Explanation of the Nature Journal - First CPDN Results

From Unofficial BOINC Wiki

Jump to: navigation, search

Contents

[edit] General

Nature is a very prestigeous scientific journal. CPDN's has had their paper go through the extensive peer review process necessary for it to be approved for publication. It was published on 27th January 2005. The paper is available here.

The authors are D. A. Stainforth, T. Aina, C. Christensen, M. Collins, N. Faull, D. J. Frame, J. A. Kettleborough, S. Knight, A. Martin, J. M. Murphy, C. Piani, D. Sexton, L. A. Smith, R. A. Spicer, A. J. Thorpe & M. R. Allen

[edit] Abstract

Most papers begin with an abstract of what they are going to say. In this case:

The range of possibilities for future climate evolution needs to be taken into account when planning climate change mitigation and adaptation strategies. This requires Ensembles of multidecadal simulations to assess both chaotic climate variability and model response uncertainty. Statistical estimates of model response uncertainty, based on observations of recent climate change, admit climate sensitivities—defined as the equilibrium response of global mean temperature to doubling levels of atmospheric carbon dioxide—substantially greater than 5K. But such strong responses are not used in ranges for future climate change because they have not been seen in general circulation models. Here we present results from the ‘climateprediction.net’ experiment, the first multi-thousand-member grand ensemble of simulations using a general circulation model and thereby explicitly resolving regional details. We find model versions as realistic as other state-of-the-art climate models but with climate sensitivities ranging from less than 2K to more than 11 K. Models with such extreme sensitivities are critical for the study of the full range of possible responses of the climate system to rising greenhouse gas levels, and for assessing the risks associated with specific targets for stabilizing these levels.

[edit] Abstract context

The first models were Slab Models. This means there was no ocean with a slow response time. Consequently, the model reponds quickly and this makes it good for studying the equilibrium response because you do not need to run the model for a long time before a steady state is reached. The forcing applied in these models was a doubling of the carbon dioxide level. The Climate Sensitivity is defined in the above abstract. So it sounds like the models are designed to be perfect for measuring Climate Sensitivity. (There is actually a complication that may get a mention later.)

The range for climate sensitivity was less than 2K to more than 11K. Let's provide some background for that. A 'K' represents Kelvin; this is a similar temperature scale to the celcius scale. The size of 1K is equal to 1C. Kelvin is used by scientists because zero on that scale is equal to absolute zero which is -273C. The IPCC predict a temperature increase of 1.4 to 5.8C from 1990 to 2100. This is based on the climate sensitivity being in the range of 1.5 to 4.5C.

It was headline news that CPDN had found model versions as realistic as other state-of-the-art climate models with such high sensitivities. If the climate sensitivity was much more than 4.5 then the 5.8 suggested upper limit for the temperature rise this century would be understated. Note however that a sensitivity of 2C is also possible and the paper stated that they are not able to give probabilities for different sensitivities.

Other studies had reported the possibility of high sensitivity being consistent with observations but this is the first time such behaviour had been seen in a General Circulation Model.

[edit] What has Climateprediction.net got from each participant and how are results combined?

Global mean temperature for model 234404 - a stable run.

Amongst a lot of other things each Participant's model can produce a graph of global mean temperature. Further explanation of how to see this graph and what it means are in Checking on How Your Climateprediction.Net Work Is Progressing.

The climate sensitivity is the difference in temperature between Phase 2 and where Phase 3 would reach equilibrium temperature. This model hasn't quite reached equilibrium, however it is clear enough where it is heading. Processing more than 15 years in Phase 3 could be done to be sure about it but the error in estimating the equilibrium temperature from the shape of the graph so far is considered to be small. Therefore, it is considered better to process more models than continue processing for more than 15 years. An exponential extrapolation is used to calculate the equilibrium temperature. Details of this extrapolation are in the methods section of the Nature Paper.

[edit] How do the scientists combine the results?

If there are only a few runs they can be plotted onto a single graph like

http://www.climateprediction.net/science/sci_images/thcts2.jpg

However when there are too many to see clearly what is going on, they can be converted to density type graph like:

http://www.climateprediction.net/science/sci_images/results1.jpg

The more frequently occurring the pattern the more red the colour used. On this type of graph, the most frequently occurring sensitivity can be seen to be about 3.4 degrees (16.9-13.5). Some are getting over 22 degrees which is at least 8 degrees warmer (22-14) but clearly those haven't reached equilibrium yet. When the extrapolations are done, the sensitivities can be calculated to be up to at least 11 degrees C.

[edit] Why do the results cluster around 3.4 degrees?

There are a few possible explanations:

  • The relevant processes governed by the Parameters varied may only have a limited impact on sensitivity.
  • The parameter ranges used may have been too small.
  • Multiple perturbations may have effects which compensate when you look at a global average. There could be significant regional effects that disappear in the averaging process.

[edit] What other processing of results do Climateprediction.net do?

