Page 1 of 1

How to improve convection triggering?

Posted: Mon Jul 24, 2017 7:51 am
by meteoadriatic
Does anybody know what can improve model sensitivity in triggering convection? Aside from trying different physics combos (none seems to work as it should). The problem is that summertime convection in unstable conditions is not triggered in model very well... in reality convection is much more widespread than model forecasts. And GFS is much better really.

Any ideas?

Re: How to improve convection triggering?

Posted: Tue Jul 25, 2017 1:01 pm
by norulz
Unfortunately no ideas...

I share your feeling for the problem. Last night I, once more, read many of the existing studies into use of the various convection schemes for WRF but found no clear guidance on how to improve convection triggering. In the East Med the problem with convection triggering is less in summer unstable conditions but rather more in the relatively warm humid winters. I concur that GFS is often more accuarate.

My production runs are using 12km grid with 4km nested domains, so it appears that I am in "no-man's-land" as far as convection is concerned.

Re: How to improve convection triggering?

Posted: Tue Jul 25, 2017 1:17 pm
by alfe
As far as I know from my trials, at least 60 levels are required and Kain Fritsch multiscale (CU = 11). Better but not perfect.

Re: How to improve convection triggering?

Posted: Tue Jul 25, 2017 2:40 pm
by meteoadriatic
I even tried 80 levels for test if I remember correctly and it was not much different than "standard" amount of ~45. However, not only because there is more levels to slow model down but more levels require smaller time step to keep model stable so this reduces it's speed even further.

What I somehow feel is that main problem is not somewhere within Cu scheme used (which would logically be first guess) but rather either within PBL or SFCLAY scheme, or just model dynamics as they are. Why do I think that? There are multiple reaasons:
    • The first thing is that diffusion in model seems to affect convection triggering. My knowledge about that is limited but there are two types of diffusion in WRF that are represented in different ways. One is horizontal and other is vertical. Horizontal difussion is basically controlled by WRF dynamics, and particularly namelist diff_opt.
    • There are three settings for diff_opt, either default and simpler solution (1) or more advanced and probably better in areas with high terrain slopes (2), but both of them (in my opinion) should REDUCE probability of triggering convection. Why I think that?
    • Years ago we had problems in NMM that were developing convection without much reason. Basically we had pretty much opposite situation than we have now in ARW. The fix that worked excellent to "damp" these false convection triggers was to actually INCREASE diffusion. So this leads to conclusion that now we might want to decrease diffusion and then model might trigger convection easier. However, setting diff_opt=0 does not help much (if at all).
    • But there is also vertical difussion and this one is controlled in PBL scheme and I don't think there is any PBL scheme that can change this via namelist so there is not much run-time control there.
    • I tried to run 10km grid distance WITHOUT PBL scheme (bl_pbl_physics=0). I know this does not make sense because PBL is pretty much required for all grid distances >~1km to have reasonable good simulation, BUT here the result of experiment is I believe important, because when I turn off PBL scheme on such simulation there is absolutely no convection within domain at all. This strongly suggests that convection triggering is caused by processes within PBL scheme, at least when we talk about mesoscale grid distances.
    • Further, as PBL scheme controls mentioned vertical diffusion, there might be a chance that this has something to do with it... not sure though.
    • And lastly, PBL and SFCLAY schemes are very closely related to each other, so part of problem (or whole problem) might actually be within SFCLAY scheme although I still don't think it is.
Unfortunatelly the source code of PBL schemes are pretty criptic and not well commented so it is hard to know how they work.

Let me know if anything gets on your mind, we might brainstorm something useful if we share ideas, no matter how "stupid" they seems to be at the moment.

Re: How to improve convection triggering?

Posted: Tue Jul 25, 2017 3:06 pm
by norulz
I'm not as intimate as you are with the internals of the schemes and the interactions between them but I will be playing around with them and will let you know if I think anything might be interesting.

Re: How to improve convection triggering?

Posted: Fri Oct 27, 2017 7:59 pm
by norulz

I have spent some time trying some options to see the effect on the triggering of convection in the model and to the extent of convection shown.
I run a 12km outer domain with one 4 km nest and did the comparisons on a specific date (25 October 2015) when there was a very active Red Sea Trough with a cold Siberian trough at 500mb. The GFS for this morning showed very high values of CAPE - up to around 3000 kj.
The actual weather was extremely stormy with several supercells moving rapidly from Alexandria in Egypt to Beirut in Lebanon and scraping the Israeli coast.

I tried the following combinations:

1. MSKF cu_physics = 11,11, cudt = 5, YSU pbl bl_pbl_physics = 1
This was the baseline

2. MSKF cu_physics = 11,11, cudt = 2, YSU pbl bl_pbl_physics = 1
I found that the more frequent calls to cu did not show any effect

3. GRELL cu_physics = 5, pbl = 1
This was not as good as the baseline and showed less triggering and extent of convection

4. KF cu = 1,1 kfeta trigger = 1, pbl =1
Not as good as the baseline

5 KF cu = 1,1 kfeta trigger =2, pbl =1
Much improved. Showed earlier triggering and better convection coverage than the baseline

6 KF cu = 1,0, kfeta trigger = 2, pbl =1
This was even better. It had earlier triggering and gave coverage of convection which was closest to the real time radar images

The combination of the regular KF on the 12km domain and no scheme for the 4km sub domain worked best on this particular test data.
The kfeta_trigger option 2 does indeed seem to have affect in the warmer Mediterranean climate.

Unfortunately there appears to be no way to adjust triggering for MSKF

It needs to be said that all the above combinations identified several large supercells with gusts of up to 100kts and sustained winds of above 70kts. These indeed occurred whereas GFS gave no indication of this and only showed 30 kts of gradient wind at 10m for the Red Sea Trough

Re: How to improve convection triggering?

Posted: Thu Nov 02, 2017 11:21 am
by norulz
I spent some more time on this issue after poor rain forecasts and managed to obtain some improvement with the following configuration of a 12km/4km domain set.

I set cu_physics = 1,1 to force the cumulus module on the high resolution sub domain in spite of the uems warning. The 4km sub domain is not simulating convection without the help of a parameterization module.

I found that kfeta trigger = 3 triggered more convection that options 1 and 2.


Re: How to improve convection triggering?

Posted: Sun Nov 19, 2017 1:24 pm
by alfe
Hello norulz,
I had a look into the cumulus physics routine code (WRF 3.8.1) and here is what I have seen :

There are 2 modules that we are interested in : module_cu_kfeta.F and module_cu_mskf.F.

Both routines have the same argument called "trigger" when called. Both are testing if trigger = 2 and do the same treatment. So I assume that MSKF cu physics is treating the case kfeta = 2 in the same way as KF.
kfeta = 3 is treated only by KF. No code for testing trigger = 3 in MSKF.


Re: How to improve convection triggering?

Posted: Sun Nov 19, 2017 2:14 pm
by norulz
Hi alfe,

You are correct. There is no possible selection of trigger options for MSKF which is a pity at it appears to not trigger when it should.
Rainy season is starting here in the East Med and I'm closely monitoring forecast performance with the regular KF module using trigger option 3.

Interesting point is that the 4km domain is not resolving convection on it's own and it needs the parameterization module (KF) to get convection to kick in. This even though KF is designed to be used with grids > 10km.

In general I'm disappointed with MSKF as it under estimates precipitation.

Re: How to improve convection triggering?

Posted: Sun Nov 19, 2017 2:26 pm
by alfe
@norulz :
are you tracking the "Medicane" case ? If yes it would be interesting if you could share what you have got as a comparison between WRF and observation !
For me, it is too far to the east to track it...