 |
 |
| Pielke Home | People | Publications | Images | Courses | News | Links | Contact |
| RESEARCH GROUPS @ CIRES > |
Dynamic Downscaling Group
Welcome to the discussion group on dynamic downscaling! This site provides a chronological listing of e-mails on the subject of dynamic downscaling using regional atmospheric models and recent research on this topic. If you are interested in joining and are working on research on this topic, please contact Dallas Staley at dallas [at] cires.colorado.edu to be added.
Date: Mon, 21 May 2007 12:21:15 -0600 (MDT)
From: Roger A Pielke Sr.
To: downscale group
Subject: list
Hi All, This is the inaugural e-mail to a discussion group on the value-added of dynamic downscaling using regional climate models. Using the class of downscaling presented in
Castro, C.L., R.A. Pielke Sr., and G. Leoncini, 2005: Dynamical downscaling: Assessment of value retained and added using the Regional Atmospheric Modeling System (RAMS). J. Geophys. Res. - Atmospheres, 110, No. D5, D05108, doi:10.1029/2004JD004721.
these are Type 2, 3 and 4 downscaling (i.e. where initial conditions are "forgotten").
The questions are:
1. Does dynamic downscaling add value (skill) to a climate simulation over and beyond what is achieved by interpolating global climate model predictions (or reanalyses) to a finer grid resolution of terrain and other landscape features?
2. While we know dynamic downscaling adds spatial structure, are
i) the atmospheric features at the spatial resolution of the global model improved in accuracy within the regional climate model domain?
ii) are skillful atmospheric features at a finer spatial resolution than the global model resolution obtained by the regional climate model?
3. As a function of the size of the regional climate model domain, what is the relative role of lateral boundary conditions and surface boundary conditions in the simulation?
Please add to as necessary. Also, let Dallas and I know who should be added to this list (and when you reply, please reply to all listed above).
I look forward to a fruitful exchange of ideas on this subject.
Chris Castro will be taking the lead to prepare a review article on this subject; as these e-mail exchanges evolve, please let us know who would want to join us as co-authors (and run model tests to evaluate the questions that are written above and others that are added).
Roger
Date: Tue, 29 May 2007 13:20:30 -0600 (MDT)
From: Roger A Pielke Sr.
To: downscaling
Subject: Re: [Tropical-storms] New BAMS article on regional downscaling of Atlantic hurricane
----------------------------------------
Hi All, FYI; a valuable test for them to do is to increase the size of their model domain to see if the results change. Roger
--------------
Date: Tue, 29 May 2007 14:23:54 -0400
From: Brian Mapes
To: Tom Knutson
Cc: tropical-storms@tstorms.org
Subject: Re: [Tropical-storms] New BAMS article on regional downscaling of Atlantic hurricane activity
Congratulations GFDL group on a very interesting study!
If I I understand right, the regional model is given fields of humidity and circulation with observed (analyzed) structure on the scale of this (attached fig) box's average (~1/16 of your domain), with realistic time evolution.
The model then creates all the internal details, and after tuning (for both total storminess and the magnitude of year-to-year variations) it reproduces ~80% of the interannual variance in hurricane count and ~50% of ACE and PDI variance.
I'd be interested to hear how surprising this is to list members. I have no intuition on these scales.
Might a longtime Atlantic watcher be able to look at animations of, say, low-level winds and precipitable water, truncated at this fuzzy scale, and guess seasonal hurricane counts about as well? It'd be interesting to see such animations, if they wouldn't be too hard to make...
Or is there more to the large-scale flow that is important (vertical structure details? like a shear measure at least? ...or might that end up implicit in horizontal circulation and humidity, since storm fate becomes part of the box average?)
Or is there more to the large-scale flow that is important (vertical structure details? like a shear measure at least? ...or might that end up implicit in horizontal circulation and humidity, since storm fate becomes part of the box average?)
Could a statistical model be tuned do as well or better from (a subset of) the same inputs? How simple could it be?
Or are there subtler skill sources in the dynamical model that could actually be mined to advance understanding of storm-environment interactions on these scales?
Perhaps we will also hear soon about that big NCAR regional model simulation, with even less information input on regional scales (whole tropical belt running free).
Interesting times, the era of dynamical seasonal TC prediction seems to be on the horizon...(or is it here already?)
Brian Mapes
On May 24, 2007, at 5:33 PM, Tom Knutson wrote:
> Dear Colleagues,
>
> Not to try upstage Jim Kossin, but we also have a paper just accepted at BAMS on Atlantic hurricane activity.
>
We report that using a regional modeling framework, nested within NCEP reanalyses and using interior large-scale nudging, we can reproduce annual hurricane counts (1980-2006) in the Atlantic to a surprising degree (correlation of 0.87), including the rise in hurricane activity during the period. Other metrics such as ACE are also simulated rather well, as is the dependence of activity on the state of ENSO.
>
> We feel this is promising tool for exploring mechanisms behind past changes in Atlantic hurricane activity, as well as for making future projections based on climate model projections of large-scale changes in Atlantic climate (e.g., as recently documented by Vecchi and Soden).
>
Comments welcome.
>
> http://www.gfdl.noaa.gov/~tk/manuscripts/BAMS_2337_Rev.pdf
Date: Wed, 19 Sep 2007 08:26:05 -0700
From: Yongkang Xue
To: Downscale Group
Subject: Re: dynamic downscaling
Hi Roger, Thanks for your information and ideas on downscaling issues. Our downscaling paper has recently been published in J. of Climate (Xue et al., 2007: Assessment of Dynamic Downscaling of the Continental U.S. Regional Climate Using the Eta/SSiB Regional Climate Model, J. Climate, 20, 4172-4193). I attach it in this email for your information.
Best, Yongkang
Date: Tue, 2 Oct 2007 01:17:32 -0700 (PDT)
From: Barry Lynn
To: Yongkang Xue , Roger A. Pielke Sr., daniela.jocob
Cc: downscaling
Subject: Re: dynamic downscaling
Dear Yongkang: Thank you for sending your paper as an e-mail attachment.
You write that the model simulation was accutely sensitive to the position of the southern boundary, as
this determined the moisture transport from the southerly low level jet.
I am wondering if what you think your result would be if you had perfect lateral boundary conditions. Would the position of the boundary really matter to the same degree, or is what is important the moisture transport across the boundary -- which you don't simulate well without simulating the higher frequency waves in your regional domain?
Thank you for your reply,
Barry Lynn
Date: Tue, 02 Oct 2007 08:43:36 -0700
From: Yongkang Xue
To: Barry Lynn , Roger A. Pielke Sr., daniela.jocob
Cc: downscaling
Subject: Re: dynamic downscaling
Dear Barry, Thanks for raising this interesting question. I thought this issue for a while and still do not know if I complete understanding this. The following is my thoughts on this issue for your reference.
Looking at our results in Table 2, you could see when North American regional reanalysis (NARR) was imposed as LBC, Case 6 (its southern boundary was over Caribbean Sea) had much better results than Case 2 (same southern boundary position but with global reanalysis as LBC). So this result supports your speculation that better LBC may affect the importance of selection of boundary position. However, if comparing with Case 5 (also suing NARR as LBC but the southern boundary was over the Gulf of Mexico), Case 6 is still not as good as Case 5, which indicates even with an better LBC data set, the selection of boundary condition still matters.
I think the following factors may still have effects: (1) NARR is better than global reanalysis, but it is still not a perfect data set, especially over Caribbean Sea. (2) Eta's ability to simulate the moisture transport over ocean. We have not done any rigorous tests to evaluate this; I think Fedor and Zavisa may have better ideas on these two factors;(3) Dynamic effects. If this is the case, then nudge may be the only way to avoid this problem.
Please let me know if you have further questions/comments.
Best, Yongkang
Date: Tue, 2 Oct 2007 10:47:39 -0700 (PDT)
From: Barry Lynn
To: Yongkang Xue , Roger A. Pielke Sr. , daniela.jocob
Cc: downscaling
Subject: Re: dynamic downscaling
Dear Yongkang: (I think I have someone's (Daniela's) e-mail incorrect. Please correct.)
I think that your e-mail pretty much gets to the heart of the issues: boundary conditions do matter, and what you are really trying to do is to place the boundary of the model in the best location to either incorporate realistic (global or regional reanalysis) boundary conditions or in lieu of this to have the regional model "create them" for, let us say, an inner nest over your area of interest.
The additional problem we encounter, regardless, is that the regional model simulation is itself dependent not only on the large scale boundary conditions but also on how it "converts" those boundary conditions into meteorology (by choice of model parameterizations, etc).
The question I still have, though is this: global model might be able to reproduce characteristics of the general circulation on average, but can they realistically produce the higher frequency (and smaller scale) synoptic meteorology that drives regional model solutions and mesoscale model weather? If they cannot, it may not help us to use a general circulation at all to do dynamic downscaling, at least how they are currently used.
Barry
Date: Tue, 2 Oct 2007 15:06:46 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Barry Lynn
Cc: Downscale Group
Subject: Re: dynamic downscaling
Hi Barry I have copied this e-mail to Dallas to correct Daniela's e-mail address; she certainly should be in on this discussion. [I think I have corrected but have copied to Dallas to make sure]
My conclusion, based now on several papers, is that the RCMs cannot skillfully create the synoptic structures that we need. Spectral nudging, in particular, can assure that the large scale faithfully replicates the global model (or reanalysis) but the RCM is not going to be able to add
value to these synoptic features.
Roger
Date: Tue, 02 Oct 2007 17:29:09 -0700
From: Yongkang Xue
To: Barry Lynn , Roger A. Pielke Sr., daniela.jocob
Cc: downscaling
Subject: Re: dynamic downscaling
Dear Barry, Sorry for the delay in response because I had two classes today. We had one paper in JGR using the NCEP GCM as LBC (De Sales, F., and Y. Xue, 2006, Investigation of seasonal prediction of the South American regional climate using the nested model system, J. Geophys. Res., 111, D20107, doi:10.1029/2005JD006989.). That study shows the regional model produces better South American low level Jet and then the precipitation in La Plate Basin. But the precipitation in Amazon was deteriorated. We are just doing some further tests trying to see if the improvement in the regional model is purely due to better topography in the regional model.
Yongkang
Date: Tue, 2 Oct 2007 20:55:47 -0700 (PDT)
From: Barry Lynn
To: Yongkang Xue, Roger A. Pielke Sr., daniela.jocob
Cc: downscaling
Subject: Re: dynamic downscaling
Dear Yongkang, Roger: Roger wrote: "My conclusion, based now on several papers, is that the RCMs cannot skillfully create the synoptic structures that we need. Spectral nudging, in particular, can assure that the large scale faithfully replicates the global model (or reanalysis) but the RCM is not going to be able to add value to these synoptic features."
My response is that we need to better define what we mean by "value." Clearly, the WRF is used to downscale the Eta or GFS, and it add value by giving more realistic meteorology at higher grid resolution than either.
I think what Roger means (?) is that the RCM is not going to be able to produce more realistic meteorology if the synoptic meteorology by itself is not realistic -- unless (as Yongkang hinted) other factors such as topography, land use, provide additional information. Clearly, nudging wouldn't help if the boundary conditions from a global GCM were not realistic.
The assumption has always been that an RCM could take any run of the mill longwave pattern and improve upon it, just because it produces higher resolution information. Roger seems to say no it can't, but we have to be sure we have reconciled how the WRF can improve on the GFS (I run the WRF operationally from the GFS, and it does produce higher resolution, accurate information: see www.weather-it-israel.com/5kmaps.php). Is this solely because the boundary conditions are more realistic in the GFS than in the Reanalysis or Global Models, or because of higher resolution surface information? Or, can the RCM produce better information even when the longwave pattern meteorology is not more realistic (without nudging)?
Roger: Please e-mail me a paper you think I should read that best makes your point of the last e-mail. I probably have a reference, but the paper would be good to have in hand.
Thanks, Barry
Date: Wed, 3 Oct 2007 06:44:52 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Barry Lynn
Cc: Downscale Group
Subject: Re: dynamic downscaling
Hi Barry, My comments on downscaling are with respect to Type 2, 3 and 4 regional model runs. The downscaling from the weather prediction models is a Type 1 downscaling (as we define in Castro et al), and, I agree, does add significant predictive skill (i.e. value). The reason is that with Type 1 downscaling (say from the GFS), the regional model runs (e.g. from WRF) include observed initial conditions within its domain. This iniital condition knowledge is lost in Type 2, 3 and 4 downscaling.
For the Type 2, 3, and 4 downscaling, the regional model only has the lateral and bottom boundary conditions to insert information. With Type 2, the lateral boundary condition information is based on observations (i.e. reanalyses), yet we show that the larger scale (synoptic information) deteriorates unless nudging, particularly spectral nudging, is used. With this nudging, however, the regional model is now significantly constrained by the larger scale reanalysis. When you apply spectral nudging to the Type 3 and Type 4 simulations, the regional model, at the synoptic scale (i.e. the scale resolved by the larger models) is constrained by the larger model (with all of its biases).
I hope this clarifies. The new paper that has been submitted related to this topic is
Lo, J.C.-F., Z.-L. Yang, and R.A. Pielke Sr., 2007: Assessment of dynamical climate downscaling methodologies using the Weather Research and Forecasting (WRF) Model. J. Geophys. Res., submitted. http://www.climatesci.org/publications/pdf/R-332.pdf
Best Regards, Roger
P.S. I use the term "value-added" for the synoptic scale to mean that the accuracy (fidelity) of the simulation of features that are resolved by the larger scale model as improved using the regional model. For Types 2, 3, and 4, this does not, unfortunately, appear to be the case.
Roger
Date: Wed, 03 Oct 2007 15:05:42 +0200
From: Hans von Storch
To: Barry Lynn , Roger A. Pielke Sr.
Cc: Downscale
Subject: Re: dynamic downscaling
Roger, how would you see the Big Brother exp. and Feser's results in terms of your assessment?
Feser, F., 2006: Enhanced detectability of added value in limited area model results separated into different spatial scales. Mon. Wea. Rev. 134(8), 2180-2190. I did not follow too carefully this discussion, and I may have missed something.
We should be aware of the possibility that failure to reproduce observed states in type 2-simulations (climate mode run with re-analysis) may not be a malfunction of the model but a result to the chaotic behavior of the dynamics, namely that different interior solutions are consistent with the same lateral boundary conditions set (spectral nudging suppresses this effect). Thus, failure to reproduce observed conditions is not equivalent in falsifying the model. Instead we have to rethink "predictability", as we are used to do with global simulations.
We need a workshop to discuss the word "added value" generated (or not) by RCMs.
Cheers, Hans
Date: Wed, 3 Oct 2007 09:42:19 -0500
From: Zong-Liang Yang
To: Hans von Storch , Barry Lynn ,
Roger A. Pielke Sr.
Cc: Downscale Group
Subject: Re: dynamic downscaling
Dear Roger, Barry, Yongkang, Hans et al., Thank you for nice discussions. I agree with Hans that a workshop needs to be organized to further discuss regional climate modeling and dynamic downscaling. Fillipo Giorgi has a recent review article on regional climate modeling (see attached) which is very interesting.
Attached also for your information is my talk entitled "Challenges in Downscaling Global Climate Information for Regional Water Resources and Air Quality Studies" presented at the GEOSS Workshop "Regional Decisions for Climate Change" (http://www.opengeospatial.org/event/070921workshop).
Best regards,
Liang
Date: Wed, 3 Oct 2007 10:00:38 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Hans von Storch
Cc: Downscale Group
Subject: Re: dynamic downscaling
Hi Hans, The idea of ensembles of regional climate model (and, of course, regional prediction model runs) is an excellent idea. However, your hypothesis that the same lateral boundary conditions could produce significantly different internal fields needs to be tested.
Moreover, since we have the observed interior information (i.e. from the North American Regional Reanalysis), an ensemble of runs, without spectral nudging, would still need to bracket the observed results on both the scales that are resolved on the spatial scale of the lateral boundary conditions and the finer scales of the regional models. If the regional model had skill, it would need to do this regardless of where the lateral boundary conditions are placed.
On the big/little brother experiments, I will re-review the Feser paper but suspect that spatial structure added is being confused with value added. Value added means that the improved spatial structure is accurate and this can only be assessed in comparison with observations (i.e. reanalysis).
A workshop on this topic is an very good idea!
Best Regards, Roger
Date: Wed, 3 Oct 2007 10:16:12 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Zong-Liang Yang
Cc: Downscale Group
Subject: Re: dynamic downscaling
Hi Liang, Thank you for sharing. On your powerpoint talk, you state that the regional climate model adds value, but this is the issue that needs to be further assessed. My conclusion is that spatial structure is added, but since it is so dependent on the larger scale model, biases in the input from that model cannot be corrected by the regional model. This is shown convincingly to me in our paper with you and Jeff, since the interior results are so dependent on how frequently the regional model is reinitialized.
On the Giorgi paper, it is disappointing that he did not comment on the Castro et al paper on downscaling issues.
Best Regards, Roger
Date: Wed, 03 Oct 2007 10:54:04 -0700
From: Yongkang Xue
To: Downscale Group
Subject: Re: dynamic downscaling
Dear All, I think there are a few issues in this discussion: (1). how to evaluate the RCM results. We have to have some quantitative means to assess RCMs and define what "add value" means. I think in addition to RMS, correlation, Equitable Threat Score, the spectrum analysis, which was described in Castro et al (2005) and used in our paper, show very promising. As to the spatial structure, I am not that sure if there is any better methods other than spatial correlation. ETC and spectrum analysis also implicitly provide information in this aspect.
(2). As to RCM downscaling ability, it seems think there are three opinions: add, not add, or conditional add. For the first type (i.e., weather forecast type), there is no different opinions. For the other types, especially the GCM LBC, opinions are quite different. Flipo seems at one end of spectrum. Due to RCM's special characteristics, .i.e., limited domain, imposed LBC, etc, I think RCM's ability is tied to some conditions.
(3). Workshop is a good idea so different opinions could exchange ideas. Roger may contact NOAA (since NOAA CPPA) is making efforts on that direction) to see if there is possible to get some support.