Figure 1b Frequency distributions of Temperature (colours indicate density of trajectories per 0.1 K interval) through the three phases of the simulation. Frequency distribution of the 414 model versions. In this graph temperature is shown relative to the value at the end of the calibration phase and where initial condition ensemble members exist, their mean has been taken for each time point.
Figure 1b Frequency distributions of Temperature (colours indicate density of trajectories per 0.1 K interval) through the three phases of the simulation. Frequency distribution of the 414 model versions. In this graph temperature is shown relative to the value at the end of the calibration phase and where initial condition ensemble members exist, their mean has been taken for each time point.

This graph is quite similar to the last but there are three differences. Some models have been rejected because they were not stable in the control Phase (2). This has removed the Cold Equator Models and others where the temperature just drifts too much.

Also temperatures are now being plotted relative to the temperature at the end of the calibration Phase (1).

Finally, this is using the average for each Initial Condition Ensemble. Therefore 414 Ensembles are being used instead of 2017 models (see below).


[edit] More about the models used

  • For this paper 2578 models were used with variations in 6 Parameters.
  • 461 were duplicates leaving 2017 different models.
  • 864 did not pass quality control tests on the control Phase (2) leaving 1154 models. Some of these are Cold Equator Models others exhibit just a slow drift. A higher proportion of models could probably have got though quality control tests by having a longer calibration Phase (1). However this would mean that they would take more time to complete so less models would have been completed. (It would also be less interesting for Participants.)
  • 6 Models passed quality control on the second Phase but exhibited 'Cold Equator Models' effects leaving 1148 stable simulations.
  • These 1148 stable simulations were averaged into 414 model versions or Initial Condition Ensembles. This means that within each of those 414 model versions all the Parameters were the same except for the initial conditions Parameter. The average Initial Condition Ensemble size was therefore less than 3.

[edit] What about the effects of Parameters?

Fig 2a The response to parameter perturbations. The frequency distribution of simulated climate sensitivity using all model versions (black), all model versions except those with perturbations to the cloud-to-rain conversion threshold (red), and all model versions except those with perturbations to the entrainment coefficient (blue).
Fig 2a The response to parameter perturbations. The frequency distribution of simulated climate sensitivity using all model versions (black), all model versions except those with perturbations to the cloud-to-rain conversion threshold (red), and all model versions except those with perturbations to the entrainment coefficient (blue).

This graph shows the frequency of different climate sensitivity models. The different distribution with different parameter perturbations may have roughly the same sort of shape but the differences are very significant. The blue distribution pattern has no models with sensitivity over 6.5 degrees; the red distribution has 4.9% of model versions with sensitivity of over 8 degrees celcius. To put this in perspective, it is widely agreed that a climate sensitivity of 5 degrees would be catastrophic.

This means that the distribution of sensitivities obtained by these models cannot be relied upon. This is because only 6 Parameters have been varied. There are other poorly constrained Parameters and we have not varied some of these yet. Therefore we do not know whether these would have the effect of making the distribution look more like the red or the blue distribution.


The next step is to try to look at the quality of the models:

Fig 2b, Variations in the relative root mean square error of model versions. The unperturbed model is shown by the red diamond. Model versions with only a single parameter perturbed are highlighted by yellow diamonds. The triangles show the CMIP II models for which data are available; HadCM3 (having the same atmosphere as the unperturbed model but with a dynamic ocean) is shown in red and the others in blue.
Fig 2b, Variations in the relative root mean square error of model versions. The unperturbed model is shown by the red diamond. Model versions with only a single parameter perturbed are highlighted by yellow diamonds. The triangles show the CMIP II models for which data are available; HadCM3 (having the same atmosphere as the unperturbed model but with a dynamic ocean) is shown in red and the others in blue.

R.M.S.E. stands for Root Mean Square Error. This is a measure of quality of the model. The operations are done in reverse order. So first the temperature of each cell is deducted from the observed value. These are the squared. This removes the negative signs; the square of both negative and positive numbers are positive. Mean is another word for average (this is area weighted). The square root is then taken. This is largely cosmetic to get the answer into an understandable unit of 'degrees celcius' rather than 'degrees celcius squared'. The lower the answer the higher the quality of the model.

So what does the graph show? The models generally appear to be of good quality compared to a different set of models the CMIP II model. The plan was to use the higher quality models rather than the low quality model. You might expect extreme versions of the model to be of lower quality than other models. This does not appear to be the case; the models appear of similar quality. Therefore this part of the plan to use the higher quality models cannot yet be put into action.

This combined with the choice of parameters changing the shape of the distribution means that the project is not yet able to reach its goal of producing an objective probability distribution function for simulated climate sensitivity.

It should be noted that this is at an early stage and the quality test being applied here is quite poor. It only uses annual averages and seasonal data might be a better test. Also this is only using the Slab Model so we are not yet using a dynamic ocean. Different types of climate data may also provide more stringient tests.


