raguse — 27 May 2007 @ 8:40 am
6.

What is meant by LOCAL climate?

[Response:I'm getting the message, don't worry. To me, the local climate is the climatic characteristics which have are directly relevant to my perceptions. This would normally be on a smaller scale than a grid-box for a GCM and smaller than 'meso scales' referred to in meteorology (more like 'meso-gamma') and smaller than minimum scale of most RCMs (which typically have spatial scales of ~50km, although some go down to ~10km). I define regional scales as somewhat larger, that whch characterises a larger region (e.g. at meso scale to synoptic scales). -rasmus]

Comment by Charles Raguse — 27 May 2007 @ 8:41 am
7.

How do you manage to convert heat, as in watts m-2, into temperature?
What is the relationship between the steady state input of energy, in watt m2, and total global atmospheric volume, pressure and temperature?

[Response:First law of thermodynamics, but this is done in the models. -rasmus]

Comment by DocMartyn — 27 May 2007 @ 9:17 am
8.

Thanks for this post.

You say:
"Most GCMs are able to provide a reasonable representation of regional climatic features such as ENSO, the NAO, the Hadley cell, the Trade winds and jets in the atmosphere. They also provide a realistic description of so-called teleconnection patterns, such as wave propagation in the atmosphere and the ocean."

I would like to know, given that these (and other) regional features have not been studied all that long, how stable they are, in the mathematical sense of the term.

A second question is: what do you mean by "*reasonable* representation"? Any links to the primary literature on either question would be appreciated.

Thanks again. I look forward to more on the topic of how GCMs are constructed and parameterised.

P.S. There are a couple of typos in the article: aggrigated, parametereisation

re #5: "local" is an intentionally flexible concept. Often sub-continental, sometimes smaller, somewhere between 100 and 1000 miles.

[Response:The best link is probably to the IPCC AR4 chapter 8. -rasmus]

Comment by bender — 27 May 2007 @ 9:26 am
9.

There are distributed computing projects for SETI, protein folding, and crypto cracking ... but none for global climate change? OK, I've Googled after writing that line and I find http://ClimatePrediction.Net ... but I don't have the skill to determine if what they're doing is valid or not.

What doesn't RealClimate have a distributed computing initiative? This place certainly has the pull to get people interested in providing computing power.

And thusly (and quite tangentially) the model resolution could improve.

[Response:Climateprediction.net is a SETI-inspired initiative where GCMs are run as screen-savers. These GCMs are coarser than the 'normal' GCM run on super-computers, but the vast number of runs provide high 'statistical power' (a very large ensemble yields a large statistical sample). I don't know of any initiative where distributed computing has been used to for one high-resolution GCM (i.e. splitting the world into managable chunks of computation), and I think that would be very unpractible as this requires a high rate of data exchange. -rasmus]

Comment by neal rauhauser — 27 May 2007 @ 9:54 am
10.

Thanks for this post. You say:
"Most GCMs are able to provide a reasonable representation of regional climatic features such as ENSO, the NAO, the Hadley cell, the Trade winds and jets in the atmosphere. They also provide a realistic description of so-called teleconnection patterns, such as wave propagation in the atmosphere and the ocean."

I would like to know, given that these (and other) regional features have not been studied all that long, how stable they are, in the mathematical sense of the term. A second question is: what do you mean by "*reasonable* representation"? Any links to the primary literature on either question would be appreciated.

Comment by bender — 27 May 2007 @ 10:08 am
11.

Don't some of the GCMs (e.g. NCAR's CCSM3) try to address coarse grids by having models for different terrain within the grids, and then weighting them according to their area for the grid cell? Does this not help things?

You don't say much about what "skilful scale" means. Is this the same as some other supercomputer techniques when the grid size is not constant but instead varies as needed?

[Response:I don't think the literature is very clear on this (try googeling 'skilful scale'; I only got 20 hits!), and I thought it would be interesting to bring it up in this forum. -rasmus]

Comment by Earl Killian — 27 May 2007 @ 10:11 am
12.

Re #5: "local" is an intentionally flexible concept, something smaller in scale than "regional". The author defines this term in the opening paragraph in human terms, as the spatial extent over which most of us live most of our lives. Seems to me he's talking about something the size of a state or smaller. Doesn't much matter, as his point is that GCMs don't work well at those small scales. Hence the attempt to define the minimum scale over which the models are skillful. "Local" is therefore anything smaller than that.

Comment by bender — 27 May 2007 @ 10:19 am
13.

Thank you for the great job of this site.
I've got a question somewhat related to this article. You gave insights on the spatial resolution of the climate models. I am wondering what you can say on their temporal resolution. More precisely, climate models are predicting the mean temperature evolution. Are there any precision on the evolution of the thermal amplitudes, and on what time scale?
Thank you again.
G. Gay

[Response:One important consideration is the size of the model's time-step (order of minutes), and then the type of time-stepping (integration) scheme matters. But I'm not sure what the exact answer is to this (others?). -rasmus]

Comment by Guillaume Gay — 27 May 2007 @ 10:40 am
14.

Re #6: [There are distributed computing projects for SETI, protein folding, and crypto cracking...]

This comes down to a fundamental difference in computational methods. The problems you mention can be broken down into many computationally independent pieces. Each of the pieces can be worked on independently, and the results collected and combined whenever they're done.

Climate models (and many other problems) work on a grid. At each timestep, values are computed within each cell. Those computations depend on the values of the previous timestep, and the values in adjacent cells. That means there's a lot of data exchange going on. This might happen in memory on a single processor machine (very fast), or via a dedicated high-speed network on a parallel machine (e.g. IBM BlueGene). But if you tried to do this over the internet, the communication time would be very much larger than the time needed to do the computations themselves.