Yongkang
Date: Wed, 3 Oct 2007 12:31:35 -0500
From: Zong-Liang Yang
To: Roger A. Pielke Sr.
Cc: Downscale Group
Subject: Re: dynamic downscaling
Hi Roger et al., In my ppt talk, I said RCMs add value, but mainly in two cases as shown in Giorgi's (2006) review. He gave two examples, both for wintertime: (1) precipitation is improved over GCMs in mountainous regions, or areas of complex coastal lines, and (2) the frequency of occurrence of high intensity daily precipitation events is better simulated by RCMs than GCMs. I agree with you that in both cases the high spatial and temporal resolutions in RCMs are necessary conditions for the improved simulations. Of course, if possible biases in GCM forcing fields (e.g., winds, temperature, moisture) are corrected further, the above-mentioned high-resolution simulations are expected to be improved further.
Do RCMs add value in other regions or orther seasons? Clearly, more studies are needed.
In our recent study, we focus on the CONUS, and we use global reanalysis to drive WRF and compare with NARR. For precipitation, we find that dynamic downscaling does not provide good precipitation simulations other than over western coastal regions and the Rocky Mountains. Using nudging, however, good simulations can be obtained over almost the entire CONUS. If we want to have high-resolution (spatially and temporally) precipitation fields for water resource applications, dynamic downscaling with nudging is recommended, especially for the regions with limited rain gauge data. For other meteorological fields (e.g., winds and geopotential hights) which are citical for air quality applications, dynamic downscaling alone produces larger errors than a simple bi-linear interpolation from the global reanalysis, while downscaling with nudging is better than the bi-linear interpolation. This indicates that dynamic downscaling alone does not add value (relative to the interpolation) but it does when nudging is considered.
Best regards, Liang
Date: Wed, 3 Oct 2007 13:05:33 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Zong-Liang Yang
Cc: Downscale Group
Subject: Re: dynamic downscaling
Hi Liang (with respect to both of your e-mails), Thank you for the follow up. I have thought of a thought experiment (that could also be completed as a real simulation test).
Assume the region we are interested in is over open ocean with no surface SST gradients. Can the regional model provide increased accurate spatial and temporal structure?
For this experiment, it means the added information that the model has for Type 1 downscaling are the initial and lateral boundary conditions. For this case, there is a lot of evidence that the regional (weather) model adds skill at both the small and larger scale features without nudging of any sort even over the open ocean. As you noted in your second e-mail, we all agree on this.
For Type 2 downscaling, the reanalysis (which provides a relatively coarse sampling of continuous atmospheric fields) provides the only added information. If nudging is used, this real world data can be used to constrain the larger scale features, and let the finer scale features evolve as accurately as possible given the regional model dynamics and parameterized physics. Of course, the regional model than becomes dependent on the reanalysis for the larger spatial features even within the regional model domain.
For Type 3 downscaling, a GCM is run but with prescribed surface forcings (e.g. SSTs), so the constraint by the real world is much less than in Type 2 downscaling. For Type 4 downscaling, all real world constraints are removed except, for example, the atmospheric well-mixed greenhouse gas concentrations. Even if nudging is not used, the regional model is still dependent on the GCM for the larger scale features through the lateral boundary conditions, and, thus, in that sense is a slave to it. Biases/errors in the lateral boundary condition values cannot be corrected by the regional model.
The first research question for the open ocean experiment is the degree of dependence of the regional scale features on the lateral boundary conditions versus the surface forcing (in this case the uniform SSTs in the regional model). This issue focuses on what Hans has called the "flushing rate" of the regional model. This experiment should be done without any nudging, of course. The ensemble idea suggested by Hans should be tested.
The second question is can the dynamics and parameterized physics provide increased model skill? This is a lot to expect of the regional model, as it only has the parameterized (generally column model) physics with just the advection, Coriolis effect and pressure gradient forces on the regional model resolved scale available to process the incoming information from the lateral boundary and SST conditions. This question should be answered using real world data, not big brother/little brother experiments.
When terrain or other surface features are present, we do see additional spatial structure, but this makes the assessment of the actual value added much more difficult than for the open ocean experiment.
On a workshop, I suggest instead of NOAA, that, if we agree, one of you take the lead in preparing an article for EOS or BAMS. This would provide the framework for a workshop.
Best Regards, Roger
Date: Wed, 3 Oct 2007 16:44:40 -0500
From: Zong-Liang Yang
To: Roger A. Pielke Sr.
Cc: Downscale Group
Subject: Re: dynamic downscaling
Hi Roger, An EOS or BAMS article prior to a workshop is a great idea. I would like to take the lead in preparing a review article summarizing current literature and ideas (especially the new papers in the past three years), and circulate among us before submitting to BAMS or EOS. Using the Web of Science, I am impressed by the number and quality of the papers on regional climate modeling in the past few years. I would like to have the first draft by Oct 16. Any suggestions are welcome.
Best regards, Liang
Date: Wed, 3 Oct 2007 17:53:39 -0400
From: "Niyogi, Dev"
To: Downscale Group
Subject: RE: dynamic downscaling
Hi Liang, and Roger (and all), Thanks for the very informative discussions. We have been testing domain size, lateral and surface boundary conditions and grid spacing issues for the Indian monsoon region for simulating heavy rain episodes. If pertinent we can provide you summary of experiments for that region. One issue we have been struggling (which I am not sure was discussed in todays email exchange) is no matter what we decide as "optimal" lateral boundary conditions, the physical parameterizations and (both from details of surface processes as well as convection/ boundary layer parameterization) changed our results substantially. I am not sure if this is an issue that is specific over the Indian monsoon region though.
Thanks, Dev
Date: Wed, 3 Oct 2007 17:30:55 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Zong-Liang Yang
Cc: Downscale Group
Hi Liang, I am glad you are taking the lead! I suggest a focus on the questions we
are raising on the isse of value-added and how is this tested.
Best Regards, Roger
Date: Wed, 03 Oct 2007 16:34:13 -0700
From: Yongkang Xue
To: Roger A. Pielke Sr., Zong-Liang Yang
Cc: Downscale Group
Subject: Re: dynamic downscaling
Roger's idea is very good to have an article first to provide a better review on RCM issues. Flipo's review is too much one side story. I will support Zong-Liang's effort on this article.
I also agree that, we should work on real world rather than big-brother type of experiment although it might be necessary at the early stage. As to the ocean experiment, it is a very good idea to conduct ideal tests in a real world. There are two issues: (1) how to choose an area with no SST gradient; (2) how to verify results since the observation over ocean is not that reliable, so do reanalysis. These technical issues need to be discussed.
Date: Wed, 03 Oct 2007 16:42:15 -0700
From: Yongkang Xue
To: "Niyogi, Dev" , Zong-Liang Yang ,
Roger A. Pielke Sr.
Cc: Downscale Group
Subject: RE: dynamic downscaling
Dev, Physical parameterizations, such as convective schemes and land surface processes, always have
big impacts. I think over Indian monsoon regions, these processes are crucial. The LBC can not replace basic dynamics and physics in the model. You may try a relative stable period to see if the model is able to have a reasonable climate, or you may reinitialize your model every day to see if there are any problem for the weather forecast type of simulations.
Yongkang
Date: Thu, 4 Oct 2007 06:39:27 +0200
From: Hans von Storch
To: Downscaling Group
Subject: AW: dynamic downscaling
Let me just say that I am very much in favour of this approach. We have to make sure that the strong Skandinavian groups, in particular SMHI from Sweden, are joining. Also Rene Laprise is always a source of inspiration. My lab may be able to throw in a few kEURs in support for a workshop.
We had a very useful workshop on added value and related issues in Lund in 2004; ICTP wnted to have a follow up in March 2008, but I have never heard any details of that workshop again.
Roger's categories '1-4 of different types of simulations with RCMs are really useful, even if #3 is not really relevant in my view. Now we should try to assemble a list of issues, which form added value can take. One is the phenomena/structures approach of Rogers: are there new spatial features evolving? (applicable for #1-4)- another is what we in GKSS do: are we better in simulating medium scale variability (the Feser approach, applicable to #1-2)? A third one: do we get better distributions, in particular with respect to the tails of relevant variables? (#1-4, but with the problem that the provision of such added value in #3 and #4 depends very much on the mother GCM.)
Also - is this added value "trivial", i.e., we we get it also with considerably simpler methods, in particular geostatistical spatial interpolation (co-kriging topography)?
With respect to #3-4 we have to keep in mind the fact that, conditional upon the area size and region, the solution in the interior becomes intermittently independent from lateral boundary conditions - showing in an ensemble of simulation a variety of possibly developments in the interior, which are all consistent with a set of boundary values. This may be a virtue, given the purpose of #3/4 simulations.
Hans
Date: Thu, 04 Oct 2007 08:02:52 +0200
From: hvonstorch
To: Downscaling Group
Subject: Re: dynamic downscaling
Dear Roger, that the same lbc's can produce intermittently different solutions in the interior has been demonstrated several times by now, e.g.,
Ji, Y.M, and A.D. Vernekar, 1997: Simulation of the Asian summer monsoons of 1987 and 1988 with a regional model nested in a global GCM, J. Clim. 10: 1965-1979
Rinke, A., and K. Dethloff, 2000: On the sensitivity of a regional Arctic climate model to initial and boundary conditions. Clim. Res. 14, 101-113.
Weisse, R., H. Heyen and H. von Storch, 2000: Sensitivity of a regional atmospheric model to a sea state dependent roughness and the need of ensemble calculations. Mon. Wea. Rev. 128(10), 3631-3642
von Storch, H., H. Langenberg and F. Feser, 2000: A spectral nudging technique for dynamical downscaling purposes. Mon. Wea. Rev. 128: 3664-3673
Feser, F., and H. von Storch, 2007: A dynamical downscaling case study for typhoons in SE Asia using a regional climate model. Mon. Wea. Rev. (in press)
The Feser study (Feser, F., 2006: Enhanced detectability of added value in limited area model results separated into different spatial scales. Mon. Wea. Rev. 134(8), 2180-2190.): an intercomparison is made with high-resolution regional analyses done by the German Weather Service (DWD) for a limited period of years. After separating the RCM and analysis fields into different scales, pattern correlation coefficients are calculated for NCEP vs. DWD, RCM/no nuding vs DWD and RCM/spectral n vs. DWD. Two variables are considered: temp and SLP. The RCM is worse than NCEP on large scales, but better than NCEP for medium scales, at least for spectral nudging.
When talking about added value, one needs two well defined references Feser has used as references the DWD analyses (which should be replicated) and the NCEP reanalyses (compared to which the tested model must be better). Something like a Brier-score.
Regards, Hans
Date: Thu, 04 Oct 2007 07:48:50 +0200
From: Hans von Storch
To: Downscaling Group
Subject: Re: dynamic downscaling
The idea of correcting large-scale errors by the use of better resolution in a RCM has long, and unsuccessfully, tried by Bennert Machenhauer. To have this idea work we would need processes, which need high resolution representation, but have large-scale effects. What conditions would cause these processes to respond with a spatially uniform/large-scale? The atmospheric/oceanic large scale state or the soil state, which again depends on the large scale atmos state. Thus, such processes can, and are, parameterized in the coarser resolution (global) models.
If we find an RCM to correct systematic errors of a global model, we likely could strip the RCM down to a much smaller code, which may act as a parameterization in a GCM.
Hans
Date: Thu, 4 Oct 2007 06:17:26 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Downscaling Group
Subject: Re: dynamic downscaling
Hi Hans, This is slightly off focus, but in developing parameterizations, we have recommended a new approach, as we described in
Pielke Sr., R.A., D. Stokowski, J.-W. Wang, T. Vukicevic, G. Leoncini, T. Matsui, C. Castro, D. Niyogi, C.M. Kishtawal, A. Biazar, K. Doty, R.T. McNider, U. Nair, and W.K. Tao, 2007: Satellite-based model parameterization of diabatic heating. EOS, Vol. 88, No. 8, 20 February, 96-97.
We discuss diabatic heating in that paper, but the methodology can be used for all processes with models that are not basic physics.
With our approach, ultimately observations are used to create the parameterizations, thus bypassing the errors in the traditional way we parameterize processes.
Roger
Date: Thu, 4 Oct 2007 06:28:13 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Hans von Storch
Cc: Downscaling Group
Subject: Re: AW: dynamic downscaling
Hi Hans, Thank you for your offer of support for a Workshop! I suggest that in preparing the EOS/BAMS paper, that can be used to provide the foundation to justify the Workshop, that Liang contact SMHI and Rene (and Filippo and others) to join, if they chose, in such a paper where the issues are identified, and the different perspectives presented.
I do have two questions in response to your e-mail below:
1. Why is Type 3 not relevant? This is the case where exterior surface climate information (e.g. SSTs; which changes on time periods corresponding to the length of time of the climate simulation) is
prescribed from observations (as the AMIP runs have done). In Type 4 runs, the SSTs are predicted.
2. I do not see any reason that the interior can intermittently become independent from the lateral boundary conditions, unless you mean, for example, when a large, stationary subtropical ridge overlays a region. Please clarify. In a long term model simulation, the interior cannot remain independent, so that this special case is only applicable to short term model runs.
Roger
Date: Thu, 4 Oct 2007 06:42:42 -0600 (MDT)
From: Roger A. Pielke Sr.
To: hvonstorch
Cc: Downscaling Group
Subject: Re: dynamic downscaling
Hi Hans, I agree with you. The creation of different interior solutions with almost identical lateral boundary conditions is further evidence of the nonlinearity of the climate system. The use of ensembles with perturbed lateral B.C.s (and parameterizations, such as Barry has done and re initializations as Jeff and Liang have done) should be priority to propose in the paper that Liang is leading, and for a Workshop.
Your studies and methodology to evaluate value added will be central to this discussion. In my view, runs without any short of nudging represent an essential benchmark in such evaluations, as they illustrates the limitations in the regional models to create real world structure in the absence of external constraints except for the lateral boundary conditions. The assessment using Type 2 regional climate runs is needed in order to make statements in our paper on value-added from Type 3 and Type
4 regional climate model runs. Your Brier-score like concept is an excellent idea for such assessments with the Type 2 runs!
Roger
Date: Thu, 04 Oct 2007 14:48:48 +0200
From: hvonstorch
To: Downscale Group
Subject: Re: AW: dynamic downscaling
Roger, I find #3 pretty irrelevant, because I consider the AMIP simulatons are purely technical with little scientific interest; the output is useful in most cases only for testing the models, not for generating data, which may be later used in other studies, or testing dynamical hypotheses.
The intermittent indeterminacy - maybe we misunderstand each other; we have spoken about this issue before, and had agreed, if I remember correctly. But for the clarity of the discussion, also for the others participating in the e-mail exchange, I repeat the argument here.
of course, the lateral boundary conditions exert an influence, but they do not determine what happens in the interior (if there is not strong flushing, so that the information is really advected quickly enough through the model domain). Thus the interior is not "independent" but also not "determined", but only conditioned by the LBC. An area where this is not the case is the Arctic, and it is no wonder that Rinke and Dethloff got strong variability there in an ensemble of simulations. Also in the tropics this may be a frequent phenomenon. In westerly regime in mid latitudes the phenomenon is rare, but not absent. We have now submitted an publication with polar lows in the #2-set-up. You get very different cyclogenesis when you begin your integration one day earlier or later (when you start the integrations weeks before the polar low forms).
In the literature I gave earlier, several such cases are documented.
It is really surprising that the regional modellers needed 25 years longer than the global modellers that there is ensemble variability - there are many papers testing parameterizations with just one pair of a control and a "treatment" (changed param.) simulation - without any reference to uncertainty in response due to this effect. Already in the early 1970s Chervin and Schneider, among others, introduced the need to do statistical testing in such analyses - but the RCM people happily continued (and often still do). This is not issue in #1, but in #2-4.
Regards, Hans
Date: Thu, 4 Oct 2007 09:48:35 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Hans von Storch
Cc: downscaling
Subject: Re: AW: dynamic downscaling
Hi Hans, We certainly agree on this; thank you for your follow up clarification. As we discussed previously we need to quantify further this issue of "flushing" (i.e. the degree to which the lateral boundary conditions "condition" the interior) for different regions, and you already have publications on this subject, as do we. The relative role of the nudging (including spectral nudging) in this conditioning also needs further exploration with real data.
On the AMIP runs, I agree with you that the scientific value is limited, but feel it is still necessary to identify these types of model runs as a category, as there has been quite a bit of effort and funding of them in this mode.
With Best Regards, Roger
Date: Thu, 4 Oct 2007 12:18:20 -0700 (PDT)
From: Barry Lynn
To: Roger A. Pielke Sr. , Zong-Liang Yang
Cc: Downscale Group
Subject: Re: dynamic downscaling
Dear Roger, Hans, Liang: Perhaps you mention this (in a later e-mail), but I would like to add:
Compare my 4 km weather maps at http://www.weather-it-is-israel.com/4kmaps.php to
http://www.weather-it-is-israel.com/lcities.php (maps at 36 km grid resolution. The maps are not up to date because I had a computer failure, but the point is the same. Note, the maps at 36 km resolution
include the 4 km nest in the middle of their domains (where you should focus your attention).
1) A comparison of the maps show that the higher resolution simulation clearly adds spatial resolution
of the meteorology.
2) The meteorology at the higher resolution scale should be more accurate because the higher resolution domain resolves the underlying topography (including land elevation and surface type). This information is included in the 4 km simulation, but would not be included in a 45 km resolution simulation.
---> Without higher resolution surface information, I am still not sure I would concede Roger's point simply because the higher resolution model can better reproduce physical processes that lead to
meteorologically observable circulations and clouds, something the coarse model cannot do.
Barry
Date: Thu, 4 Oct 2007 12:39:20 -0700 (PDT)
From: Barry Lynn
To: Downscale group
Subject: Re: dynamic downscaling library?
Dear Roger/rest: We have a lot of very interesting e-mails be passed along and papers as well. Could someone make a record, and perhaps agree to host pdf versions of papers on downscaling that we could download at our leisure? Barry
Date: Thu, 4 Oct 2007 12:22:27 -0700 (PDT)
From: Barry Lynn
To: Downscale Group
Subject: Fwd: Re: PDF of our paper?