Fig 2c, Linear prediction of climate sensitivity based on summing the change in l for the relevant single-parameter perturbation model versions, to estimate l when multiple perturbations are combined. Error bars show the resulting uncertainty (^ one sigma) caused by the combination of a number of Dl values where each l has an uncertainty deduced from the initial-condition ensembles having only a single parameter perturbed. Linear predictions within one sigma of the simulated value are shown in green, between one and two sigma in black, and above two sigma in red. Mean uncertainties in the simulated value (two-sigma range, inferred from the initial-condition ensembles) are shown at the bottom for four regions of sensitivity (0–3, 3–6, 6–9, 9–12).
Fig 2c, Linear prediction of climate sensitivity based on summing the change in l for the relevant single-parameter perturbation model versions, to estimate l when multiple perturbations are combined. Error bars show the resulting uncertainty (^ one sigma) caused by the combination of a number of Dl values where each l has an uncertainty deduced from the initial-condition ensembles having only a single parameter perturbed. Linear predictions within one sigma of the simulated value are shown in green, between one and two sigma in black, and above two sigma in red. Mean uncertainties in the simulated value (two-sigma range, inferred from the initial-condition ensembles) are shown at the bottom for four regions of sensitivity (0–3, 3–6, 6–9, 9–12).

It would be nice if a way could be found to use a small Ensemble size rather than a large one as it would save a lot of crunching effort. It was believed that this is not possible due to non-linearities in the response but this had not been demonstrated before this paper.

This graph demonstrates that a simple linear assumption does not work. If it did work, you would expect to see as many red and black error bars above the line as below the line. Also the ratio of green to black to red error bars should be approximately in the ratio of xxx:yy:z.Update me Therefore the number of red bars representing large errors should be x% of the total but in fact over 20% of the bars are in fact red.

The red bars are predominately below the line. This means that if this linear technique was used, climate sensitivity would be more likely to be higher than the technique would indicate. This effect looks like it will become more pronounced the more Parameters are varied and only six had been varied for the models used in this paper.


[edit] What about regional effects?

Figure 3 The temperature (left panels) and precipitation (right panels) anomaly fields in response to doubling the CO2 concentrations. a, b, The unperturbed model (simulated climate sensitivity, 3.4 K). c, d, A model version with low simulated climate sensitivity (2.5 K). e, f, A model version with high simulated climate sensitivity (10.5 K). These fields are the means of years eight to fifteen after the change of forcing is applied, averaged over initial-condition ensemble members; they are not the equilibrium response.
Figure 3 The temperature (left panels) and precipitation (right panels) anomaly fields in response to doubling the CO2 concentrations. a, b, The unperturbed model (simulated climate sensitivity, 3.4 K). c, d, A model version with low simulated climate sensitivity (2.5 K). e, f, A model version with high simulated climate sensitivity (10.5 K). These fields are the means of years eight to fifteen after the change of forcing is applied, averaged over initial-condition ensemble members; they are not the equilibrium response.

Figure 3 shows the Initial-Condition Ensemble average of the temperature and precipitation changes for years 8–15 of Phase 3 which has doubled CO2 concentrations. This is shown for for three model versions: (1) the unperturbed model; (2) a version with low sensitivity; and (3) a version with high sensitivity.

There is a known pattern for temperature changes where the polar regions get warmed by more than other parts of the world. This is known as polar amplication (see here for some explanation). Even in the low-sensitivity model version the warming in certain regions is substantial, exceeding 3 K in Amazonia and 4K in much of North America.

The precipitation field shows a greater variety of response. For instance, this particular low-sensitivity model version shows a region of substantially reduced precipitation east of the Mediterranean; something not evident in either the standard or high sensitivity model versions shown. It is critical to note that model versions with similar sensitivities often also show differences in such regional details. If only a small Ensemble was used it would not be possible to see how consistent the patterns are. With a large Grand Ensemble, the significance of such patterns can be considered.


[edit] Concluding Comments

The paper ends with some concluding comments thanking the Participants and stating they have "been able to discover models that have comparatively realistic control climates and with sensitivities covering a much wider range than has ever been seen before. This is a critical step towards a better understanding standing of the potential responses to increasing levels of greenhouse gases, regional and seasonal impacts, our models and internal variability." Future plans include doing a grand ensemble of Transient Simulations.

[edit] Concluding Comments Context

So why is discovering such models so critical? If the models sampled are only from part of the Parameter space where models are of good quality, then we could not rely on the probability information obtained being objective. The experiment needs a wide range of behaviours so that unrealistic behaviour models can be ruled out. The quality tests at Figure 2b indicate that this has not happened yet. However the sensitivity range of 2 to 11 degrees is very wide and better quality tests are expected to rule out some of the higher sensitivities. If anything it is worrying that there are not models with lower sensitivities. However, this was with early models that did not have a sulphur cycle built in. When these are included the range is expected to be extended downwards.


[edit] Also See

Retrieved from "http://www.boinc-wiki.info/Explanation_of_the_Nature_Journal_-_First_CPDN_Results"
Personal tools