Comment by James — 27 May 2007 @ 10:44 am
15.

I have been interested in the regional and local climate changes partially because as Rasmus writes that is what effects people's daily lives, but also because regional and climate changes will need to be better predicted to determine the ecological changes caused by climate change.

Are there any barriers that prevent better regional and local climate predictions? For example are there problems in regional vs global models like the difference between climate and weather prediction (ie weather is chaotic and therefore less predictable than climate) that make it impossible to make better local/regional predictions, or is it just a question of researching more and developing better models?

[Response:Good question. I suppose in theory, one could always go down in scale, and when taking it to the extreme, to the scale of atoms (quantum physics). At one point, I expect the downscaling will becom impractical, at least. -rasmus]

Comment by Joseph O’Sullivan — 27 May 2007 @ 10:48 am
16.

Google: climate distributed computing
First two Results of about 1,130,000 for climate distributed computing. (0.20 seconds)

BBC - Science & Nature - Climate Change ...experiment used a technique called distributed computing to utilise users' spare computing power to predict future climate. http://www.bbc.co.uk/sn/climateexperiment/theexperiment/distributedcomputing.shtml

Distributed computing tackles climate change. Posted by Stephen Shankland. Oxford University and the BBC have launched a partnership .... news.com.com/8301-10784_3-6041683-7.html

Comment by Hank Roberts — 27 May 2007 @ 11:05 am
17.

re: #6
#10 is right, but in addition:

The problems for which @home distribution works have some other characteristics as well:

EITHER:
1) There are a large number of independent input cases, each of which can be analyzed separately, using a modest amount of input, and yielding a simple answer that is easy to verify:

a) YES/NO ("hello Earth. Why aren't you answering us?"
or INTERESTING/NOTINTERSTING, as particle physicists have long done, i.e., send an event to each free machine, have it crunch for a while, and say whether or not something interesting happened worth further analysis.

b) A few numbers, as in crypto-cracking: "here are the prime factors", which make take a long time to find, but are trivial to verify by multiplying them back together.

c) A number, which is mainly of interest if it's the bet found so far, i.e., as in Monte Carlo approaches to protein folding or Traveling Salesman routing problems.

OR
2) One is doing a Monte Carlo simulation where there is no right answer, but one is interested in generating an ensemble of results, and analyzing the distribution, i.e., "Do you have enough money to retire?" A delightful short piece on such is Sam Savage's "The Flaw of Averages": http://www.stanford.edu/~savage/flaw/

I think the ClimatePrediction.net effort is of this sort, and it may be useful, but it doesn't help the topic discussed in the (nice) original posting.

Note that going from a 100km grid to a 10km grid means a 10x10 = 100X more elements (2D), and if it were general 3D, that would be 1000X. In many disciplines, people using such methods have had to do non-uniform-sized subgridding to improve results with a given level of computing. I.e., for some parts of a model, a coarse grid gives reasonable results, but for others, one needs a much finer grid.

Comment by John Mashey — 27 May 2007 @ 12:23 pm
18.

This is all very interesting but as far as I'm concerned fails to give an accurate enough picture of the processes at work behind the differences in predictions.

What I would like to see is a much more detailed run-through of the differences in local predictions by models and concurrently, the unique ways in which they simulate natural processes.

If 8 grid points represents a skillful representation and 1 does not, can you give a relevant example of how regional dynamics cancel each other out at that scale?

[Response:I don't think it's necessary a matter of 'cancelling out' but rather giving a 'blurred' picture or for instance a spurious geographical shift of a few grid points due to e.g. approximations of local grid-point scale spatial gradients. -rasmus]

Comment by J Bloom — 27 May 2007 @ 12:59 pm
19.

I am running a distributed model for climateprediction.net. I would just as happily run one for realclimate. I am not a scientist and real climate has given me the insight and the links to further education which enable me to refute the flat earth rightwingers on the political board I frequent. They are no longer as dismissive and abusive as they were two years ago, when I began studying the information on your website. Thanks and if you need some of my cpu cycles, I will be glad to donate.

[Response: Watch this space... - gavin]

Comment by John — 27 May 2007 @ 1:37 pm
20.

I'm curious about this example of a regional-local prediction:

Model Projections of an Imminent Transition to a More Arid Climate in Southwestern North America, Seager et al. Science 2007
Abstract: "How anthropogenic climate change will impact hydroclimate in the arid regions of Southwestern North America has implications for the allocation of water resources and the course of regional development. Here we show that there is a broad consensus amongst climate models that this region will dry significantly in the 21st century and that the transition to a more arid climate should already be underway. If these models are correct, the levels of aridity of the recent multiyear drought, or the Dust Bowl and 1950s droughts, will, within the coming years to decades, become the new climatology of the American Southwest."

This is within the 'skillful scale'. However, if one wants to know the effect on California's Central Valley (an area of massive agricultural production) and on Sierra Nevada snowpack levels, it seems the models still lag behind the observations - but water districts should probably be focusing on long-term conservation strategies right now.

This paper, based on observations, seems to provide support for Seager et al :Summertime Moisture Divergence over The Southwestern US and Northwestern Mexico, Anderson et al GRL 2001

They note that moisture divergence over the American Southwest increased in 1994, in line with model predictions of persistent drought.

Droughts are a common feature of the 'unperturbed climate system' but this change appears attributable to anthropogenic climate change. Seager et al report that the persistent drying is due to increased humidity, which is changing atmospheric circulation patterns and leading to a poleward expansion of the subtropical dry zones. Increases in atmospheric humidity are a long-standing prediction of the effects of increased anthropogenic greenhouse forcing.

Comment by Ike Solem — 27 May 2007 @ 1:45 pm
21.

Surely very high resolution experiments were done on smaller time scales covering wide regions, wasn't there any significant change in results?