Hi: I have attached a paper we should have published soon in Climatic Change (Quantifying the sensitivity of simulated climate change to model configuration by Lynn et al. 2007). It details the possible sensitivity of present and future climate prediction to regional model configuration.
Barry Lynn
See also: Lynn et al., 2007: An Analysis of the Potential for Extreme Temperature Change Based on Observations and Model Simulations. J. Climate, DOI: 10.1175/JCLI4219.1
Date: Thu, 4 Oct 2007 14:54:27 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Downscale Group
Subject: Re: dynamic downscaling
Hi Barry, Are you referring to weather prediction or long term climate prediction? If the former, I agree with you. If the later, how did you test the results with observations?
Roger
Date: Thu, 4 Oct 2007 14:24:07 -0700 (PDT)
From: Barry Lynn
To: Roger A. Pielke Sr.
Cc: downscaling
Subject: Re: dynamic downscaling
Dear Roger: I was referring to weather prediction.
We also did regional model simulations of climate with a 108 vs 36 km grid domain driven by the GISS GCM. Comparing to temperature and precipitation observations, the 36 km domain did better, but I am not sure that this means that the regional model produced more realistic wave patterns, or was just better tied to the topography.
I did note that the regional wave pattern was very sensitive to the choice of cumulus parameterization in the regional model. When one scheme produced too frequent rain, this led to a relative cooling and
building of a trough, rather than a ridge (which occurred when the other scheme produced less frequent
rain and warmer surface temperatures).
One thing to keep in mind, though, good climatology consists of good meteorology, so I am not sure how that latter one can be true but not the other.
Barry
Date: Thu, 4 Oct 2007 16:00:06 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Barry Lynn
Cc: Downscale Group
Subject: Re: dynamic downscaling
Hi Barry, Thank you for the clarification. We agree completely on the added skill of downscaling for weather prediction (i.e. where the reigonal model has initial conditions to work with).
On the GISS runs, how did you verify?
I like your statement of good climatology consists of good meteorology! This means that the Type 2, 3 and 4 must faithfully replicate the synoptic features on an event basis, and that the statitical recreation of the climatology is not sufficient.
Best Regards, Roger
Date: Thu, 4 Oct 2007 17:28:47 -0500
From: Zong-Liang Yang
To: Downscale Group
Subject: Re: dynamic downscaling
Hi Roger and Barry et al.,
I also like the statement of good climatology consists of good meteorology (as Climate is the average and variations of weather over long periods of time). For quite a while, climate modelers are not very concerned about their models' ability or inability to model individual storms or weather events. Generally speaking, they are happy as long as their models can reproduce the mean climatology, mean circulation patterns, monthly mean rainfall, etc. Now there is a tendency for modelers to look at the
higher-order statistics of modeled variables. This trend is likely due to the fact that the society cares more and more about extreme events and their impacts. However, if we are doing a paleo-type of climate simulations, perhaps we care less about the individual weather events, partly because of
coarse-resolution nature of the proxy data.
Liang
Date: Thu, 4 Oct 2007 16:31:48 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Zong-Liang Yang
Cc: Downscale Group
Subject: Re: dynamic downscaling
Hi Liang, I add two comments here to the discussion. Climatology (long term weather statistics) is just part of what climate models need to simulate. Climate involves the oceans, land, and cryosphere in addition to the atmosphere, and involves physical, chemical and biological processes (NRC,
2005). Also, weather involves just a different statistical averaging time than climatology.
Roger
Date: Thu, 4 Oct 2007 22:02:45 -0700 (PDT)
From: Barry Lynn
To: Roger A. Pielke Sr.
Cc: Downscale Group
Subject: Re: dynamic downscaling
Dear Roger: Regarding: "good climatology consists of good meteorology:"
The conclusion, I suggest, implies that any type of downscaling (type 2, 3, and 4) would fail unless the
global model also simulates the meteorology, as well as climatology.
Regarding how we verified the GISS RCM results: we compared the mean, root mean square, and bias, to verify the RCM. It was clear to us that going from a 4 X 5 degree resolution of the GCM to a 1 X 1 degree RCM gave improvement in the simulated results. It was less clear that going to a 0.5 X 0.5 also did so. This raises the question whether the improvement gained from the RCM was also related to the use of better physics parameterization.
I also suggest that the horizontal scale of the topophography (surface boundary forcing) can perhaps
provide an upper limit to the ability of the RCM to downscale local meteorology, although the RCM might still produce better results on a more regional (average) scale.
Barry
Date: Thu, 4 Oct 2007 22:26:43 -0700 (PDT)
From: Barry Lynn
To: Zong-Liang Yang , Roger A. Pielke Sr.
Cc: Downscale Group
Subject: Re: dynamic downscaling
Dear Liang: I appreciate your receptivity to my suggestion.
My concern would be that simulating only climatology, but not higher order statistics is a tuneable
exercise. The problem with this, of course, is that we can't tune the model to accurately simulate changes in external forcings. This is because since the model actually fails to simulate the meteorology, it cannot be counted on to accurately simulate change in meteorology due to external forcings. Hence, its climate projections would be invalid, or at least suspect.
Barry
Date: Fri, 5 Oct 2007 06:20:07 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Barry Lynn
Cc: Downscaling Group
Subject: Re: dynamic downscaling
Hi Barry, In regard to your paragraph on the topic "good climatology consists of good meteorology", your conclusion is correct. This needs to be a requirement, since if a climate model can not be skillful when it has initial conditions, it certainly is not accurate when it does not have this observational constraint.
This is why tests comparing Type 1 and Type 2 downscaling with the same model are so useful.
Roger
Date: Fri, 5 Oct 2007 06:27:40 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Hans von Storch
Cc: Downscaling Group
Subject: Re: dynamic downscaling
Hi Hans, The requirement for applyng the regional models in a weather prediction mode means that they are as skillful as current operational codes. This would then mean that their representation of dynamics and physics is realistic on the weather prediction time scale. If they fail, however, to properly handle the dynamics of a propagating upper level trof, for example, even when the model is initialized with observed dats, we can have no confidence in any results they present with they are run in Type 2, 3, or 4 runs.
As to the types of model applications, I offer the breakdown that I published on Climate Science:
http://www.climatesci.org/2005/07/15/what-are-climate-models-what-do-they-do/
"What Are Climate Models? What Do They Do?"
"Climate models are comprised of fundamental concepts and parameterizations of physical, biological, and chemical components of the climate system, expressed as mathematical formulations, and then averaged over grid volumes. These formulations are then converted to a programming language so that they can be solved on a computer and integrated forward in discrete time steps over the chosen model domain. A global climate model needs to include component models to represent the oceans, atmosphere, land, and continental ice and the interfacial fluxes between each other. Weather models are clearly a subset of a climate model (a discussion of mesoscale weather models is given in Pielke, R.A., Sr., 2002: Mesoscale meteorological modeling. 2nd Edition, Academic Press, San Diego, CA, 676 pp), where the basic framework of all scales of weather models is presented). On the global scale, it is very important to distinguish global atmospheric-ocean circulation models (AOGCMs) from global climate models. Global climate models need to include all important components of the climate system as discussed in a 2005 National Research Council report, while AOGCMs up the present have not.
There are three types of applications of these models: for process studies, for diagnosis and for forecasting.
Process studies: The application of climate models to improve our understanding of how the system works is a valuable application of these tools. In an essay, I used the term sensitivity study to characterize a process study. In a sensitivity study, a subset of the forcings and/or feedback of the climate system may be perturbed to examine its response. The model of the climate system might be incomplete and not include each of the important feedbacks and forcings.
Diagnosis: The application of climate models, in which observed data is assimilated into the model, to produce an observational analysis that is consistent with our best understanding of the climate system as represented by the manner in which the fundamental concepts and parameterizations are represented. Although not yet applied to climate models, this procedure is used for weather reanalyses (see the NCEP/NCAR 40-Year Reanalysis Project).
Forecasting: The application of climate models to predict the future state of the climate system. Forecasts can be made from a single realization, or from an ensemble of forecasts which are produced by slightly perturbing the initial conditions and/or other aspects of the model. Mike MacCracken, in his very informative response to my Climatic Change essay seeks to differentiate between a prediction and a projection.
With these definitions, the question is where does the IPCC and US National Assessment Models fit? Since the General Circulation Models do not contain all of the important climate forcings and feedbacks (as given in the aforementioned 2005 NRC report) the models results must not be interpreted as forecasts. Since they have been applied to project the decadal-averaged weather conditions in the next 50-100 years and more, they cannot be considered as diagnostic models since we do not yet have the observed data to insert into the models. The term projection needs to be reserved for forecasts, as recommended in Figure 6 in Pielke Sr., R.A., 2002: Overlooked issues in the U.S. National Climate and IPCC assessments. Climatic Change, 52, 1-11.-225.
Therefore, the IPCC and US National Assessments appropriately should be communicated as process studies in the context that they are sensitivity studies. It is a very convoluted argument to state that a projection is not a prediction. The specification to periods of time in the future (e.g., 2050-2059) and the communication in this format is very misleading to the users of this information. This is a very important distinction which has been missed by impact scientists who study climate impacts using the output from these models and by policymakers."
Roger
Date: Fri, 5 Oct 2007 05:35:48 -0700 (PDT)
From: Barry Lynn
To: Roger A. Pielke Sr.
Cc: Downscaling group
Subject: Re: dynamic downscaling: meteorology vs climatology
Parts/Attachments:
----------------------------------------
Dear Roger: In my opinion, even climate models can be skillful when they are initialized with observations, or just tuned to mimic reality. Still, this won't help them in the long run when the tuning parameters have little relevance to the actual systems being simulated. So, I pretty much agree with you.
I was defining meteorology as those systems or scales that we observe on a weather map from day to day. A regional model could simulate the meteorology well, even if it drifts from reanalysis after a number of days without nudging. I know that if this is my only criteria we might then have differences in the mean fields of the regional model and reanalysis, but this is another problem we need to address in detail.
Another participant asked to what effort (downscale) we need to go to simulate meteorology. This is a very good question for a potential meeting. My suggested answer is that we can only reduce the resolution of the model to the point where it can accurately simulate synoptic systems (which probably means it requires meso-scale grids). Some would argue, though, that parameterized clouds are not realistic in how they distribute hydrometeors. In this case, we need explicit simulations of clouds to simulate even synoptic systems.
If we don't simulate the clouds correctly, then the radiation balance will not be realistic, let alone other feedbacks that will occur to move the model simulation away from "reality."
Barry
Date: Fri, 5 Oct 2007 09:07:27 -0400
From: "Niyogi, Dev"
To: Downscaling group
Subject: Re: dynamic downscaling: meteorology vs climatology
All, Barrys email below reminded me of a meeting we had here re ag sector needs. So here's another aspect we may additionally consider re model performance,skill and value. We have been having discussions with agriculture users on what they would find useful/skillful from a seasonal to more long term climate model/outlook. The overwhelming response we received for growing season is,
1. Temperature accuracy is good but not critical unless it is for major changes in below feezing and frost occurrence probability. The users seem to be ok with large errors in temperatures away from freezing.
2. Both for temperature and rainfall, they were quite satisfied to only know a range of how much will it be below normal or above normal for the season. Further they desired this at monthly resolution. An error or upto 3F seemed acceptable! So the user requirements are in some sense very modest as they look to us for information beyond weather forecast as they good make operational decisions.
So my point is that in the discussion we have on what consitutes skill, we need to have the user or sector needs beyond the traditional statistical measures.
Dev .
Date: Fri, 05 Oct 2007 11:48:39 -0700
From: Yongkang Xue
To: Downscaling Group
Subject: Re: AW: dynamic downscaling
Hi Roger, Hans, Barry, and others, This is indeed a very interesting discussion. Sorry I could not catch up most of your discussions. My two classes this quarter really keep me busy. The following is my thoughts on two issues:
(1). For the AMIP type experiments, I agree with Han that there is not much scientific interest in
AMIP comparisons. This type of comparisons just states the model results difference or shows some
good or bad parts of model results. It is not very interesting. However, if the model comparison is issue (or hypothesis)-orientated, it would be different. Thus far, there are different results from RCMs to show add or not add values. These are model dependent. The development of RCM always has some specific regions in mind and may be good for certain regions but not for others. We need to have a series of tests with a set of RCMs to test some hypotheses to reduce results' model dependence. This kind of tests would have scientific value.
(2). The discussion about ^good climatology consists of good meteorology^ implies a serious
RCM downscaling issue. I believe the statement is very good. But if we apply this statement to
the model, we will conclude none of GCMs will be qualify for climate study because none of current
GCM can simulate ^meteorology^ (as defined by Barry) well. My feeling is other way around more
likely to be true, i.e. the model capable to simulate meteorology well should produce good climate.
The RCM climate downscaling intends to have both meterology and climate be well simulated. If RCM
is not able to add value, then type 2, 3, and 4 downscaling will deem to fail, and statistic
downscaling is more meaningful. From some current RCM studies, I think one issue needs to be considered, i.e., whether the added value is due to RCM^Òs high resolution and/or better
topography. If this is the case, then RCM will have limited future since more and more high resolution GCMs will be in use.
Yongkang
Date: Fri, 05 Oct 2007 22:33:56 +0200
From: hvonstorch
To: Roger A. Pielke Sr.
Cc: Downscaling Group
Subject: Re: dynamic downscaling
Good evening, Roger, I would not really consider the exercise "construction of scenarios" (i.e. possible futures) as a subgrup of process studies. In my view, they are conditional predictions, with the conditioning factors given by ad-hoc assumptions.
The purpose of such simulations MAY be to enhance our understanding, but FOREMOST it is an exercise to span the space of possible futures, an effort of great practical importance, also known from other fields (e.g., Schwartz, P., 1991:The art of the long view. John Wiley & Sons, 272 pp).
Hans
Date: Fri, 5 Oct 2007 15:22:50 -0600 (MDT)
From: Roger A. Pielke Sr.
To: hvonstorch
Cc: Downscaling Group
Subject: Re: dynamic downscaling
Hi Hans, I do not see how they can even be viewed as conditonial predictions if all of the first order human climate forcings are not included (e.g. see the discussion of this in
National Research Council, 2005: Radiative forcing of climate change: Expanding the concept and addressing uncertainties. Committee on Radiative Forcing Effects on Climate Change, Climate Research Committee, Board on Atmospheric Sciences and Climate, Division on Earth and Life Studies, The National Academies Press, Washington, D.C., 208 pp. http://www.nap.edu/openbook/0309095069/html/
Such a requirement (to include all of the forcings) is a necessary condition for a skillful forecast.
Moreover, the global models when applied in a retopspective view, have not even been able to accurately predict such coarse climate change metrics as the linear trend of the tropical zonally averaged tropospheric temperatures (see the CCSP report Temperature Trends in the Lower Atmosphere: Steps for Understanding and Reconciling Differences. Unfortunately, as I discuss in my Public Comment on CCSP Report).
Even one of the lead IPCC authors has stated the limitations of the models as predictions; see
Lots to debate on this group! :-)
Have a good weekend! Best Regards, Roger
Date: Sat, 06 Oct 2007 12:01:59 +0200
From: hvonstorch
To: Roger A. Pielke Sr.
Cc: Downscaling Group
Subject: Re: dynamic downscaling
Dear Roger, there are two issues,
1) "Conditional forecast" - this is the task to determine the/a probable/possible future state if certain a-priori given conditions will prevail. These simulations are NOT made for improving understanding, as you associate with process studies - thus they are not process studies. Their purpose is to explore possible future paths of development (see reference to Schwarz, which I gave before) Conditional forecasts are prepared with similar tools as forecasts, with the difference that the former is processing a-priori assumed forcing conditions, while forecasts (Weather, ENSO) exploit initial conditions. In both cases the models are incomplete compared to reality - which is unavoidable as they are simplifications of reality (see our book Müller, P., and H. von Storch, 2004: Computer Modelling in Atmospheric and Oceanic Sciences - Building Knowledge. Springer Verlag Berlin - Heidelberg - New York, 304pp, ISN 1437-028X). The question is, in both cases - which dynamics must be included?, - but the answer in the two cases may be different.
2) Validity of the assumed conditions is a different issue. In some cases, an d in particular in teh IPCC scenario simulations, you may find the assumed conditions to be incomplete, internally inconsistent, impossible or highly improbable, or otherwise unsatisfying. In case of climate these assumed conditions are given in the form of socio-economic scenarios (A2, B2 ....), which some (e.g., Richard Tol) find very questionable. I personally consider myself not qualified to judge the SRES scenarios with respect to the required plausibility, internal consistence and possibility - and completeness.
For our present discussion, about methodical aspects and the added value/purpose of models, only issue #1 is of relevance. For issue #2 we have to extend the group of participants so to cover the required competence sufficiently deeply.
Regards, Hans
Date: Sat, 6 Oct 2007 10:37:37 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Downscaling Group
Subject: Re: dynamic downscaling
Hi Hans (sorry for the long e-mail on thwe weekend! :-))
The term "conditional forecast" could be used for any simulation! In any case we clearly disagree on this. Indeed, support for my view comes from unexpected places; e.g.
1. The first one is
"In fact there are no predictions by IPCC at all. And there never have been. The IPCC instead proffers "what if" projections of future climate that correspond to certain emissions scenarios....None of the models used by IPCC are initialized to the observed state and none of the climate states in the models correspond even remotely to the current observed climate.... the science is not done because we do not have reliable or regional predictions of climate. " [by Kevin Trenberth]
http://blogs.nature.com/climatefeedback/
as I discussed in
http://www.climatesci.org/2007/06/18/comment-on-the-nature-weblog-by-kevin-trenberth-entitled-predictions-of-climate/ [you should, of course, read the entire posting by Kevin Trenberth to
place it in the complete context of his weblog; however, his comments above are remarkably candid from an IPCC author].
2. The second quote is
"In the late 1960s and mid 1970s the chaotic nature of the climate system was first recognized. Lorenz (1969, 1975) defined two types of predictability problems:
1) Predictability of the first kind, which is essentially the prediction of the evolution of the atmosphere, or more generally the climate system, given some knowledge of its initial state. Predictability of the first kind is therefore primarily an initial value problem, and numerical weather prediction is a typical example of it.
