Welcome, Guest
Username: Password: Remember me
General topics
  • Page:
  • 1
  • 2

TOPIC: Spectral grid

Spectral grid 4 years 5 months ago #1472

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
I would like to have a bit wider discussion on the spectral grid, and involve dynamics people a bit more. I know Alvaro Subias, Xiaohua Yang and Ole Vignes has done several tests with various spectral grids, and we have looked at the KE spectra and temperature spectra to understand a bit more the behavior at the highest wave-numbers.

We also have a new HIRLAM member, Colm Clancy from Met Eireann who is starting to work on dynamics and numerics, perhaps a forum discussion on this can help him get up to speed on some of the discussions.

I will below paste in some of the earlier e-mail conversation that's been going on regarding the "spectral grid series", it would be helpful if you could post your experiences regarding this topic here from now on. Thank you!

Spectral grid 4 years 5 months ago #1473

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
From Xiaohua 1/26/2016:

Attached please find some summary slides from tests of cubic grid with 40h1b1. I was to send you an update of my previous summary slides about validation of 40h1 but the file looks to be too big to send, so please try the link

docs.google.com/presentation/d/18PdJrU5-...aNM/edit?usp=sharing

Spectral grid 4 years 5 months ago #1474

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
500hpa_spectra.png


CUBICAROME_precip.png


LINEARAROME_precip.png


Lisa Bengtsson 2/2/2016

I looked at the kinetic energy spectra for the cubic vs the linear grid from a case from Xiaohua's experiment. As can be seen the two spectra follow each other exactly, but the cubic grid has a truncation of 300 waves and the linear of 600 waves. Essentially this means that for this particular domain DX=1200 = 1200*2.5 km = 3000 km. With linear grid this gives a smallest resolved wave of 3000/600 = 5 km (=2*dx), but for the cubic grid 3000/300 = 10 km (=4*dx).

We saw rather large impact on convective cells "resolved" at 2.5 km resolution when applying more horizontal diffusion on the smallest waves, therefore it would be interesting to see the impact on convective behavior if we now change the effective horizontal resolution, I would like to look at this, but I am not able to convert the history files IC* to grib using girb_api from cycle 40, has anyone done this successfully?
Last Edit: 4 years 5 months ago by Lisa Bengtsson.
The following user(s) said Thank You: Jana Sanchez Arriola, Emily M. Gleeson

Spectral grid 4 years 5 months ago #1475

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
Xiaohua Yang, 3/2/2016:

I am a bit uncertain about direction here. I don't assume there will be much more work on research on cubic grid now that Mariano has retired. For the time being linear option is certainly a safer one, but leaving cubic grid out is really a pity as it is quite attractive for operational purpose. The things is, at least for operational services, selection of cubic grid or not will have different consequence for future tuning, say on wind speed. Right now the HARMONIE surface wind for some nordic area is non-optimal, and with the larger wind bias in cubic runs, sso option again become a tuning tool. Using it or not has an impact on HARMONIE reference.

Spectral grid 4 years 5 months ago #1476

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
Xiaohua, 8/2/2016

In view of uncertainty about cubic grid I've done some new test lately with use of quadratic grid for the MEtCoOP and IMO domains, two of which cubic grid seem to deviate quite a bit on wind speed etc compared to linear grid. The test results so far showed quadratic grid seem to have good potential. ALso, it is quite much cheaper and stable than linear one. Actually it looks even cheaper than cubic grid. Will check on the last fact and report back later.

hirlam.org/portal/validation/38h1/MCPquadratic

hirlam.org/portal/validation/38h1/IMOquadratic

Spectral grid 4 years 5 months ago #1477

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
Ole Vignes, 8/2/2016:

results [with quadratic grid] are now under metcoop.met.no/verif/201507_cy40_lin_quad_cubic_export/

Note that here linear grid is run with 75 sec. tstep, quad and cubic with 90 sec.

