The SSMI 2 degree test runs is back up and running. It was started on Jan 1 1984 and completed 1984 yesterday. It will be a couple weeks before it catches up to the first SSMI observations. In the mean time we are scanning it's precipitation, temperature and heights. Most look reasonable. However, we know from previous experiments that the GEOS5 precipitation is much improved at 1/2 degrees resolution.
The system put toward validation is being finalized. We are checking and rechecking the implementation of a correction to the observational error statistics. Also, the system to be used in validation should have the entire file spec being produced. One component still being tested is the grid resolution reduction. The most voluminous data are the 3-D tendencies. I think just about every physical tendency has been included in the file spec. The 72 layer 1/2 degree data would be far too enormous for anyone to use, ( and uncomfortable to store). So, we are reducing the tendencies to 1.25x1.25 degrees and 42 pressure levels. The resolution reduction is being build inline to conserve the time doing I/O.
Recently, we found an oscillation between 6 hour analyses, with and without the radiosondes. The problem is local to high winter latitudes over land. It appears that the model has a cold surface bias, and the surface radiances are being rejected. Mid troposphere radiances still think the atmosphere is cold, and try to warm it up, but in the next analysis, the radiosondes draw the temperatures back. So, it seems to be a disagreement between a model bias (at high winter land latitudes) and differences in the radiosondes and radiances. Here, this has become known as the "sawtooth" problem, as the 500 mb time series shows a 6 hour zigzag pattern.
One possible source of the model bias is that there is no limit on the stability of the surface layer. So, Richardson numbers are very high in these conditions. In free running model, it doesn't appear to be a problem, but it decouples the land and atmosphere when observations are included. So, a limit on the Richardson number is being tested. There is also a strict criteria that limits the surface radiance increments. So that when the forecast and observation differences are large, the increment is rejected. This is strict for very good reasons (cloud clearing problems, for example). We are cautiously looking into relaxing that restriction. However, this could be risky and would only go forward if remarkable improvements can be obtained.
We also had some good news in this respect. The analysis group turned on the bias correction routines for surface temperature in GSI. This will apply a bias correction to the model background, before the surface temperatures are analyzed. So, the GSI sees a better surface temperature. These bias corrected temperaturesdo not feed back to the model forecasts (a much more difficult problem). After 15 days, the sawtooth has been corrected in many regions. East Siberia is spinning up to this a bit slower. In any event, it seems to be a reasonable fix as well.
We had not seen a problem in the system like this until recently, when running Jan 2006 as a test case. The previous winter case that was examined in detail was Jan 2001. In Jan 2006, we also include AIRS radiances. While these are thinned, AIRS still contributes 500K observations of the 1.2M observations assimilated. Away from the AIRS period, we still see the problem when we are looking for it, but it is reduced. We can also see this periodically in other operational analyses, but not nearly the degree seen in our Jan 2006 experiment.