2) Predictability of the second kind, in which the objective is to predict the evolution of the statistical properties of the climate system in response to changes in external forcings. Predictability of the second kind is thus essentially a boundary value problem..
...because of the long time scales involved in ocean, cryosphere and biosphere processes a first kind predictability component also arises. The slower components of the climate system (e.g. the ocean and biosphere) affect the statistics of climate variables (e.g. precipitation) and since they may feel the influence of their initial state at multi decadal time scales, it is possible that climate changes also depend on the initial state of the climate system (e.g. Collins, 2002; Pielke, 1998). For example, the evolution of the THC in response to GHG forcing can depend on the THC initial state, and this evolution will in general affect the full climate system. As a result, the climate change prediction problem has components of both first and second kind which are deeply intertwined.
This concept is illustrated in Figure 2, which shows two hypothetical future climate evolutions as simulated by a climate model. In each simulation the GHG concentration increases in the same way but starting from different times of the Control run, and thus different initial ocean, sea ice and land surface conditions. As illustrated, the two climate evolutions can potentially differ both in their mean and variability characteristics. The relevance of the first kind predictability aspect of climate change is that we do not know what the initial conditions of the climate system were at the beginning of the "industrialization experiment" and this adds an element of uncertainty to the climate prediction.....
.....To add difficulty to a prediction is the fact that the predictability of a system is strongly affected by non-linearities. A system that responds linearly to forcings is highly predictable, i.e. doubling of the forcing results in a doubling of the response. Non-linear behaviors are much less predictable and several factors increase the non-linearity of the climate system as a whole, thereby decreasing its predictability (e.g. Rial et al., 2004)."
[quotes above from F. Giorgi, 2005 : Climate Change Prediction: Climatic Change (2005) 73: 239.265 DOI: 10.1007/s10584-005-6857-4]
Thus, if the climate system is both a boundary value and an initial value problem (which I agree with), it is an initial value problem!
I discussed why climate prediction is an initial value problem in Pielke, R.A., 1998: Climate prediction as an initial value problem. Bull. Amer. Meteor. Soc., 79, 2743-2746
This is a topic we should debate further as it is central to the IPCC perspective, but is a view that places an obstacle in the recognition of how difficult climate prediction is.
In terms of dynamic downscaling, we do not need to resolve this disagreement. We do, however, need to agree on how Type 1, 2, 3 and 4 dynamic downscaling are tested scientifically.
For Type 1, this is done every day in weather forecasting, and we all agree on this approach.
For Type 2, my recommended track is to use coarse reanalysis for the lateral boundary conditions and nudging and validate with finer scale reanalyses on an event basis. We should also continue to do the statistical evaluations that have been completed by several of us.
For Type 3, we need to resolve, but I suggest we use seasonal forecasts where the SSTs are prescribed, as we have done in
Castro, C.L., R.A. Pielke Sr., J. Adegoke, S.D. Schubert, and P.J. Pegion, 2007: Investigation of the summer climate of the contiguous U.S. and Mexico using the Regional Atmospheric Modeling System (RAMS). Part II: Model climate variability. J. Climate, 20, 3866-3887.
For Type 4, suggestions are welcome, but I recommend that we use the IPCC model runs made for the last 20-30 years, and also for the next ten years.
The fundamental requirement is that we should agree on a set of quantifiable tests of regional downscaling using real world data. I have concluded we need to do this on an event basis for Type 2 and Type 3 downscaling, with Type 4 have predictive skill that is necessarily the least of all three types.
Finally, the reason that we should interpret "conditional forecasts" as process studies is that they are essentially to doing a partial derivative with a differential equation. This permits us to assess, for example, whether a particular climate forcing is important (e.g. the radiative effect of added atmospheric CO2 - which is shown to be important). The latter is, in my view, all that the IPCC has shown.
Roger
Date: Sat, 6 Oct 2007 19:09:16 +0200
From: Hans von Storch
To: Downscaling Group
Subject: AW: dynamic downscaling
Roger, I never made the claim that the scenarios are "predictions" or "forecasts"; indeed, I object when people, in particular from the UK, Hadley Center, use this misleading language. I also know the differences between first and second kind of predictions. In climate applications, scenario construction take the FORM of conditional predictions with GCMS or RCMs, and this can be done better or worse.
But, please, read what I write. My English may be litmited. We may look at theses thing from rather different contexts, which is a perfect chance for misunderstandings.
I am talking about an exercise with a certain purpose. You may find that people are not good in implementing the models for this purpose - but just look at the purpose, and agree with me that this purpsoe is not among the purposes you have listed. We can then begin to talk about how to implement the models better for THIS purpose.
Usually I describe scenarios by an example: When the city of Hamburg is planning for next Christmas, whether municipal employees may go on vacations or have to be stand-by to clear streets with heavy machinery. The decision depends on the weather in more than 2 months. Nobody can make a forecast, but we can make a forecast of the decision conditional upon the unknown weather. This weather is known in a certain probabilistic/plausibility range. Thus, we may deal with the scenario "snow" or "no snow" etc. Heat-wave would not be an admissable scenario. Thus, we have two or more possible futures. The city makes contingeny plans for these two or more (scenarios), and asks science - "when will science be able to give as robust statements about likely conditions on Christmas Eve?". -- This is standard approach, and you may question if the METHOD to describe plausible/probable scenarios is valid or needs improvement, ok - but that is not what you have argued against. The exercise to describe possibilities for Christmas Eve is not a process study but an "exploration of possible futures".
Regards, Hans
Date: Sat, 6 Oct 2007 13:49:16 -0700 (PDT)
From: Barry Lynn
To: Downscaling Group
Subject: Re: AW: dynamic downscaling
Hi: I will try to address issue (1) after I have read more of the papers suggested.
Regarding (2), Xue wrote:
"But if we apply this statement to the model, we will conclude none of GCMs will qualify for climate study because none of current GCM can simulate ^Ómeteorology^Ô (as defined by Barry) well."
I have to admit that I agree with this statement. Even so, my approach is that the GCMs can produce some general climate statistics (mean height field) that the RCM can modify to produce some kind of synoptic pattern (not always the same as the GCM). In this case, the RCM's domain should be much larger than the GCM, to give it enough "room" to produce more realistic meteorology.
Given the mismatch between the RCM and GCM, though, this implies that long term climate studies using RCMs (rather than timeslice studies) may not have much real value. This is for two reasons: i) when using a large domain the RCM solution diverges from that of the GCM so a mismatch is created in the meteorological and soil fields. This grows over time; ii) if we use a small RCM domain than the meteorology from the GCM creates unrealistic meteorology in the RCM, also rendering the RCM results suspect.
The end result is that I am firm believer in the prospects for global warming, but I don't think we really know any details about how this will pan out.
Barry
Date: Sat, 6 Oct 2007 13:36:01 -0700 (PDT)
From: Barry Lynn
To: "Niyogi, Dev" , pielke
Cc: Downscaling group
Subject: Re: dynamic downscaling: meteorology vs climatology
Dear Dev: I appreciate your comment and agree with you that we must know our user.
I have a few comments.
1) The paper I sent around details the possible sensitivity of RCMs to model configuration (really, choice of CU, radiation scheme, and boundary layer scheme).
2) The modeled climate and projected changes depend substantially on choice of model configuration.
3) The number of calculated frost days, heavy rain events, or heating degree days can be as or more sensitive to choice of model configuration as to changes in greenhouse gases.
So, my conclusion is that our current level of sophistication is not high enough to draw detailed (and regarding precipitation:meaningful) conclusions. (Our JCli paper on the potential for extreme temperature change used RCMs in timeslice mode with what we hoped was a more sophisticated CU parameterization to draw conclusions about the potential for extreme temperatures -- I wouldn't want to extend the conclusions of this work further than this to answer how many frost days, etc).
Barry
Date: Sat, 6 Oct 2007 15:53:10 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Hans von Storch
Cc: Downscaling Group
Subject: Re: AW: dynamic downscaling
Hi Hans, Thank you for the clarification.
After reading your latest e-mail, I feel our difference is now mainly semantic. The term "conditional forecast", however, is misleading, as it is equivalent to the phrase "conditional prediction". I read conditional forecast as meaning that this is the forecast if the excluded forcings and feedbacks are relatively unimportant. I would suggest "conditional scenario" although in my article (see Figure 6 for my suggested terminology)
Pielke Sr., R.A., 2002: Overlooked issues in the U.S. National Climate and IPCC assessments. Climatic Change, 52, 1-11.
I reserve the term scenario for the simulations if all of the important direct and feedback effects on climate are included. A scenario is then a realization out of an ensemble of possible simulations. This is, of course, just my proposed terminology but if we do not use this, we need to all agree on what to use to includes each situation.
In terms of how the perspective of how risk is assessed, we have utilized the concept of vulnerability, which has significant advantages over the scenario (and "conditional forecast") approaches as we summarize in Table E.7 in
Pielke, R.A. Sr., and L. Bravo de Guenni, 2004: Conclusions. Chapter E.7 In: Vegetation, Water, Humans and the Climate: A New Perspective on an Interactive System. Global Change - The IGBP Series, P. Kabat et al., Eds., Springer, 537-538.
with more details in
Pielke, R.A. Sr., and T.J. Stohlgren, 2004: Contrast between predictive and vulnerability approaches. Chapter E.3 In: Vegetation, Water, Humans and the Climate: A New Perspective on an Interactive System. Global Change - The IGBP Series, P. Kabat et al., Eds., Springer, 491-495.
The vulnerability perspective is more inclusive than with the conditional forecasts.
The question for your illustrative example in the vulnerability framework, is what are the thresholds for activities that are affected by snow and no snow conditions, and what can be done to reduce the risk of negative consequences if either of these conditions occur? The starting point is the impact (i.e. the endpoint). We appear to agree on this, but please confirm.
In terms of the IPCC process, this would mean the starting point would be the threats to society and the environment from the spectrum of human and natural effects, and not by starting from the GCMs and downscaling. With the use of conditional forecasts (scenarios) the problem is that they may not capture all of the potential risks.
Thanks for engaging in this disucssion this weekend!
Best Regards, Roger
Date: Sun, 7 Oct 2007 03:06:14 -0700 (PDT)
From: Barry Lynn
To: Yongkang Xue , Roger A. Pielke Sr.
Hans von Storch
Cc: downscaling
Dear Xue, Roger, Hans, Dev: I am reading Xue's paper on the possible impacts of the position of the boundary on the modeled precipitation. His quest, so to speak, is to find the best location for the boundary to provide the RCM with the most realistic input.
Our paper (already attached) shows how given the same boundary conditions, the same model with different configurations can give quite different downscaling climates (and climate change).
Roger's experiments seem designed to investigate the best way to downscale (i.e, incorporate boundary conditions into the RCM domain). I think we agree that if the GCM provides unrealistic meteorology, then we pour results our suspect -- even if we had the "perfect" way to incorporate the information. Since the GCMs have never been designed to produce realistic meteorology, we may be hindered in our quest to realistically simulate regional/local climate/change.
From my end, our research (among others) shows that choice of model physics will also impact the "answer" the model gives in both the present day climate, and in response to changes in greenhouse gases, for example.
I guess my point is that realistic downscaling is not only about obtaining perfect boundary conditions, but also simulating the meteorology using model physics that realistically captures precipitation/radiation feedbacks (something Roger has tried to stress in his e-mails about process studies), for example, It also means finding the correct horizontal resolution to simulate such processes.
In summary, if we are really going to make progress in downscaling, then we need to attack both sides of the problem (and if GCMs ever are run at "high-enough" resolution, their results will still be suspect until we can be sure they produce realistic meteorology (e.g., frequency of precipitation, time of maximum precipitation, precip amounts, profile of cloud induced heating, etc).
So, where do we begin? Barry
Date: Sun, 07 Oct 2007 08:24:05 -0700
From: Yongkang Xue
To: Barry Lynn , Roger A. Pielke Sr., Hans von Storch
Cc: downscaling
Subject: Re: Downscaling: Perfect Model Vs Perfect Boundary Conditions
Dear Roger, Hans, and Barry, Roger and Hans's discussions are very enlightening. I agree with Roger's final statement that the differences between Roger and Hans arguments are mostly semantic. We all agree IPCC experiments have serious shortcoming and it is hard to view as a real future prediction (or projection as someone call to make other scientists feel a little bit better and/or make themselves have less responsibility in case climate goes differently). But nevertheless, the report based on the IPCC experiment creates a shock way across the globe and make climate prediction one of the central issues in today's world.
As to the downscaling, I think there are a few issues which need to be think about:
(1) For the first type, we all agree that this type downscaling adds values. However, there are still issues: (a). In this type of experiment, initial conditions are dominant. What about LBCs? Does GCM outputs could be used for LBC? As a matter of fact, SST is not changed in this type of experiment. (b). Although RCM as a group has the ability for this type of downscaling. Does every RCM have the ability to downscale properly for every region in the world? My feeling is not. As Barry pointed out, physical processes play a role. So the dominant physical processes in a specific region, such as convective processes, moisture transport processes, and/or land processes, have to be properly represented in a specific RCM for this region. (c). The successful downscaling is due to the memory of initial conditions. How long this dominant effect will last? It seems we have no question for initializing every 24 hours then looking at long term mean for climate (as did in Xue et al., 2001, Mon. Wea. Rev.). Zong-liang has shown in Lo et al's study, they restart every 7(?) days. I am not that clear what the conclusion from that experiment. It leads to the second type downscaling because the second type is an extension of first type in terms of model experiment design with less initial condition effect but more and more LBC effect.
(2). For type 2, we certainly need to explore the nudge effect. However, if nudge is the only way to make RCM downscaling right, I am a little bit pessimistic about the dynamic downscaling since statistical downscaling plus nudge may be a more efficient way for downscaling. As Barry's, ours, and others have shown many factors may affect the dynamic downscaling, I think we also need to set some ideal experiments to test them (including nudge) with more RCMs since previous results were all model dependent.
(3). For type 3, we may use GCM forecast (say from ECMWF or NCEP) for LBC.
(4).For type 4, NOAA is preparing a big experiment on this type with NCEP CFS. We may need to discuss with Jin on this.
Yongkang
Date: Sun, 07 Oct 2007 19:24:22 +0200
From: hvonstorch
To: Hans von Storch , Barry Lynn, Roger A. Pielke Sr. , Yongkang Xue
Cc: Downscaling
Subject: Re: Downscaling: Perfect Model Vs Perfect Boundary Conditions
Barry, may I challenge you?
why do you think your list of variable is so important - "In summary, if we are really going to make progress in downscaling, then we need to attack both sides of the problem (and if GCMs ever are run at "high-enough" resolution, their results will still be suspect until we can be sure they produce realistic meteorology (e.g., frequency of precipitation, time of maximum precipitation, precip amounts, profile of cloud induced heating, etc)." Why don't you list other properties such as the turbulence spectrum in the boundary layer, the wave growth on the ocean surface as interacting feature of the ocean surface and the atmopsheric boundary layer? (Just examples, of course.)
You know that all models will always disregard some processes; some of which may be parameterized, others neglected (which is a special of "parameterization"). How do you select?
Parameterizations specify the effect on resolved scales of non-resolved processes conditional upon the large-scale state (as conditional mean (averaged across many similar large-scale states) or even randomized). Do you think that is possible, given your statement below?
Does it precipitate" in LAMs? I mean, are drops formed, interacting with each other? How many? Which size spectrum? It does not rain in LAMs - water is moved from the atmosphere to the surface; we do some calculations that determine the amount of water which must have fallen down. We describe statistics of precipitation, but not precipitation it self - thus do we include "real meteorology"? We may call this information in in our "slang" prpecipitation, but it is not, is is statistics of precipitation.
Hans
Date: Sun, 7 Oct 2007 12:33:50 -0700 (PDT)
From: Barry Lynn
To: hvonstorch, Roger A. Pielke Sr., Yongkang Xue,
Cc: downscaling
Subject: Re: Downscaling: Perfect Model Vs Perfect Boundary Conditions
Dear Hans: Thank you for helping me to clarify my thinking. I answer you below (---->).
hvonstorch: Barry, may I challenge you?
why do you think your list of variable is so important -
"In summary, if we are really going to make progress in downscaling, then we need to attack both sides of the problem (and if GCMs ever are run at "high-enough" resolution, their results will still be suspect
until we can be sure they produce realistic meteorology (e.g., frequency of precipitation, time of maximum precipitation, precip amounts, profile of cloud induced heating, etc)."
Why don't you list other properties such as the turbulence spectrum in the boundary layer, the wave growth on the ocean surface as interacting feature of the ocean surface and the atmopsheric boundary layer? (Just examples, of course.)
----> I know that the variables that I have mentioned are crucial to simulating radiation/precipitation
feedbacks, and hence surface temperatures. I cannot comment if the variables you mention are
of the same "level" of importance.
You know that all models will always disregard some processes; some of which may be parameterized, others neglected (which is a special of "parameterization"). How do you select?
-----> I only selected/listed those that I am familiar with. There may be others, like parameterized land
surface processes, that may be of similar importance over time, not to mention melting of ice!
Parameterizations specify the effect on resolved scales of non-resolved processes conditional upon the
large-scale state (as conditional mean (averaged across many similar large-scale states) or even
randomized). Do you think that is possible, given your statement below?
------> We need to develop our best parameterizations. If our parameterizations do not reproduce reality, then we may have introduced large errors into the system. Quantifying those errors will help us better understand the potential uncertainty in the model error.
-----> I am not sure that cumulus parameterization can adequately reproduce explicit moisture/ice
processes. Unfortunately, even models with bulk microphysics introduce fairly large errors into the
hydrometeor field.
-----> I don't know if it is possible to develop realistic parameterizations, but we should try, so as to
reduce potential sources of model error.
Does it precipitate" in LAMs? I mean, are drops formed, interacting with each other? How many? Which size spectrum? It does not rain in LAMs - water is moved from the atmosphere to the surface; we do some calculations that determine the amount of water which must have fallen down. We describe statistics of precipitation, but not precipitation it self - thus do we include "real meteorology"? We may call this information in in our "slang" precipitation, but it is not, is is statistics of precipitation.