Spectral grid 4 years 5 months ago #1478

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
Xiaohua, 8/2/2016:

Thanks for the results. Your test seem to confirm the fact that use of quadratic grid indeed results in model behavior that is well in between those from linear and cubic ones. Actually, such trend seems to be clearer in your runs than my runs for winter episode, where the wind bias for Norweigian station lists goes through distinctly different route in terms of daily error. Will you also be doing test for a winter episode?

hirlam.org/portal/validation/38h1/MCPquadratic

I used high order orographic smoothing but removed explicit horizontal diffusion in my quadratic test runs.

My tests are in
~nhe/hm_home/MCP_20150120_75_quadratic
~nhe/hm_home/IMO_201211_quadratic

Spectral grid 4 years 5 months ago #1479

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
Ole Vignes, 8/2/2016:

The run with linear grid (and 75 sec. tstep) was with standard settings
> regarding diffusion and vorticity dealiasing (LGRADSP=yes).
> The runs with quadratic and cubic grid both deviated from the reference in
> the following two ways: (except for a different spectral truncation,
> obviously):
> 1) LGRADSP=no (as I expect no aliasing in these cases)
> 2) RDAMP* = 200000. in NAMDYN
> as opposed to RDAMPPD = 200000., but all others (VOR,DIV,T,Q,VD) = 20. ( =
> ref.)
> Also LSPSMORO=F for the quad and cubic truncation runs, LSPSMORO=T for
> linear,
> this is the default behaviour.
>
> Typical execution times for a +66h forecast on 37 nodes on frost (36 comp. +
> 1 I/O):
> - Linear/75s: 4500 sec. (but with 1 OMP thread for historic reasons, might
> be down to ~4000
> with two threads (?))
> - Quadratic/90s : 3150 sec.
> - Cubic/90s: 3000 sec.
>
> So, quadratic is ~5% more expensive than cubic, but much cheaper than linear.
> Linear can maybe also be run with 90s timestep, not sure.
> Before deciding on operational configuration we should probably also run a
> winter
> episode, as Xiaohua suggests.
>
> Ole

Spectral grid 4 years 5 months ago #1480

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
Ole Vignes, 8/2/2016:

I have started one more experiment in the "spectral truncation" series,
this time with a "superlinear" grid. What does it mean?
On the MetCoOp domain (750x960) linear grid means (trunc=2) that NMSMAX=374,NSMAX=479
Cubic grid is a quite severe truncation (trunc=4), so that NMSMAX=187,NSMAX=239
(giving an effective spectral resolution of 5 km)
Quadratic grid has trunc=3, and thus NMSMAX=249,NSMAX=319 (which is still a quite severe
truncation, spectral resolution is 3.75 km)
In my "superlinear" grid I have chosen trunc=2.4, giving NMSMAX=311,NSMAX=399
This cuts "only" the upper 20% of the spectrum, giving a spectral resolution of 3.0 km.
The cost of the superlinear is approx. 10% higher than cubic (quadratic was 5% higher).
I don't expect problems with a 90s timestep. In fact, the number 2.4 was chosen to give
roughly equivalent stability to a linear grid with 75s timestep (I think).
I still use LGRADSP=no, aliasing might be a problem in theory, but I think not in practice.
About horisontal diffusion I have kept RDAMP{PD,T,Q} = 200000., but have set RDAMP{VOR,DIV,VD} = 2000.,
as I think the wind variables are more sensitive to noise than the mass variables (since we're following
S-L trajectories).
Will keep you informed how it goes. MetCoOp has seen quite good verification earlier both from
a 3.1km grid (linear truncation) and 90s timestep (2.5 km). With gridpoint resolution of 2.5 and
spectral 3.0 I expect (hope) to see verification scores very close to 2.5 km linear (dt=75s).
Time will show.

Ole

Spectral grid 4 years 5 months ago #1481

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
lev20_spectra.png


lev50_spectra.png


Lisa Bengtsson, 11/02/2016:

Attached is KE spectra for level 20 and level 50. At level 50 super-linear looks attractive (cubic does not), but some ailiasing at level 20 with super linear. Maybe some more horizontal diffusion is needed?
Last Edit: 4 years 5 months ago by Lisa Bengtsson.

