Short report on the work on "DIAGPACK" performed in September 1999
by Vincent Casse (Meteo France), Jure Jerman (Hydromet. Institute of Slovenia) and Gabor Radnoti (Hungarian Met. Service)
in the framework of French-Hungarian ans Slovene-Hungarian bilateral cooperations
(a complete paper version -with figures- can be asked to Gabor)
DIAGPACK (diagnostic package) is meant to be a software package that enables the user to analyse and diagnose some mesoscale features of the atmospheric boundary layer in such a way that the emphasis is not on the subsequent model forecast, but on the most accurate description of the instantaneous state of the boundary layer. The first component of such a package, that our work was fully dedicated to, is to create a version of "CANARI" objective analysis (hereafter CANARI-DIAGPACK) that produces analysed fields of boundary layer by the use of an earlier model forecast and boundary layer observations. In the forecast model these quantities are derived from the model state vector and this is the reason why a different approach is adopted for directly analyzing them fully independently from the model state and observations of the upper air. To achieve this goal some modifications of the code of analysis and some new setting of namelist parameters were required. This was designed by V. Casse in early 1999 (DIAGPACK by V. Casse, February, 1999) and, based on this, the first version of the modified CANARI code was developed by G. Radnoti in March, 1999 in Prague. This development was on the AL09/CY19T2 version of the code. The purposes of this 2 week common action were:
This work consisted of the merging of the modifications (bug corrections) that have been done in Toulouse and Budapest since the first code version. We have to mention that during this 2 weeks we were working on the AL11/CY21T1+CANARI-DIAGPACK version of the code, while in Toulouse the base cycle for this code is AL11/CY21T2. Nevertheless, from the scientific point of view these two versions are equivalent and porting the most recent version of the code to the main library AL11/CY21T2 will be straightforward. Experiments for reproducing CANARI-DATASS with the new code and proper namelist will be done later on in Toulouse in a parallel suite. The reference results are not fully reproducible because of diminishing vertical correlations between boundary layer data and the corresponding upper air both for model vs. obs and for obs vs. obs (see CACOVA and CATRMA) , the expectation is neutral or slightly improved scores for boundary layer parameters. As a first check for quasi-reproducibility we made a single observation test (one simulated TEMP observation containing boundary layer observations and upper air observations at a single pressure level). The analysis increments are realistic.
In these experiments we were using a simulated SYNOP observation in the middle of our analysis domain. We tried different combinations of T2m, Hu2m, U10m and V10m. Simultaneously we examined the impact of the characteristic lengths of correlations between obs and guess and standard deviations of guess errors for the different variables. In the new code these are namelist driven (NAMCANAPE). (Here we have to remark that the standard deviations are taken by the CANARI code as it is only in the case if the stretching parameter RCALPH in NALORI is set to zero.) To see clearly the analysis increments free of noise we also executed an analysis without any observations which eliminates from the analysis increments the impact of compacting fields and spectral fitting. From these we can conclude, that in accordance with our expectations, in case of T2m and Hu2m observation the increment fields are circular with the correct center; maximum value and breakdown; (the impact of characteristic length and standard deviation is also as expected). In the old reference CANARI-DATASS version it was not the case because of the impact of pressure difference between the points because of the impact of pressure difference between the points through the vertical correlation. In the case of wind observation we examined separately the impact of U10m and V10m such that one component was changed with respect to the guess while the other was exactly the guess value. In this case the increment field of the modified component had symmetrically a primary maximum center and additional secondary maxima in accordance with the fact that the characteristic lengths correspond to vorticity and divergence and not for the wind components themselves. The increments of the opposite components have symmetrically 4 maxima (2 with positive, 2 with negative sign). When the characteristic lengths of divergence and vorticity are the same these will merge into one circular pattern. The experiments showed these features exactly according to the expectations.
Some initial efforts were done in direction of tuning DIAGPACK. For two different situations tests were performed with different tuning of namelist parameters.
Available GTS SYNOP data together with data from Hungarian automatic station were used. Experiments were performed on Aladin/HU domain with the guesses taken from Aladin/HU model. 2m temperature fields, 10m wind fields and 2m humidity fields were analysed.
Tests were done with respect to two groups of namelist parameters, both coming from namelist NAMCANAPE:
First the impact of horizontal characteristic lengths was studied. Values used were -25000m, 50000m (reference test), 100000m for ref_a_t2, ref_a_h2, ref_a_vor10, ref_a_div10 (with ref_sigt2=3K, ref_sigh2=30%, ref_sigvl=6m/s)
Second, standard deviations of the guess were varied
Behaviour of CANARI-DIAGPACK was as expected, analyses better fit to the observations if bigger value of guess standard deviation is used.
With smaller values of guess standard deviations the differences between analyses and guess are smoothed, but the the pattern in difference fields is preserved.
In order not to introduce too much noise into the guess with bigger value of guess standard deviation or on the other hand to use the observations in proper extent it is worth to respect the rule that
(Sigma)_guess / (Sigma)_obs should remain between 1/2 and 2.
One can very easily conclude that experiments performed with smaller characteristic length better fit to the observations. But this is not what we always want, too good fit to observations might not be realistic and it may introduce some noise into the analyses. Choice of characteristic horizontal lengths depends on scale on which we want to introduce new informations into the guess.
Better fit to the observations should be achived with decreasing of horizontal characteristic lengths rather than with increasing standard deviation of the guess.
In this experiment the guess of analysis was a very good forecast. This experiment should be repeated for some less successful case, where we expect larger departure between analysis and guess.
Our time was partly dedicated to make tests with distributed memory version of CANARI. That is important if we consider a possible operational application of DIAGPACK on any distributed memory platform in the future. To do this we successfully implemented the so called obsort package that splits the observation files and creates separate observation files for the individual processors. Some bug corrections were also necessary in the basic code of CANARI that will enter the next cycle.
Comparision timings for analyses of boundary layer parameters (T2m, RH2m, wind10m) with full set of observations for shared (CANARI) and distributed memory (OBSORT + CANARI) were done. User times in seconds were: SM: 131 s, DM(1 processor): 162 s, DM(2 processors): 90 s
The results of our modifications as confirmed by the performed tests give a good chance that a stable and safe code can enter the next cycle that can be a good start for building up a diagnostic package DIAGPACK. Some further modifications are planned to make possible that boundary layer parameters have a direct impact on upper air analysis through vertical cross correlations (e.g. T2m obs on T upper air analysis) without the counterpart. This would be useful for later CAPE (convective available potential energy) computations that require an accurate description of the lower atmosphere. On the other hand it can result in better reproducibility of the CANARI results in data assimilation mode. This development will involve only some simple modification on routine CACOVA and introduction of some vertical correlation function between boundary layer observations and upper air model fields. Also the routine CATRMA will be modifyed for the safety reason with introduction of correlation function between boundary layer and upper air observations. We plan to introduce these modifications together with the rest of the code into AL11/CY21T2.
|
Home |