----> You are correct: it is the statistics of precipitation. The question is: how accurate are these
statistics? Up to now, GCM modelers have mostly cared about getting the precipitation amounts "correct," while ignoring the frequency and timing of precipitation (the frequency of precipitation and timing of precipitation determine how many days and how much of the day of sunlight reach the ground, or how much longwave radiation escapes to space). They also tune the radiation function so that they get pretty good surface shortwave radiation for their given precipitation scheme. Is this tuning at all malleable to new circumstances?
----> I think it less important how you parameterize processes than whether these parameterizations
actually "work," and can respond
correctly to changes in the initial conditions or forcing fields.
---->> I think GCMs and regional model studies are useful and have given us ample warning about the
potential for future warming. I just think the uncertainty is large. Given this uncertainty, though, we
should probably be more cautious than less.
Barry
Date: Sun, 28 Oct 2007 22:19:51 -0400
From: Fedor.Mesinger
To: Downcaling Group
Subject: Downscaling: Value added at large scales
Hi all, feel guilty being privileged to get all of these most interesting e-mails, and not contributing! But it is still October and not much beyond Liang’s target date of October 16; thus – for those who still have some energy left - here are a few words, addressing
Value added at large scales
Are we facing the very basic problem of the impossibility of proving that something is NOT possible? Before the space age, could one prove that there were no elephants on the other side of the Moon? What I am trying to do below is much easier, to demonstrate that value added at large scales IS possible. Here I go:
Please have a look at plots on top of the back cover (page 16) of Mitchell, K., M. J. Fennessy, E. Rogers, J. Shukla, T. Black, J. Kinter, F. Mesinger, Z. Janjic, and E. Altshuler, 2001: Simulation of North American summertime climate with the NCEP Eta Model nested in the COLA GCM. GEWEX News, 11, No. 1, 3-6.
Is there anyone who does not agree that the Eta/RCM has improved on the driver model large scales, both in winter and in summer?
A still more impressive result for the precip difference 1993-1988 was obtained by Fennessy and Altshuler later, running 9 ensemble members for each of 1993 and 1988, shown at the December 2002 AGU Assembly. Will mail the plot separately, following this e-mail. The GCM, in a predictive mode (Type 3 of Castro et al.), produced a minimum of precip difference roughly where the maximum ought to be. In spite of using the GCMs LBCs, the Eta/RCM placed the maximum quite well, and its intensity was not bad either.
An MM5 Type 3 result showing an RCM improvement compared to the driver GCM is shown in Gustafson and Leung, BAMS, August this year. Using the NARR, upper right panel of Fig. 4 (page 1222) as truth, the MM5 result shown in Fig. 4, bottom right, is a visible improvement compared to the driver GCM result, Fig. 5 right panel. The discussion of this is on page 1224, lower right column.
There is also a ton of Type 1 operational Eta model results suggesting that its value added compared to its driver global NWP model increases somewhat with time out to 3.5 days, as the initial condition is being flushed out of its domain, and LBCs of the global NWP model run of 6-hr prior to the Eta initial time are advected in. There is strong indication that this is due primarily to largest scales {my various conference papers, e.g. Mesinger, F., 2004: Dynamical core design: A neglected thrust toward increasing NWP skill several days ahead. The First THORPEX International Science Symposium, Montreal, Canada, 6-10 December 2004.
It is hard to discount this as due to the Eta not yet having forgotten its initial condition, because 1) if the initial condition is the source of the Eta value added, the advantage should be greatest in the beginning, and 2) of the strong indication that the operational Eta initial condition was considerably inferior to that of its driver global model, same reference.
So how come we are in this state of disagreement? First, I believe that the Type 2 experiments are not very useful in assessing the RCMs performance at large scales. To see why, consider a thought experiment in which we have a perfect RCM, performing as the real atmosphere. We are now driving this perfect RCM with the reanalysis LBCs which although sampled from the real atmosphere are sampled with an error; compared to the real atmosphere they are only approximate. While in a run long enough our perfect RCM will forget its initial condition, it will not forget the approximate LBCs that are constantly fed to it. To quote Ed Lorenz (acknowledgment: posting on Eugenia Kalnay office’s door at UMd) chaos: when the initial condition determines the future but approximate initial condition does not approximately determine the future. Approximate LBCs will inflict the same behavior on our perfect RCM, the more so the bigger its domain is. Thus, our perfect RCM will not pass the test of emulating the reanalysis “truth”, which does approximately describe the “future” of the real atmosphere.
Second, a considerable contribution (perhaps all of it?) to the idea that RCMs might be unable to improve on large scales comes from the spectral analysis, reanalysis vs the RCM. I feel the results of such an analysis may or are even likely to be deceiving. Suppose an RCM produces a jet stream which is somewhat more accurate than that of the reanalysis, being in a slightly better position here and there. Although the jet stream is the largest scale feature there is, the difference of the two will show as the difference in smaller scales.
But of course an RCM should properly reproduce statistical features of the driver fields giving the LBCs, e.g., the level of kinetic energy (Castro et al.). If it does not, rather than suggesting an inability of RCMs to do that, shouldn’t we better try to figure out why? (I am of course painfully aware of the
funding problem with this attitude!) RCM being an approximation to the real atmosphere, it ought to be able to do that; if it is not, there has to be a reason or reasons. “... ineffectiveness of the nesting for controlling the large scales over the whole domain” (quote from a published paper) I find a strange statement to find in a scientific paper. One should rather have said “... the nesting we have used ...”, and also note that the goal of producing large scales the same as those of the driver solution is appropriate only if applied to statistical features, unless the RCM domain is rather small.
My candidate reasons for failure of an internally well-designed RCM to reproduce statistical features of the driver fields (e.g., the level of kinetic energy) are in smoothing that takes place in boundary interpolations, and in the relaxation LBC scheme that most RCMs use. Note that the relaxation scheme is in conflict with the mathematics of the limited area problem we have. First, we are solving an initial-boundary value problem, and not an initial-boundary band values problem. Second, in view of the characteristics of the differential equations we are solving, one less condition ought to prescribed at the outflow than at the inflow parts of the boundary. This was pointed out already by Jule Charney in his 1962 (!) Tokyo symposium paper.
It is in Type 3 and Type 4 experiments that tests free of the above discussed Type 2 difficulties can be done. An RCM driven by forecast GCM fields then has a chance to produce large scales with verifiable value added compared to those of the driver fields. Hopefully the plots I refer to are convincing enough of this actually having been done! To demonstrate that something IS possible already a single (believable!) experiment suffices, a task much easier than that that of demonstrating that something is not possible. Of course, some territory for argument remains, such as that value added is possible only due to a major weakness in some sense of the driver model. But even so, the demonstration of RCMs creating value added in large scales has been made, provided you accept that plots can serve that purpose.
Regret being so long; hope you do not mind. Can I use the weekend for an excuse, following Roger’s example?
Looking forward to further discussions, or the intended draft, best, Fedor
Date: Sun, 28 Oct 2007 22:21:57 -0400
From: Fedor.Mesinger
To: Downscaling group
Subject: Downscaling: Value added ..., the plot
the plot. f.
Date: Sun, 28 Oct 2007 23:38:55 -0500
From: liang
To: Fedor.Mesinger
Cc: Downscaling Group
Subject: Re: Downscaling: Value added at large scales
Hi Fedor, Thank you for your long but very informative email. You not only provide a few good examples of value added downscaling simulations, but also remind me of the targeted overdue summary. Sorry that I am still working on it -- I have been too busy with several theme hire interviews going on at our Jackson School of Geosciences; see http://www.jsg.utexas.edu/hiring/climate.html. I view your comments as a nice nudging towards the synoptic forcing fields :-)
Cheers, Liang
Date: Mon, 29 Oct 2007 03:58:24 -0700 (PDT)
From: Barry Lynn
To: Fedor.Mesinger
Cc: Downscaling Group
Subject: Re: Downscaling: Value added ..., the plot
Dear Fedor: Thank you for sending me the picture of your RCM modeling results.
It certainly appears that the RCM has added value to the GCM simulation.
I have a few questions:
1) What is the resolution of the GCM?
2) Why do think the RCM succeeded where the GCM failed? Is the RCM producing better synoptic-scale/mesoscale meteorology than the GCM (at least in terms of changes in the height field, frequency of storms?).
3) Really, I want to know if the RCM produced better meteorology as well as a better overall seasonal signal.
Last, I agree generally with the letter that followed: in fact, I will make the statement stronger. If an RCM produces more realistic meteorology (as well as seasonal precipitation/temperatures) than the GCM, then most likely it is not reproducing the longwave/synoptic pattern simulated by the GCM, but something closer (in the mean/std) to observed conditions.
Barry
Date: Mon, 29 Oct 2007 07:13:19 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Fedor.Mesinger
Cc: Downscaling Group
Subject: Re: Downscaling: Value added at large scales
Hi Fedor, Thank you for your significant contribution on this subject. It provides a very effective focus for our discussions.
In terms of value-added from RCMs (Type 2 and 3), I agree that they can improve model accuracy compared with observations when the more detailed surface forcing forces mesoscale/regional weather patterns. We certainly see this, for example, when we used a reanalysis as lateral boundary conditions to force a RAMS over Florida. The reanalysis does not resolve the sea breeze. This is why your results (that you provide documentation for) improve on the larger scale model, since the surface forcing is important [incidently, Tomislava Vukicevic has proposed using her adjoint work to quantify what fraction of the model results come from surface forcing and what fraction comes from lateral boundary conditions; we just need to get this funded..:-)).
However, this is not the appropriate test in my view. We need to perform the downscaling over ocean areas where there is minimal surface forcing on a spatial scale smaller than the larger scale model. If you can produce more realistic weather features on the regional domain than with the parent reanalysis (Type 2 run) and GCM (Type 3 run), this would be convincing.
With respect to the use of the reanalysis, the power of this approach is that we have observed analysis data to compare with the RCM results. For instance, using the NARR to evaluate model performance when the lateral boundary conditions are from the NNR provides such a test.
I do agree that Type 3 runs (with prescribed SSTs) such as performed with the COLA/ETA set are very effective. However, ETA can add value (accuracy) due to its finer spatial resolution of the surface forcing over land, in the example that you provi8ded. As I mentioned above, we need to evaluate this value-added over a region where surface forcing is not going to add information beyond that in the parent model.
Finally, while I agree with your thought experiment on the Type 2 experiments (i.e. that approximate lateral boundary conditions from the reanalysis, sampled from the real world, will create deviations in the interior of the "perfect" RCM), this deviation is what we assessed with our papers.
In the Type 3 (and even more so in a Type 4 run), the lateral boundary conditions have no information at all on smaller scales than the parent model can resolve. Thus, they must provide poorer (i.e., less accurate) lateral boundary conditions than when reanalyses are used to force the RCM. Even with a "perfect" RCM, how can the RCM correct for erroneous information that is being fed into the RCM? We concluded, therefore, that the Type 2 downscaling provides an upper limit on the accuracy of the RCM.
I hope these comments clarify and that we continue to discuss ways to develop a consensus on this very important subject.
With My Best Regards, Roger
Date: Tue, 30 Oct 2007 12:50:16 -0700 (PDT)
From: Barry Lynn
To: liang, Fedor.Mesinger
Cc: Downscaling Group
----------------------------------------
A question/thought experiment for all those that believe that downscaling cannot add any greater value than the information present in the boundary conditions.
Let us say that the lateral boundary information is at 36 km grid resolution (the same as our outer grid). The model's finest grid is at 4 km grid resolution. The 36 km grid uses a cumulus parameterization to simulate moist convection. The 4 km grid uses an explicit scheme that produces realistic looking precipitating clouds. Has not the 4 km grid added value to the forecast?
Ok, I understand that if the highest resolution grid is large enough and the simulation time long enough, then the initial information is lost, and the timing of simulated precipitating events might deviate from the information implied by the boundary conditions.
1a) But, isn't it possible that on the mean, the 4 km grid will produce more realistic meteorology on average than the 36 km grid (would have by itself)?
1b) Could this type of idea explain the better results obtained using the COLA RCM than with the COLA GCM by itself?
2) In fact, couldn't the higher resolution grid produce more realistic meteorology than the coarse resolution grid if it was large enough to correct the errors from the coarse boundary conditions?
If the answer to my questions is yes, then the importance of simulating the meteorology over a wide area better outweighs any "errors" (or deviations) induced in the downscaling of the coarse information.
Barry
Date: Tue, 30 Oct 2007 16:50:32 -0600 (MDT)
From: Roger A. Pielke Sr
To: Barry Lynn
Cc: Downscaling Group
Subject: Re: Downscaling: Value added at large scales
Hi Barry, A fundamental question is how can the higher spatial resolution and more detailed parameterizations "correct" errors in the lateral boundary conditions if the internal regional model dynamics and physics significantly depend on the lateral boundary conditions? I propose that we need to look at this issue over the ocean with no SST gradients.
I agree that when surface forcing is better resolved, that the regional model can still provide added skill, providing the errors in the lateral boundary conditions do not exert too large of an effect on the internal model results (thus directly affecting how the effect on the atmospheric dynamics and physics of the surface forcing is simulated). The skill added by the COLA/RCM appears to be due to the more spatially resolved surface forcing. We need to test this to see which of our views is correct.
Roger
Date: Tue, 30 Oct 2007 22:35:00 -0400
From: Fedor.Mesinger
To: Roger A. Pielke Sr.
Cc: Downscaling Group
Subject: Re: Downscaling: Value added at large scales
Hi all, Appreciate the comments and questions sent! Here are some of my responses, in the order things came to me.
Barry, questions:
1) The resolution of the GCM is R40/18 L. More information on the two models can be found in the reference I gave, freely downloadable.
2) Why do I think the RCM succeeded where the GCM failed? Ruby Leung gave a talk at the NCAR regional climate modeling workshop two and a half years ago. She and her postdoc were running the same 1993-1988 case, using the NCAR WRF as an RCM, and had results that were less than satisfactory. As I recall she interpreted the requirements for a successful seasonal forecast of the 1993 flood as the correct jet stream over the Rockies in combination with an also correct low-level jet from the Gulf bringing in moisture for all these huge rains.
If this is so, consequently both of these must have been done better by the Eta RCM than either the COLA GCM, or the NCAR WRF.
If this is so, consequently both of these must have been done better by the Eta RCM than either the COLA GCM, or the NCAR WRF.
I expect Roger will be uncomfortable with this kind of talk: I am talking about models, as opposed to GCM vs RCM. If we want to be certain that the value added comes exclusively from the higher resolution of the RCM, even experiments with an RCM over water will not do. We^Òd have to do experiments with an RCM that is a limited area version of the global model used to drive the RCM. This can be done - pairs of such models are available - I am not aware if this has in fact been done by someone already.
3) Better meteorology or a better overall seasonal signal? The first I suppose. SSTs were the same.
I agree with the ^Óstronger statement:-)
Roger,
Date: Wed, 31 Oct 2007 08:06:49 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Fedor.Mesinger
Cc: Downscaling Scales
Subject: Re: Downscaling: Value added at large scales
Hi Fedor, On the reply to Barry, I agree that IF the precursors of a jet stream over the Rockies and the low-level jet from the Gulf can be realistically simulated, the RCM could improve on this. However this is the crux of the issue. The coarse model must be able to skillfully insert the winds, temperature and moisture through the lateral boundary conditions as a necessary condition for the skillful RCM run with these synoptic features. Without this, how can the ETA RCM skillfully produce (and improve on) these synoptic features if the input of larger scale synoptic information is flawed? Clearly, it can only add value (skill) if the surface forcing is a dominate effect despite poor synoptic features.
On the tests over water, our studies suggest that ALL of the RCM skill comes from more detailed terrain and landscape information, as you mostly agree with in your e-mail. This is why I suggest a test over open water as we can isolate the value-added (if any) with respect the ability of the RCM to improve on the skill of a coarse GCM (or Reanalysis). This would show whether or not the RCM can add skill at both the GCM (or Reanalysis) resolved scale and the finer scale. We need to run Type 2 and 3 experiments to test this. [I agree Type 2 experiments can answer only part of this question, but it can provide an upper limit on RCM skill, as recommended in my earleir e-mail).
The paper among us to focus these questions will make a very effective research proposal. Thank you for your, as always, insightful comments!
With My Best Regards, Roger
Date: Wed, 31 Oct 2007 08:09:00 -0700 (PDT)
From: Barry Lynn
To: Roger A. Pielke Sr. , Fedor.Mesinger
Cc: Downscaling Group
Subject: Re: Downscaling: Value added at large scales
----------------------------------------
Dear Roger: You wrote: "The coarse model must be able to skillfully insert the winds, temperature and moisture through the lateral boundary conditions as a necessary condition for the skillful RCM run with these synoptic features."
I agree with this statement. However, it is possible that the domain of the RCM domain is large enough to reproduce these features (even if they are not reproduced in the coarse (GCM) domain), and also of high enough resolution to reproduce both synoptic and mesoscale meteorology required to accurately simulate precipitation.
Alternatively, if a high resolution domain does not receive realistic boundary conditions and is not large enough to reproduce them, then the larger scales will not realistically force the mesoscale (and turbulent scale).
Barry
Date: Wed, 31 Oct 2007 10:06:38 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Barry Lynn
Cc: Downscaling Group
Subject: Re: Downscaling: Value added at large scales
Hi Barry, How can the RCM generate more accurate propagating synoptic features in its domain when it is made larger? This assumes that the lateral boundary conditions do not matter (i.e. become progressively less important). If this is true, the RCM could be run with any information on its boundaries. This is easy to test.
When the RCM "remembers" its initial conditions (i.e. a Type 1 simulation), I agree the RCM model can do add very significant skill. However, after this initial condition information is lost, except for surface forcings, what information does even a "perfect" RCM have available to accurately create these propagating synoptic features? How would a larger domain help? And, if the atmospheric response to surface forcing depends significantly on these propagating features, this forcing will not add much value either. From Fedor's summary, of course, it is clear there are cases where surface forcing dominates. We need to better define when this is the case (and when it is not). The open ocean simulations is one way to assess this question.
The reason that we need to work with Type 2 downscaling is to have observed data (NNR and NAAR) to evaluate quantitatively this issue.