Spectral grid 4 years 5 months ago #1482

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
Lisa Bengtsson and Ole Vignes... lots of Norwegian and Swedish discussion... but to summarize the question is what should the diffusive damping be in the "super-linear" experiment, the spectral linear diffusion is determined by the NAMDYN parameters RRDXTAU/RDAMP[x]. RRDXTAU is default 123 and not given in the default harmonie namelist:

> Default:
>
> 'RDAMPDIV' => '20.,',
> 'RDAMPPD' => '200000.,', (This large value is the same as setting it to 0)
> 'RDAMPQ' => '20.,', (actually not used, can be removed)
> 'RDAMPT' => '20.,',
> 'RDAMPVD' => '20.,',
> 'RDAMPVOR' => '20.,'

I also attach the temperature spectra at level 20 and 50, here the diffusion seem sufficient.

lev20T_spectra.png


lev50T_spectra.png

Spectral grid 4 years 5 months ago #1483

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
Ole Vignes, 15/2/2016:

Hi all,

the experiment is now finished. Verification scores can be found here:

metcoop.met.no/verif/201507_cy40_lin_quad_cubic_slin_export/

As expected, scores are somewhere between those of linear and quadratic grid.

Spectral grid 4 years 5 months ago #1484

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
From Ole Vignes and Xiaohua Yang 10/2/2016:

> Actually, I ended up running without (extra) smoothing, but was uncertain.
> I guess LSPSMORO is designed/tuned for the linear grid, therefore I was not
> certain what effect it would have. I have a feeling it does not make a lot
> of difference on the grids that does extra truncation, but it should probably
> be investigated more deeply.
>
> Ole
>
>
>
Original Message
> > From: "Xiaohua Yang" < This email address is being protected from spambots. You need JavaScript enabled to view it. >
> > To: "ole vignes" < This email address is being protected from spambots. You need JavaScript enabled to view it. >, "Andrae Ulf" < This email address is being protected from spambots. You need JavaScript enabled to view it. >
> > Cc: "Bent Hansen Sass" < This email address is being protected from spambots. You need JavaScript enabled to view it. >
> > Sent: Tuesday, 9 February, 2016 7:18:27 PM
> > Subject: RE: SV: quadratic grid: a way out?
> >
> > Wow, very interesting
> > I guess you will keep the orographic smoothing in the super-linear
> > scenario?

Spectral grid 4 years 5 months ago #1485

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
So I guess the question is where do we go from here, in regards to cycle 40? We have seen a degradation in wind-speed even with super-linear grid, but more efficient computational time.

Do we need to look further into the horizontal diffusion? The orographic smoothing?

Spectral grid 4 years 5 months ago #1486

  • Lisa Bengtsson
  • Lisa Bengtsson's Avatar
  • OFFLINE
  • Administrator
  • Posts: 60
  • Thank you received: 5
I got some questions on how to plot the spectra. I am sure there are many ways to do this, but I use the method from Jean-Daniel Gril at Meteo-France using the "Ectoplasm" program which allows you to compute the spectral energy on ellipses referenced by integer ( between 0 to NMSMAX ). Then I usually plot the output using Matlab.

These are the steps:
1) Copy the binary "ALTO" to "ECTO" (ALTO is compiled by default in HARMONIE and should be in your bin/ directory).

2) Link to your history file:
ln -s $DIR1/$FILE infile${EXP}.fa ($FILE is ICMSHHARM+00${LL})

3) ./ECTO namlist infile${EXP}.fa (See below how to construct the namlist)

4) Plot wavenumbers 2:end, I usually plot using loglog plot in Matlab, e.g:
loglog((2:lenl),LINEAR(2:end,5),'b','linewidth',2);

* namlist:

&NAMNAM
C(1)='S***TEMPE???????'
L(:,1)=24,41,50
CKE(1)='S0**WIND.?.PHYS'
LKE(:,1)=24,41,50
CVD(1)='S0**WIND.?.PHYS'
LVD(:,1)=24,41,50
&END

With this namelist KE for lev 41 ends up in column 5, that is why I select "5" in the plot command

Here is the documentation from Jean-Daniel on how to construct the namelist:
www.cnrm.meteo.fr/gmapdoc/meshtml/ecto3.0
The following user(s) said Thank You: Emily M. Gleeson, Alvaro Subias
  • Page:
  • 1
  • 2
Time to create page: 0.130 seconds