FAQ 1 What is the difference between gfs and fnl in GDAS scripts
The difference between 'gdas' and 'gfs' is actually in the
analysis. In operations, we create 2 analyses per cycle. The first
analysis is run shortly after the synoptic time (i believe it's 1.5 hours right
now), and is used as the analysis for our operational, 15-day GFS
forecast. However, a later analysis is run at 3 hours past synoptic time
(when more data has trickled in), where a shorter (GDAS), 9-hr forecast is
run. This 9-hr forecast is presumably of slightly better quality, as its
analysis contains more observational information. This 9-hr GDAS forecast
is then used as the subsequent background for the next cycle's gfs AND gdas
So for any given cycle (00, 06, 12, & 18) we have two analyses (gfs and gdas), as well as two sets of forecasts (gfs, out to 15 days, and gdas out to 9 hours) from their respective analyses. (April 2008 Daryl Kleist)
What is the plan for incorporating RTTOVS to CRTM
Originally, option to select either OPTRAN, RTTOVS, or SARTA in CRTM was planned. However, the plan seems to be changed. OPTRAN, RTTOVS, SARTA and other RTM have their own strength and weakness, CRTM teams are working on combining good parts of each RTM model into CRTM. They are hoping to have some experimental version in eight month time but it will take another half year or so to become operational. Their goal is to improve CRTM for operational data assimilation.
We thought if there is an option in CRTM, we can take advantage of it and use different RTM for simulation and assimilation. It seems if you wish to use RTTOVS or SARTA to simulate observational error you may have to use existing RTTOVS and SARTA without CRTM.
(May 2008 David Groff)
FAQ 3 Moisture constrain in GSI
Basically, for various reasons (model problems, spectral to grid transforms, observation operators), unphysical values for moisture exist in the model analysis and forecast files (not the post-processed output, since these values get fixed upon processing). In the GSI, we put a weak constraint on these unphysical values (either negative or supersaturated gridpoints) to attempt to keep things from spinning out of control. It is a rather ill-conditioned contribution to the penalty function, and in an effort to not throw off the minimization, we tend to be pretty conservative about its usage.
The two subroutines for the step size and gradient calculation are stplimq and intlimq respectively. Basically, there are two penalty terms added, one each for the negative and supersaturated gridpoints. This is a highly-nonlinear term, and operates on the full analysis during each iteration (i.e. the background + increment.....and not simply the incremental contribution). There are weighting parameters set via the GSI namelist to control how much to attempt to remove these unphysical values (factqmin & factqmax). However, these parameters are highly sensitive to the grid resolution and background (i.e. number of unphysical points and their magnitude). The current parameters we use have been tested and tuned for the current operational GFS resolution and typical background.
I'm not sure how this all ties together with the Antarctic problem, except that part of this calculation involves the saturation specific humidity. I'm guessing that once the near surface temperatures got unrealistically cold (where there is still measurable moisture...near the surface and all)....this calculation started to go haywire. The calculation of qsat can be found in genqsat.f90.
If you need more information, let me know.
FAQ Question 4
Is there any test program for CRTM rel 1.0?
David Groff added some comments and update to older version to help users. Proper tested version will be available leter from CRTM team.
Run_Forward.Rel1_0.txt is posted at this directory.