Also, with respect to lateral boundary conditions, they are in effect "regional scale forcings". If the GCM is unable to accurately simulate features on the regional scale, the lateral boundary conditions are doomed to introduce significant errors. A larger RCM domain cannot remedy this problem, as there are always lateral boundaries with information being fed in, and the mechanism in which the RCM can use this added degree of freedom to correct the larger scale model errors is not clear. It just introduces more of its own "model world".
Roger
Date: Wed, 31 Oct 2007 15:40:46 -0400
From: Fedor.Mesinger
To: Roger A. Pielke Sr.
Cc: Downscaling Group
Subject: Downscaling: larger domain
Hi Roger and Barry, all,
Here is my suggestion how a larger domain helps. We seem to agree that surface forcing by topography was a crucial element enabling the Eta RCM to do well the 1993-1988 Type 3 precip case. The impact of the better representation of the mountain barrier is certainly felt downstream, but it is felt upstream as well. Recall, e.g., various theories and/or arguments what happens as a trough is moving up and down a mountain barrier. With a more accurate jet stream upstream of the barrier, even slightly, it should be better downstream as well. Thus, the more space the RCM has upstream of its major topographic barrier, the better it will do.
If you accept this, then with the same quality of the LBCs, the LBCs will indeed
become ^Óprogressively less important^Ô as the domain gets larger. ^ÓImportant^Ô may be a bit strong word here, certainly LBCs always are important and do matter; it is that the damage inflicted by the LBC errors becomes progressively somewhat smaller.
There was an inadvertent experiment run by NCEP for more than two years ~1996-1997 as follows. An operational 48 km/38 layer Eta was run on a large domain (covering considerable parts of both oceans, including e.g., Hawaii), using 12-h old global model (referred to as ^ÓAvn^Ô at the time) LBCs. The reason for using the 12-old Avn run LBCs was that the Eta was run first, before the global model run. Once the ^Ócurrent^Ô Avn run finished, a higher resolution, 29 km/50 layer ^ÓMeso Eta^Ô would run, using current Avn LBCs; but it was run on an about 2.5 times smaller domain ^Ö still however including fair-sized chunks of both oceans. Contrary to a general expectation, the higher resolution Meso Eta failed to show precip skill better than the large domain operational Eta. More details on this experiment are published in General Circulation Model Development, Ed. D. Randall, Academic Press, 2000, pp. 403-406.
I see no credible hypothesis what else could have helped the 48-km Eta compensate for clearly less accurate LBCs than its large domain. This is a Type 1 experiment, but certainly the initial condition should not have been of any help to the 48-km Eta, if anything the initial condition of the Meso-Eta should have been better too. This said, I do agree that the presence of the initial condition - and also of different resolutions - messes up the argument, and that repeating the experiment in a Type 3 framework would be great.
Best, Fedor
Date: Thu, 1 Nov 2007 07:50:51 -0700 (PDT)
From: Barry Lynn
To: Roger A. Pielke Sr.
Cc: Downscaling Group
----------------------------------------
Dear Roger: To answer your questions, schematically:
1) The larger the RCM domain, the greater the likelihood it will be able to create realistic meteorological features (just imagine an RCM at high resolution extended over more than one continent, etc.)
2) The higher the resolution of the RCM, the more realistic these features should be.
3) A GCM may send poor data into the boundaries, but if the RCM domain is large enough, it may still be able to develop more realistic synoptic meteorology in the interior of the domain -- since it is simulating higher resolution meteorology that can organize itself into larger scales.
4) When making a seasonal simulation using climate model data, a larger RCM domain allows the RCM to spin up more realistic meteorology (see Feder's letter how underlying topography can help here).
5) An RCM may never be able to faithfully reproduce reanalysis data, but over the long term it can produce more realistic meteorology than a GCM, producing a better regional product.
Barry
Date: Thu, 1 Nov 2007 09:41:42 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Barry Lynn
Cc: Downscaling Group
Subject: Re: Downscaling: Value added at large scales
Hi Barry, Please see my replies below.
Best Regards, Roger
-------------------------------
To answer your questions, schematically:
>
> 1) The larger the RCM domain, the greater the likelihood it will be
> able to create realistic meteorological features (just imagine an RCM at
> high resolution extended over more than one continent, etc.)
If the propagation and advection of precursors to meteorological features is not accurately handled at the lateral boundaries, how can the RCM create these features? In the limit of global coverage (perhaps even northern hemispheric coverage) I agree with you, but not for the domains typically used. In any case, we need to develop a way to test this.
>
> 2) The higher the resolution of the RCM, the more realistic these
> features should be.
Agreed; the meteorology is better represented.
>
> 3) A GCM may send poor data into the boundaries, but if the RCM domain
> is large enough, it may still be able to develop more realistic synoptic
> meteorology in the interior of the domain -- since it is simulating
> higher resolution meteorology that can organize itself into larger
> scales.
>
How can the RCM accommodate the correct this poor data? It can create more realistic features in the interior, but unless the lateral boundary condition has almost no effect, then the interior results are contaminated. One test of this would be to run the RCM with altered lateral boundary conditions to see how robust the interior results are to this perturbation.
> 4) When making a seasonal simulation using climate model data, a larger
> RCM domain allows the RCM to spin up more realistic meteorology (see
> Feder's letter how underlying topography can help here).
>
If the meteorology of individual weather features cannot be faithfully replicated, the RCM is generating weather in its "weather world", and not the real world. This is why running with the NNR as lateral BCs and then validating with the NARR is so valuable.
> 5) An RCM may never be able to faithfully reproduce reanalysis data,
> but over the long term it can produce more realistic meteorology than a
> GCM, producing a better regional product.
>
A necessary condition for accurate RCM runs is that they faithfully replicate the reanalyses where there is extensive observed data used in the construction of the reanalysis.
Thanks for continuing to iterate on this!
Date: Thu, 1 Nov 2007 18:27:51 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Downscaling Group
Subject: Re: Downscaling: Value added at large scales
Hi Barry,
I have a further suggestion to test your assumption about the value of expanding the RCM domain. What is the time/distance for vorticity maxima and minima (in terms of both absolute vorticity and potential vorticity) to loss their identity when they enter at the lateral boundary? The NARR (and NNR) could be used to address this question.
Having looked at weather maps for many years, I expect them to be coherent for quite a few days, as the synoptic scale vorticity maxima and minima propagate through the larger scale ridge/trof pattern.
For your hypothesis to be correct, these features need to be shown to be unimportant within the RCM domain.
Roger
Date: Fri, 2 Nov 2007 00:52:00 -0700 (PDT)
From: Barry Lynn
To: Roger A. Pielke Sr.
Cc: Downscaling Group
----------------------------------------
Dear Roger: Thank you for your suggestion.
I need to modify your statement, a bit. " these features need to be shown to be unimportant within the RCM domain." to ---> these features need to be modified by the RCM to produce more realistic structures.
I guess my point is that some of the input data will be lost, while some of it will be improved by the RCM, and some of it will be coherent enough to perhaps negatively impact the result.
Maybe we should pool all of our resources and write a gigantic proposal to study downscaling issues. Program managers like big, and it only takes one to fund it (regardless of the particulars of the reviews).
Barry
Date: Fri, 2 Nov 2007 06:29:46 -0600 (MDT)
From: Roger A. Pielke Sr.
To: Barry Lynn
Cc: Downscaling Group
Subject: Re: Downscaling: Value added at large scales/funding?
Hi Barry, I agree; a large collective proposal is an excellent idea.
On the one point you describe below;
"...need to be modified by the RCM to produce more realistic structures"
this is what we originally thought when we used the NNR to drive the RCM version of RAMS. However, when the agreement with the reanalysis (the "truth") deteriorated as the domain became larger, it alerted us to a fundamental problem with dynamic downscaling. The other papers since than [Rockel et al 2007; Lo et al 2007; Castro et al 2007] support our view on this limitation.
The combined use of NNR and NARR (for the type 2 downscaling) permits us an effective way to test this issue.
Who should take the lead on putting together the while paper for this proposal? (I suggest Liang as he agreed to take the lead on our BAMS paper as I recall).
Roger
Date: Sun, 04 Nov 2007 22:09:18 -0500
From: Fedor.Mesinger
To: Roger A. Pielke Sr.
Cc: Downscaling Group
Subject: Downscaling: domain size, more
Hi Roger, Barry, all,
Regarding
"when the agreement with the reanalysis (the "truth") deteriorated as the domain became larger, it alerted us to a fundamental problem with dynamic downscaling. The other papers since then [Rockel et al 2007; Lo et al 2007; Castro et al 2007] support our view on this limitation."
I feel it is essential to consider what it is that we are doing the downscaling for. If we are downscaling a reanalysis to a higher resolution "reanalysis", with no data input, using a parent reanalysis LBCs (e.g., Kanamaru and Kanamitsu: 57-Year California Reanalysis Downscaling at 10 km) then certainly beyond a certain domain size the bigger the domain the worse we shall do. We only need a domain big enough to describe smaller scale topography and other land surface features that should produce local phenomena which are to result in value added information compared to the driver reanalysis. When the domain becomes larger than this, we are introducing more and more of the chaotic nature of the atmosphere, and the message of the LBCs becomes weaker most likely making the result worse. This is what is tested in Type 2 experiments.
If however we are doing the downscaling in order to obtain higher resolution regional seasonal forecast or climate change information, then conclusions we can obtain from Type 2 experiments regarding the domain size are of a limited value. This is because as the domain size increases "the truth" of Type 2 experiments becomes less and less what the RCM should be expected to reproduce in a deterministic sense. If a perfect RCM, by definition performing just as the real atmosphere does, will reproduce our "truth" less and less precisely as the domain size increases, how can we expect an RCM to behave differently? No amount of Type 2 papers can change this as they do not reveal a fundamental problem with dynamic downscaling, but instead they illustrate the increased role of the chaotic nature of the model atmosphere as the domain size increases.
This said, of course real world RCMs will have their individual inadequacies which will result
in the departure of the statistical characteristics of the RCM results - e.g., the domain average kinetic energy - from those of the real atmosphere as reflected fairly well although not perfectly in the reanalysis data. Impact of such inadequacies seems to me should be expected to increase with increased domain size. Thus, Type 2 experiments do add important knowledge, in particular if and to the extent these departures with different RCMs are different. If however the departures when using two RCMs are about the same, we can only assume that the major generator of this departure is present in both RCMs. My assumption is that the relaxation LBCs are a good candidate for this role.
If you accept the explanation that the 1993 flood was due to the combination of the jet stream interaction with the low level jet moisture inflow from the Gulf, and even if you do not, could there have been anything else but the large domain which was responsible for the improvement of the RCM result compared to the driver GCM result seen in the plot I sent? The RCM domain in these runs, by the way, was about 11,000 by 9,000 km, roughly 19 % of the area of the globe. Had the RCM domain been small, this surely could not have been achieved, as in the limit of smaller and smaller domain an RCM obviously in the best case can only reproduce what the LBCs are telling it to do.
The cost/benefit issue of the domain size, how precisely and why the larger domain helps to the extent it does, and what is the best RCM strategy in seasonal forecast and climate change scenarios, are surely excellent questions which I feel justify indeed our putting together a "gigantic" proposal as proposed. I would be glad to be a part of. What I could do, with some funding for a grad student or a postdoc included, are runs using global and RCM versions of the same model, which is obviously cleaner than the COLA GCM/RCM experiments resulting in the plot I sent. My suggestion though is that emphasizing Type 3 as opposed to Type 2 experiments we ought to have a better chance of being funded, since 1) we can make a stronger case of likely arriving at positive results in the sense of demonstrated value-added, and 2) Type 3, and of course also the harder Type 4 experiments are what is needed for most of the end users of the information RCMs are meant to provide.
Regret perhaps repeating myself in some of the points made but I hope that this is a better organized and overall a more compact overall text re the domain size than those I sent so far.
Fedor
Date: Sun, 04 Nov 2007 22:15:23 -0800
From: Yongkang Xue
To: Fedor.Mesinger, Roger A. Pielke Sr.
Cc: Barry Lynn
Subject: Re: Downscaling: domain size, more
Dear All, It is indeed a very enlightening discussion on the RCM issues. I have a few thoughts based on your discussions, and would like to share with you.
(1). Reanalysis LBC versus GCM LBC. RCMs suppose to provide small scale features which are NOT exist in LBCs. Otherwise, there is no need to have RCM's downscaling. So the question is under what kind of LBCs, a perfect RCM can properly conduct downscaling. It seems one consensus is that Reanalysis LBC may be able to produce proper downscaling but not GCM (with less consensus). Since both global reanalysis and GCM have NO small scale information, the questions would be how good large scale features should be for a proper perfect RCM to conducting proper downscaling. If we give the quality of global reanalysis LBC a scale 90-100 (since there are different reanalyses with different quality), the question would be for a GCM, at what scale, (say 80, 50, etc?), their LBC are possible to help RCM to add new (small scale) value, or not at all. Here the scale is just an example, more proper criterion could be "proper jet stream", etc.
(2). The domain size. When the RCM domain becomes larger and larger, it should be able to generate more and more independent information from LBCs. The largest RCM is a GCM. The question is whether a very high resolution GCM is able to properly produce both large scale and small scale features. Very high resolution GCMs already exist and have generated quite a few results. I have not done a thorough investigation on them, but from some results I saw, it seems to me RCMs could still survive for a while. Now the question is how large the RCM domain should be. Our recent study (Xue et al., 2007, J. Climate) has indicated that when the RCM domain gets larger, the internal noise is so large which produced significant regional climate drafting. Therefore, although we wish to have a large domain for RCM to generate correct local information, but it should not be too large because of the climate drifting /noise, etc. A proper domain seems important but the real size may very much depend on individual model.
I feel these issues could be further tested in multi-RCMs.
Since I have two classes this quarter, I may not be able to respond your comments promptly.
Yongkang
Date: Mon, 5 Nov 2007 06:08:42 -0700 (MST)
From: Roger A. Pielke Sr.
To: Fedor.Mesinger
Cc: Downscaling Group
Subject: Re: Downscaling: domain size, more
Hi Fedor, I agree we should focus on Type 3 runs for the proposal, as this is what is needed by forecasters (and data is available to quantify real world skill). Your suggestion to use the same model for both the global and RCM scale for this purpose is an excellent idea.
Type 2 does permit testing the ability of the RCM to represent the dynamics and physics of the real world, however, so I still suggest we also do these experiments, since, as you suggest, we can test how chaotic the system is for different weather patterns and seasons (and ensemble runs can be made to determine if some of the realizations within the envelope capture what actually occurred). We can also assess the value-added of spectral nudging, which several of us concluded is the optimal approach, as well as an "optimal" domain size.
Type 3 runs also provide an upper bound for the skill possible for Type 4 runs (since Type 4 also has to skillfully predict SSTs, for example).
Liang- are you still going to be able to take the lead on this?
Best Regards, Roger
Date: Mon, 5 Nov 2007 08:40:16 -0700
From: Christopher L. Castro
To: Fedor.Mesinger
Cc: Downscaling Group
Subject: Re: Downscaling: domain size, more
Fedor and Roger: I have been keeping up on all these emails and the discussion. I do have some more detailed thoughts I'd like to outline at some point, but I haven't had time owing to my very busy schedule this semester.
One point I'd like to make with respect to the CA Regional Reanalysis runs that are referred to. If I'm not mistaken, I believe those runs used something of an equivalent of a spectral nudging term (that is either described within that paper or referenced in an earlier paper) similar to that used in some finite difference model studies that Roger has discussed (e.g. Miguez-Macho et al. 2005, Rockel et al. 2007, von Storch et al. 2000).
So I do agree with Fedor, of course, that 1) we need a large enough domain for RCM simulation to be physically useful, and 2) the problems that are observed with loss of kinetic energy at large scales (e.g. Castro et al. 2005, Rockel et al. 2007) can be probably be traced back to the specification of the lateral boundary forcing. I disagree with respect as to why Type 2 experiments are substantially different than Type 3 experiments. Only the type of lateral boundary forcing is being changed (from a "perfect" reanalysis to a GCM).
It would be very straightforward to see if the same type of behavior (e.g. loss of kinetic energy at large scales) is observed in a Types 3 and 4 downscaling modes. Simply force the RCM with data from a GCM, such as that of a NOAA CFS seasonal forecast (Type 3) or an IPCC projection simulation with an AOGCM (Type 4). Then do the similar analysis as I did in our 2005 paper and see if the behavior is equivalent. I've been wanting to do this myself, but I haven't had the time yet. I also encouraged our German colleagues to do this in the Rockel et al. (2007) paper in response to a similar comment as Fedor's (from one of the reviewers). I think you can see, though, from the examples I gave, how very important the issue is--especially the latter one involving climate change projections.
I agree there needs to be some comprehensive community statement on this. Perhaps a large proposal is the way to go. I don't know what the chances of it getting funded would be, though. The funding situation looks pretty bleak right now from the perspective of a young faculty member.
Regards--Chris
Date: Mon, 5 Nov 2007 09:45:28 -0600
From: Zong-Liang Yang
To: Roger A. Pielke Sr., Fedor.Mesinger
Cc: Downscaling Group
Subject: Re: Downscaling: domain size, more
Hi Roger, Yes. Give me one week as we are trying to digest and synthesize all the excellent comments to date.
Best regards, Liang
Date: Mon, 5 Nov 2007 10:25:19 -0700 (MST)
From: Roger A. Pielke Sr.
To: Zong-Liang Yang, Downscaling Group
Subject: Re: Downscaling: domain size, more
Thanks Liang! We will look for your leadership on this.
Best Regards, Roger
Date: Mon, 5 Nov 2007 11:10:05 -0700 (MST)
From: Roger A. Pielke Sr.
To: Barry Lynn, Yongkang Xue,
Cc: Downscaling Group
Subject: Re: Downscaling: domain size, more
Hi Barry and Yongkang, Thanks for your comments which are well expressed. I suggest we develop a working hypothesis; perhaps building on
"Type 3 dynamic downscaling can provide improved predictive skill on the regional and local scales"
Then we can test for different locations and times of years with different models. I really like Fedor's suggestion of using the same model for the global and regional scale (we could also then test one-way and two-way interaction across the RCM's lateral boundaries). Type 2 dynamic downscaling could be used also with the RCM for lateral boundary conditions (from the reanalysis) to see how it compares with the Type 3 runs.
Roger
Date: Mon, 5 Nov 2007 09:57:25 -0800 (PST)
From: Barry Lynn
To: Yongkang Xue, Fedor.Mesinger, Roger A. Pielke Sr.
Cc: Downscaling Group
----------------------------------------
Dear Xue, Fedor, Roger:
(From my morning: GMT + 5)
I think the arguments have been well stated. We need to (as aptly put by a coworker) better define: "where meteorology meets climatology."
An example: I was asked to downscale aGCM as fast as possible to study climate change impacts on water supply in the NY watershed. To this end, tt was requested that I set up the smallest domain possible. I argued that we need to extend our domain westward to the Rocky Mountains so as to better simulate the upstream forcing. I lost the argument, but the simulations proved me correct.
Yet, had we used Ranalysis data with accurate upstream information, then it would have been more appropriate to set up a smaller domain.
So, as noted here: a larger RCM domain has the most opportunity to correct biases in the GCM, but the greatest likelihood to drift away from the GCM climatology.
I am definitely in favor of submitting a mega-proposal. The question is: what is the hypothesis, what is the scientific question, and who will receive it. Since RCM downscaling is the fad now, maybe our focus should be on determining the optimum strategy for downscaling GCM data (Type 3 and 4 I think). Obviously, this needs to be stated more clearly in terms of a hypothesis (i.e., we propose that there is an optimum strategy for downscaling GCMs...)
Barry
Date: Mon, 05 Nov 2007 21:55:34 -0800
From: Yongkang Xue
To: Roger A. Pielke Sr., Barry Lynn
Cc: Downscaling Group
Subject: Re: Downscaling: domain size, more
Hi Roger, To submit a big proposal, we may have to do some ground work. NOAA CPPA just had a call regarding the Type 4 dynamic downscaling. I am not that sure any other agency is willing to support such big initiative at this stage. I feel we may have to work on the BAMS paper first before pursuing funding further.
Yongkang
Date: Mon, 5 Nov 2007 22:10:44 -0800 (PST)
From: Barry Lynn
To: Roger A. Pielke Sr
Cc: Downscaling Group
----------------------------------------
Hi: I would suggest using the WRF to make these simulations. NSF and NCEP already support this code, which has been and can be used to produce climatology and meteorology.
We could use the NCAR CCM3 to obtain some general climatic conditions during the present, and future decades, or just try to run the whole thing out for 30 years for select locations. I am suggesting the next thirty years because climate change impacts do seem to be important now. We would then set up the WRF to cover a large area at 108 km grid resolution, and then downscale to 36 and 12 km resolution from there.
Roger, could you please explain in more detail about what you mean concerning type 2 vs type 3. From what I gather up to now, type 2 involves running simulations over the ocean, without any impacts of the surface boundary topography. This is done to see how well the model reproduces the larger scales while it is downscaling. It seems to me, though, that when topography exists it could "help" the RCM preserve the larger scales. Also, we need to address the question how well we want the RCM to preserve the large scale (or synoptic scale) in the GCMs, when these types of systems may not be well simulated.
Barry
Date: Mon, 5 Nov 2007 22:12:26 -0800 (PST)
From: Barry Lynn
To: Yongkang Xue , Roger A. Pielke Sr.
Cc: Downscaling Group
Subject: Re: Downscaling: domain size, more
----------------------------------------
Hi: NOAA does use the WRF data. Maybe they would be interested in a WRF model focused project?
Also, please remind me what is Type 4 -- RCM climate simulations?
Barry
Date: Tue, 6 Nov 2007 05:53:00 -0700 (MST)
From: Roger A. Pielke Sr.
To: Yongkang Xue
Cc: Downscaling Group
Subject: Re: Downscaling: domain size, more
Hi Yongkang, I agree we need the white paper (BAMS) first. Liang has agreed to take the lead on this. Once we agree on the white paper (and clearly define our hypothesis (or hypotheses), then it should be very straightforward to complete a joint proposal among us.
Best Regards, Roger
Date: Tue, 6 Nov 2007 06:05:18 -0700 (MST)
From: Roger A. Pielke Sr.
To: Barry Lynn
Cc: Downscaling Group
Subject: Re: Downscaling: domain size, more
Hi Barry, I disagree on the value of future (years into the future predictions). They are not testable, and thus do not meet this requirement of a scientific study, other than as a process (sensitivity) study. This is why we need to link even the Type 4 runs (with retrospective simulations) to the reanalyses. I do agree we can do seasonal runs with Type 3 runs (described below), as we can validate them in a reasonable time (both from retrospective runs and with predictions for coming seasons in the next few years).
On the definition of Type 3 downscaling, it is not as you wrote below. With Type 3, some of the climate variables are prescribed rather than predicted (e.g. SSTs). AMIP runs have been Type 3 when an RCM is used to downscale from them. Type 4 runs predict all aspects of the climate system.
We have two tables in our paper (Tables 1 and 2 )
Castro, C.L., R.A. Pielke Sr., and G. Leoncini, 2005: Dynamical downscaling: Assessment of value retained and added using the Regional Atmospheric Modeling System (RAMS). J. Geophys. Res. - Atmospheres, 110, No. D5, D05108, doi:10.1029/2004JD004721.
that describe this distinction.
I am suggesting the ocean runs for both Type 2 and 3 RCM runs as I agree topography (and landscape) can add predictive skill if not overwhelmed by the larger scale atmospheric fields (from the lateral boundary conditions and interior nudging). Over the ocean without any significant SST gradients, we can isolate what value is added in the atmospheric fields alone by using an RCM. Using the same model for both the global scale and the regional scale (with finer spatial resolution) as Fedor suggested is a very effective way to do this.
Best Regards, Roger
Date: Tue, 06 Nov 2007 22:22:53 -0500
From: Fedor.Mesinger
To: Barry Lynn
Cc: Downscaling Group
Subject: Re: Downscaling: "using the WRF"
Hi Barry, all, presumably Barry when you say "using the WRF" you have in mind the NCAR WRF? In case you - or some of the readers - are not familiar with specifics of the WRF jargon, there is "the WRF model", but there are presently in use two dynamical cores, NCAR's so called ARW core, and NCEP's NMM core. But in ordinary language, these are two I'd say totally different models, which due to common framework rules, without much of an extra effort can use same or some of the same physics subroutines.
In my opinion, performance of the two WRF cores is likely to be fairly different in an RCM setting, because of the diferent LBC schemes they use. NCEP WRF is a modification of the Eta, and is using the same LBC scheme. NCAR WRF is the one that Ruby Leung used trying to do the 1993-1988 case, with a result less than satisfactory. This experience in comparison with that of the plot I sent I feel should not be neglected.
Fedor
Date: Wed, 7 Nov 2007 21:47:26 -0800 (PST)
From: Barry Lynn
To: Downscaling Group
Subject: Re: Downscaling: domain size, more
Dear Roger: Thank you for your e-mail.
I suggest that we design the study to include simulations that downscale GCM in the next few decades. The design of these simulations would be based on the results of our first part of the study where we try to optimize the RCM configuration. I think such an "add-on" could be useful and might be desired by policy makers who are increasingly worried about the future.
Barry
Date: Wed, 7 Nov 2007 21:50:47 -0800 (PST)
From: Barry Lynn
To: Fedor.Mesinger
Cc: Downscaling Group
Subject: Re: Downscaling: "using the WRF"
Hi: The NCAR WRF was used recently by Ruby Leung to downscale a GCM (the CCM3?). So, perhaps some of the problems Fedor wrote about have been solved.
Barry
Date: Sat, 10 Nov 2007 14:18:37 -0500
From: Fedor Mesinger
To: Barry Lynn
Cc: Downscaling Group
Subject: Re: Downscaling: "using the ... model"
Hi again Barry, all, Perhaps at this stage it is not helpful to emphasize using in an eventual project a specific model and/or specific pairs of models? I feel we might want to choose a specific set or sets of models only at some kind of a "phase 2" stage, if at all. Restricting our model choices ahead of time - or even later - might compromise our eventual results I am afraid.
On the other hand, if and/or when we do wish to move to restricting our choices as to model(s) and/or sets of models, I feel we should consider the following:
Do we have results obtained by a candidate model that demonstrate its ability to show value added skill in an RCM mode over its global driver data? What are those? If available: How does this value added compare against value added demonstrated by other models or sets of models?
Does for a given RCM exist a global version of the same model - so that we can use the pair to make clean experiments as some suggested in the discussion? For example, re the Eta RCM there is an up-to-date global version (Zhang, H., and M. Rancic: 2007: A global Eta model on quasi-uniform grids. Quart. J. Roy. Meteor. Soc., 133, 517-528); re the RAMS, I believe just as well (OLAM).
Is there a benefit in restricting our effort to the use of a single model or a single set of models? Does doing that compromise our chances of obtaining generally valid results?
Is there a satisfactory institutional support and/or community of developers/ users for a given model/ sets of models, that might be counted upon to offer assistance in case of problems encountered?
Of course, answers to all of those are not likely to be clear-cut, so that balancing the
attractiveness of specific choices vs their downsides will be needed. But then again, it might
well be too early for this sort of talk.
Fedor
Date: Mon, 12 Nov 2007 15:39:33 -0600
From: liang@mail.utexas.edu
To: Downscaling Group
Subject: Re: Downscaling: "using the ... model"
----------------------------------------
Dear All, Attached is a summary of our email changes compiled by Jeff Lo. Please let me know if there are any corrections/additions needed. In the meantime, I am trying to write an introduction and to re-organize the materials so they flow better.
Cheers, Liang
Date: Tue, 13 Nov 2007 07:48:46 -0700 (MST)
From: Roger A. Pielke Sr.
To: liang
Cc: Downscaling Group
Subject: Re: Downscaling: "using the ... model"
Hi Liang, Thank you for summarizing our e-mail discussion. I look forward to the first draft of our paper.
Best Regards, Roger
Date: Fri, 11 Apr 2008 08:20:31 -0600 (MDT)
From: Roger A Pielke Sr
To: downscale
Subject: NEW REGIONAL FOCUS ON CLIMATE CHANGE BY IPCC
Hi All, FYI; see below. I assume they will still start with global models and then dynamically (or statistically) downscale. We should contribute to the discussion (our white paper; Liang :-)), including the alternate approach of starting from the local and regional (a bottom-up perspective).
Roger
************************************
NEW REGIONAL FOCUS ON CLIMATE CHANGE BY IPCC
The Intergovernmental Panel on Climate Change (IPCC) will focus its next report, due in 2014, on a search for regional rather than global scenarios and possible solutions, and highlight the economic and social aspects of climate change, IPCC chairman Rajendra Pachauri said Thursday. Speaking at a press conference in Budapest, he added that the just-concluded IPCC session had adopted a special report on climate change and water management, pointing to the urgent need to adapt to floods and drought. The IPCC's main task is to evaluate the risk of climate change caused by human activity, and publish special reports on topics relevant to the implementation of the UN Framework Convention on Climate Change. IPCC closed its four-day session in the Hungarian capital Budapest on Thursday. It was the panel's first major meeting after it received the Nobel Peace Prize in 2007.
Date: Thu, 17 Apr 2008 11:30:38 -0600 (MDT)
From: Roger A Pielke Sr
To: downscale
Subject: 2009 Downscaling Workshop in Sweden
Hi All, While some of you know about this meeting, others do not. It is a very important opportunity to assess where the science community is with respect to downscaling from multi-decadal global climate models, as well as other downscaling time scales.
http://www.baltex-research.eu/RCM2009/
Roger
From: Dallas Staley
To: downscale
Subject: Downscale Re: Meeting
Dear All:
I recently read the papers cited below. The general conclusion is that the most realistic downscaling runs are obtained when the regional model is internally nudged (preferably with a spectral nudger) with Reanalysis. The least realism is obtained when the model is allowed to freely develop its own longwave pattern.
My questions are:
i) should the same conclusion apply to regional climate simulations with a GCM?
(Either seasonal or decadal.) Here, we are concerned that the GCM does not well
reproduce the longwave pattern and has many biases.
ii)
If not, would the most realistic longwave patterns and the statistics of the localized meteorology be obtained without any internal domain nudging?
Barry Lynn
Castro, C.L., R.A. Pielke Sr., and G. Leoncini, 2005: Dynamical downscaling: Assessment of value retained and added using the Regional Atmospheric Modeling System (RAMS). J. Geophys. Res. - Atmospheres, 110, No. D5, D05108, doi:10.1029/2004JD004721. [se Table 1 and 2].
Lo, J.C.-F., Z.-L. Yang, and R.A. Pielke Sr., 2008: Assessment of three dynamical climate downscaling methods using the Weather Research and Forecasting (WRF) Model. J. Geophys. Res., 113, D09112, doi:10.1029/2007JD009216
Rockel, B., C.L. Castro, R.A. Pielke Sr., H. von Storch, and G. Leoncini, 2008: Dynamical downscaling: Assessment of model system dependent retained and added variability for two different regional climate models. J. Geophys. Res., submitted.
[plus the Castro et al 2005 paper];
and seasonal forecasts.
Castro, C.L., R.A. Pielke Sr., J. Adegoke, S.D. Schubert, and P.J. Pegion, 2007: Investigation of the summer climate of the contiguous U.S. and Mexico using the Regional Atmospheric Modeling System (RAMS). Part II: Model climate variability. J. Climate, 20, 3866-3887
Beltran-Przekurat, A., C.A. Marshall, and R.A. Pielke Sr., 2008: Ensemble re-forecasts of recent warm-season weather: Impacts of a dynamic vegetation parameterization. J. Geophys. Res. - Atmos., submitted.
Date: Mon, 30 Jun 2008 14:22:04 -0600 (MDT)
From: Roger A Pielke Sr
To: downscale
Subject: Re: Meeting (fwd)
Hi Barry
I will add my comments here with respect to your really central questions.
My perspective is that downscaling from reanalyses (Type II) represents the upper limit of downscaling skill with respect to RCMs that have forgotten their initial conditions (what Chris Castro and I refer to as Types II, III, and IV
downscaling).
The extent to which an RCM cannot reproduce the reanalysis resolved patterns within its regional domain from a global reanalyses (without nudging), will be even greater (less accurate) for GCMs since they are not tuned with observed
data as are the reanalyses.
At least with the Type II downscaling (from the reanalyses), the lateral boundary conditions are feeding in information that has an observationally based component. With Type IV downscaling, the GCM biases are entering the model in a manner that is not constrained by any observed data.
We observed the deterioration of the RCMS, when run from the NCEP/NCAR reanalysis, as we enlarged the RCM domain. When nudging is applied, they do better (as defined by being able to reproduce the resolved spatial scales of
the global model), but then are enslaved to the parent GCM (or reanalysis) with respect to spatial features on the resolved scale of the global model with all of its biases.
I look forward to other comments on this subject!
Best Regards, Roger
Date: Mon, 30 Jun 2008 13:17:21 -0700 (PDT)
From: Barry Lynn
To: downscale
Subject: Re: Downscaling
--- On Mon, 6/30/08, Roger A Pielke Sr <pielkesr@ciresmail.colorado.edu> wrote:
Hi All: A related question is now added below.
Barry asked us to forward this to everyone on our list. Please reply to everyone on his very important questions.
Dear All:
I recently read the papers cited below. The general conclusion is that the most realistic downscaling runs are obtained when the regional model is internally nudged (preferably with a spectral nudger) with Reanalysis. The least realism is obtained when the model is allowed to freely develop its own longwave pattern.
My questions are:
i) should the same conclusion apply to regional climate simulations with a GCM? (Either seasonal or decadal.) Here, we are concerned that the GCM does not well reproduce the longwave pattern and has many biases.
ii) If not, would the most realistic longwave patterns and the statistics of the localized meteorology be obtained without any internal domain nudging?
related:
iii) Is it realistic to assume that a regional model can reproduce a longwave pattern that a GCM cannot? If so, does this have any real connection to the GCM longwave pattern? Is it at all consistent with it?
Barry Lynn
Date: Mon, 30 Jun 2008 15:09:50 -0600 (MDT)
From: Roger A Pielke Sr
To: Barry Lynn
Cc: downscale
Subject: Re: Downscaling
Hi Barry, I do not see how a regional model can reproduce realistic long wave patterns, as these are hemispheric features.
Roger
Date: Mon, 30 Jun 2008 21:16:39 -0700 (PDT)
From: Barry Lynn
To: Roger A Pielke Sr
Cc: downscale
Subject: Re: Downscaling
Hi Roger:
You suggest that regional models cannot reproduce the longwave pattern, which has hemispheric wavelength. Yet, it is the longwave pattern that forces the synoptic pattern, which then is downscaled by the regional model. Can a regional model produce a realistic synoptic pattern?
1) Can a regional model produce regional climatology, even without nudging?
2) If not, is it reasonable to use regional models to obtain local climatology in GCMs if the the boundary conditions are not realistic?
3) What grid resolution for GCMs is required to produce a reasonable longwave pattern?
Barry
Date: Tue, 1 Jul 2008 08:10:26 -0600 (MDT)
From: Roger A Pielke Sr
To: Barry Lynn
Cc: downscale
Subject: Re: Downscaling
Hi Barry, You are very effectively getting at the core of the questions that need to be quantitatively answered!
My conclusions, based on the literature and our papers, are that RCMs (Type II, III, IV) must rely on the larger scale model (or reanalysis) for the long wave pattern. A regional model can obtain a "climatology" but the realism of its representation requires the accurate representation of the large scale pattern.
The climatology also must be compared with observed data, not just statistically, but on an event basis. That is why we adopted the Type II downscaling in our papers, as we have a real world comparison. We found that, without nudging, the relation to reality (as defined by the reanalysis) deteriorated as the RCM domain was made larger.
My conclusion is that the RCMs, when used with Type II, III and IV downscaling are just sophisticated interpolators to a finer terrain and landscape grid (and are useful for this purpose), but they are still enslaved to the parent GCM (or reanalysis).
For a global model, I propose that a 10km delta-x would suffice for avoiding the need for RCMs (and, of course, may still not be realistic if all of the forcings and feedbacks are not accurately represented). For the resolution required for including the long wave features, I propose 100km both the shorter and planetary waves need to be present, and analyses of centers of mid-tropospheric vorticity are typically on the order of several hundred kilometers in wavelength.
I look forward to reading the perspective of others on our mailing group also on this issue.
Best Regards
Roger
Date: Tue, 1 Jul 2008 11:05:07 -0400
From: "Niyogi, Dev"
To: Barry Lynn
Roger A Pielke Sr
Cc: downscale
Subject: RE: Downscaling
Hi Barry and all,
We had a discussion about the last two questions you raise, a group meeting at Purdue. In particular this paper that my colleague published in JGR for doing cloud scale assessments under climate change scenarios was discussed. The "Global model to Cloud" versus "Global To Regional to Cloud" model based assessment to me was a provocative (and somewhat controversial) conclusion from this study and I would be interested in getting your thoughts.
Dev.
Telescoping, multimodel approaches to evaluate extreme convective weather under future climates
Robert J. Trapp, Brooke A. Halvorson, and Noah S. Diffenbaugh
[1] Understanding of the possible response of severe convective precipitating storms to elevated greenhouse gas concentrations remains elusive. To address this problem, telescoping, multimodel approaches are proposed, which allow representation of a broad range of processes that could regulate convective storm behavior. In the global-cloud approach (G-C), the NCEP-NCAR Reanalysis Project (NNRP) global data set provides initial and boundary conditions for short-term integrations of a mesoscale model and nested convective-cloud-permitting domain. In the global-regional-cloud approach (G-R-C), the NNRP data set provides initial and boundary conditions for long-term integrations of a regional climate model, which in turn forces short-term integrations of a mesoscale model and nested convective-cloud-permitting domain. Upon applying these approaches to historical extreme convective storm events, it was found that the global-scale data could be dynamically downscaled to produce realistic convective-scale solutions. In particular, tornado proxies computed from the model-simulated winds were shown to compare well in relative numbers to those of tornado observations on many of the days considered. This supports the telescoping modeling concept as a viable means to address effects of elevated greenhouse gas concentrations on convective-scale phenomena. In an evaluation of the two approaches, it was also found that simulations of the historical events by the G-C were superior to those by the G-R-C. Sensitivity of the convective-scale processes to details in the downscaled synoptic-scale flow, and to the placement of the mesoscale model domain within the regional climate model, reduced the effectiveness of the G-R-C.
Citation: Trapp, R. J., B. A. Halvorson, and N. S. Diffenbaugh (2007), Telescoping, multimodel approaches to evaluate extreme convective weather under future climates, J. Geophys. Res., 112, D20109, doi:10.1029/2006JD008345.
Dr. Dev Niyogi
Assistant Professor and Indiana State Climatologist
Departments of Agronomy and Earth & Atmospheric Sciences
Purdue University
Lab: http://landsurface.org
State Climate Office: http://iclimate.org
Date: Tue, 1 Jul 2008 14:35:38 -0600 (MDT)
From: Roger A Pielke Sr
To: "Niyogi, Dev"
Cc: Barry Lynn ,
downscale
Subject: RE: Downscaling
Hi Dev, This study uses a Type II downscaling approach, and, in this context, is an important study. However, its title is misleading "Telescoping, multimodel approaches to evaluate extreme convective weather under future climates". The paper is not really on the subject of how to skillfully predict future climates since that requires Type IV runs with the global model initialized with observed data, but then free to evolve as its model physics allow.
There is no testing of the predictive capability of the approach since the study uses "perfect boundary conditions" from the reanalysis. The authors write
"Although investigations of the response to 21st century climate forcing will require atmosphere-ocean general circulation model-simulated fields as part of both configurations, here we investigate the modeling approach performance given "perfect boundary conditions".
In the conclusions, they write
"Although we have assessed the ability of these two telescoping modeling approaches to simulate historical events, it is necessary to consider that in order to robustly investigate CPS behavior in response to elevated greenhouse gas forcing, any successful approach must ultimately capture critical processes independent of observed or reanalysis data. Thus far, the performance of the G-C indicates that a modeling system consisting of a GCM driving a convection-permitting model could be applied for studying CPS dynamics under elevated greenhouse gas forcing. However, in order for this configuration to be viable, the global model must accurately capture both the large-scale features and the initial surface conditions represented in the reanalysis data set."
Their study does not address the "viability" of being able to skillfully predict future severe weather. What they have done (with Type II downscaling) is a necessary but not sufficient condition.
Thank you for sharing this paper with us! It is an effective paper to focus our discussions.
Roger
Date: Thu, 3 Jul 2008 11:29:51 -0700 (PDT)
From: Barry Lynn
To: Roger A Pielke Sr
Cc: downscale
Subject: RE: Downscaling
Dear Dev/Roger:
I read the paper in question. I thought it was quite interesting in that it demonstrated that one can obtain a very good simulation of severe weather using the WRF with telescoping grids off Reanalysis data.
Yet, I don't think it really showed that running a regional model outside the WRF grids is a bad thing. Rather, it showed that the regional model will drift from the Reanalysis solution over time to produce wave patterns that are not necessarily consistent with the Reanalysis. Hence, the WRF with the RegCM3 did not produce convection on the right day and the right place because the boundary conditions to the WRF were not consistent with such an event. Hence, one might think that
i) To obtain forecasts consistent with the reanalysis one should probably nudge (papers Roger sent me also reached this conclusion.
ii) The size of the nested domain should be as small as possible to prevent drift.
I would agree with this statement in wintertime or when the flow is predominately synoptic with maybe unorganized convection.
Yet, our WRF simulations off the GFS in spring are much, much better than those of the GFS (and better than those with nudging) -- since the WRF simulates the mesoscale circulations (sea breeze, small Sharav lows, mountain breezes) than the GFS. So, the answer in my mind is neither yes or no to nudging or a big or small domain when using Reanalysis (or the GFS).
Getting back to the subject above: in a GCM the longwave pattern does not realistically account for topography and higher wave energy. Hence, it might actually be better to run a model like RegCM3 with as large a domain as possible and without nudging. The RegCM3 might not be completely consistent with the GCM, but it may actually be more realistic on a statistical, seasonal scale.
Barry Lynn
Date: Thu, 10 Jul 2008 11:27:58 -0600 (MDT)
From: Roger A Pielke Sr
To: "downscale "
Subject: weblog
Hi All, I posted today a weblog that is of relevance to our mailing group. Comments (and guest weblogs!) welcome.
Roger
Date: Fri, 3 Oct 2008 07:38:00 -0600 (MDT)
From: Roger A Pielke Sr
To: downscale
Subject: Statistical downscaling - its value
Hi All, As we plan for the Lund meeting next spring, I have urged the use of statistical downscaling from the global climate models as the standard with which we must show an improvement using the regional dynamic downscaling method. I have reproduced my comments on this issue below:
"..... on the topic of the "Comparison of downscaling using RCMs with statistical downscaling from the same GCMs". We should not delete this topic but instead should highlight it.
The fundamental issue that needs to be assessed with respect to dynamic regional downscaling is whether this adds skill beyond interpolation (i.e. of which statistical downscaling is the most sophisticated type) from the global model results. This must be our control.
With respect to the types of downscaling that are presented in
Castro, C.L., R.A. Pielke Sr., and G. Leoncini, 2005: Dynamical downscaling: Assessment of value retained and added using the Regional Atmospheric Modeling System (RAMS). J. Geophys. Res. - Atmospheres, 110, No. D5, D05108, doi:10.1029/2004JD004721.
we need to assess the value added of dynamic downscaling from Type II, III and IV applications. If the regional dynamic downscaling cannot provide added skill, as compared with statistical downscaling, we are spending a lot of funding unnecessarily and are also misleading policymakers on the actual skill of regional climate projections."
Best Regards
Roger
Roger A. Pielke, Sr.,
Senior Research Scientist CIRES and Senior Research Associate ATOC, University of Colorado at Boulder and Professor Emeritus Colorado State University http://cires.colorado.edu/science/groups/pielke/
Date: Fri, 03 Oct 2008 08:52:39 -0500
From: Raymond W. Arritt
To: Roger A Pielke Sr
Cc: downscale
Subject: Re: Statistical downscaling - its value
Roger,
I agree in part with your statement:
> The fundamental issue that needs to be assessed with respect to
> dynamic regional downscaling is whether this adds skill beyond
> interpolation (i.e. of which statistical downscaling is the most
> sophisticated type) from the global model results. This must be our
> control.
However I believe this is not "the" fundamental issue but "a" fundamental issue. The other fundamental issue is whether statistical transfer functions derived for the current climate apply in future climates. It is entirely possible to generate a statistical downscaling method that provides excellent downscaling skill for the present climate but is of little use for the problem of climate change.
Ray
Date: Fri, 03 Oct 2008 08:18:21 -0700
From: Yongkang Xue
To: Roger A Pielke Sr, downscale
Cc: yxue@geog.ucla.edu
Subject: Re: Statistical downscaling - its value
Hi, I fully support Roger's notion for a comparison between statistical downscaling and dynamic downscaling. Statistical downscaling has been developed for many years (much earlier than the dynamic downscaling) and has been widely applied and dominant for many applications. For example, most future downscaling currently uses statistical downscaling because it does not need huge amount of computer time and many people are familiar with this. RCM downscaling community is small. The computational cost is high. There are different opinions towards the utility of this method. I think Ray's concern is valid. But these need to be investigate in a scientific way to find pros and cons in either method.
Yongkang
Date: Fri, 3 Oct 2008 09:46:03 -0600 (MDT)
From: Roger A Pielke Sr
To: Raymond W. Arritt
Cc: downscale
Subject: Re: Statistical downscaling - its value
Hi Ray, Good point; I agree 100% with you. On the need for the temporal invariance of the statistical downscaling relationships, this also applies to the parameterizations that have been developed for use in our dynamic downscaling models, as they are tuned with observed data from the current climate.
Roger
Date: Sat, 11 Oct 2008 07:27:56 -0700
From: Yongkang Xue
To: Roger A Pielke Sr
Cc: "downscale"
Subject: Re: comments on downscaling
Hi Roger, I agree with your ideas but for one thing. The most RCM runs can not be used to evaluate the river flow because the model needs at least 1 year to complete spin up/down. In another word, within first year, the river runoff in RCM was largely an initial condition issue. Some hydrologists may want to argue for more than one year. This is model dependent. One feasible approach may be to use RCM's products to force hydrology model producing river runoff.
Yongkang
Date: Sat, 11 Oct 2008 08:48:23 -0600 (MDT)
From: Roger A Pielke Sr
To: Yongkang Xue
Cc: "downscale"
Subject: Re: comments on downscaling
Hi Yongkang, I agree with you. The length of time that initial conditions are "remembered" with respect to different applications of the models (i.e. their "purpose" as Hans has effectively articulated) should be one of the evaluations of the RCM studies.
Best Regards, Roger
Date: Wed, 15 Oct 2008 09:51:11 -0600 (MDT)
From: Roger A Pielke Sr
To: "downscale"
Subject: New paper
Hi All, Our new paper further documents the serious limitations of Type III dynamic downscaling. Type IV will be even worse.
http://www.agu.org/contents/journals/ViewPapersInPress.do?journalCode=JD#id2007JD009480
The paper can be obtained from
Beltran-Przekurat, A., C. H. Marshall, and R. A. Pielke Sr. (2008), Ensemble re-forecasts of recent warm-season weather: impacts of a dynamic vegetation parameterization, J. Geophys. Res., doi:10.1029/2007JD009480, in press.
Roger
Date: Sat, 14 Feb 2009 08:53:13 -0700 (MST)
From: Roger A Pielke Sr <pielkesr@ciresmail.colorado.edu>
To: "downscale
Subject: some suggestions
Hi All, As many of you plan for the Lund meeting this spring, I have a couple of overview comments that I urge you to consider. They are:
1. There is a need for the establishment of a more thorough validation methodology to quantitatively assess value-added using dynamic regional downscaling. This requires using direct statistical downscaling from the parent model (or reanalysis) as the control. The dynamic downscaling needs to add skill (value) on top of what the statistical downscaling can provide, and this needs to be presented when dynamical downscaling results are presented.
2. These tests need to be performed separately for Type II, III and IV downscaling. For Type I, there is extensive and clear evidence (i.e. from NWP) that dynamic downscaling is superior in terms of value-added (skill) to statistical downscaling. The difficulty of adding value increases from Type II through Type IV. We also know that the skill for Type IV must be less than for Type III which must be less than Type II which must be less that Type I. The tests of skill need to be with real world data and not with other models.
3. In terms of how model results are presented, we should adopt "years into the simulation", rather than specific years (e.g. 2041-2050) as the later implies these are forecasts, when they are actually sensitivity ("what if") model runs. Since there is no way we can verify the model simulations for decades into the future, we need to be more precise in terms of how such results are presented.
I look forward also to your feedback on this comments.
Best Regards, Roger
Date: Sat, 21 Feb 2009 10:29:19 -0700 (MST)
From: Roger A Pielke Sr
To: Shi SONG
Cc: "downscale
Subject: Re: On Dynamical Downscaling -- from Shi Song
Hi Shi Song, I have copied to our mailing group; in the future the easiest way to respond to our e-mails is just to do a reply. Please see my comments below, and thank you for participating!
Best Regards
Roger
=================
On Sat, 21 Feb 2009, Shi SONG wrote:
> Professor Pielke:
>
> I've read your comments on dynamical downscaling on this webpage "http://cires.colorado.edu/science/groups/pielke/links/Downscale/", but I've no idea of how to reply on the webpage. So I think it'd be OK to write directly to you.
>
> 1. I agree with your statement that "in terms of how model results are presented, we should adopt "years into the simulation", rather than specific years." However I want to make sure what do you exactly apply this claim to? All model results (including simulations of 1980~1990 for example) or only those of the sensitivity("what if") model runs¡©My opinion is the former one because I think the RCM actually plays the same role in simulations of either the past or the future. The only thing that is different is the large-scale control: reanalysis for the past and GCM predictions for the future. Please confirm yours.
For retrospective runs, if they are made as initial value climate predictions or use reanalysis, than the actual years should be used. For future runs, unless they are run as initial value climate predictions, they should not use
actual years (or decades) but years into a "what if" simulation.
> 2. The "added value" means dynamical downscaling outperforms most other methods including statistical downscaling. Let us imagine in a Type 3(or 4) downscaling, the forcing GCM has large error. So the RCM
downscaling probably has large discrepancies as claimed by lots of researchers. Given that the statistical downscaling are in better consistent with the observation, isn't it meaningless? Because under such circumstances, a satisfying result is downscaled(or tuned) from a wrong background by the statistical method! It does not justify to prove the usefulness(superiority) of statistical downscaling.
Statistical downscaling is the benchmark we should compare with for Type 2 and 3, as upper bounds on the skill of Type 4. If the statistical downscaling more closely agrees with real world observations (for a set of observations not used in the construction of the statistical downscaling), why use the dynamic downscaling? However, if the dynamic downscaling is better, quantify its improvement for each predicted climate variable.
Please let me know if further clarification is needed. I look forward to discussing this with you.
Best regards,
Shi Song
2009-02-20
Date: Tue, 22 Sep 2009 07:38:01 -0600 (MDT)
From: Roger A Pielke Sr
To: downscale
Subject: FW: [Wrf-news] Information about Climate Model Downscaling Project called CORDEX (fwd)
Hi All , Joe Eastman has provided us with this information, in case you have not already seen it.
Best Regards, Roger
--
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Roger A. Pielke Sr.
Senior Research Scientist (CIRES) and Senior Research Associate (ATOC) Stadium 255-16, University of Colorado, Boulder, CO 80309 Professor Emeritus Colorado State University
-----Original Message-----
From: wrf-news-bounces@ucar.edu [mailto:wrf-news-bounces@ucar.edu] On Behalf Of
Chris Anderson
Sent: Monday, September 21, 2009 1:40 PM
To: wrf-news@ucar.edu; wrf-users@ucar.edu
Subject: [Wrf-news] Information about Climate Model Downscaling Project called
CORDEX
Dear WRF Users,
Information below is provided on a climate model downscaling program called CORDEX. If you intend to provide WRF simulations for CORDEX, please particularly make note of the paragraph describing the CORDEX WRF Wiki.
COordinated Regional climate Downscaling EXperiment (CORDEX) is a framework aimed at improving coordination of international efforts in regional climate downscaling research. CORDEX was initiated as a result of the task Force on Regional Climate Downscaling, formed by the World Climate Research Program (WCRP). For details on CORDEX, see http://wcrp.ipsl.jussieu.fr/RCD_Projects/CORDEX/CORDEX.html
Best,
Chris
Dear CORDEX Community
A wiki has been established for those planning to do WRF simulations for CORDEX. Because the WRF users community is widely dispersed, this wiki is to help members of the WRF community communicate their intentions to each other for CORDEX simulation. It is also intended to allow CORDEX participants using WRF to coordinate efforts for a CORDEX region and perhaps engage in collaboration. Please note if others before you have already registered intentions to simulate a particular region, and work with them to ensure the most thorough set of WRF simulations for the region, without repeating work.
The wiki is at
http://www.meteo.unican.es/wiki/cordexwrf
To create an account on this wiki, click on the "login" link at the top of the page, and then "UserPreferencs" on the login page. A template providing a minimal information request appears at
http://www.meteo.unican.es/wiki/cordexwrf/GroupTemplate
You can subscribe to selected pages or to the whole wiki (in
UserPreferences) upon registering.
If you have problems, please contact me <gutowski AT iastate.edu> or Jesus Fernandez <fernandej AT unican.es>
Thanks,
Bill Gutowski
11111011001
Christopher J. Anderson, PhD
Scientist, Assistant Director Climate Science Initiative
Iowa State University
Office: 515-294-9948
Fax: 515-294-2619
http://climate.agron.iastate.edu/ResearchTeam/AndersonChristopher.html
_______________________________________________
Wrf-news mailing list
Wrf-news@ucar.edu http://mailman.ucar.edu/mailman/listinfo/wrf-news
Nondiscrimination Policy • Legal and Trademarks |
