1. Radi AJJAJI : "Incrementality deficiency in ARPEGE 4d-var assimilation system"
It has been shown that the bad behaviour of 4d-var system, in some situations, linked to the incremental formulation (the so-called incremental deficiencies) could be encountered not only over desert regions but on Europe as well. The April 24th analysis produced a great negative humidity increment over Pyrenean region at 18 h UTC, and a reasonable one 6 hours after. This situation was a good individual case to try to characterize the context of this abnormal 4d-var property by diagnosing a list of ingredients responsible of such fact.
First, and as usual in all analysis studies, observations used along the 6 hours window were fetched to identify the type which is more likely responsible of underestimating the specific humidity increment. This first study showed that most part of the impact was related to AMDAR and SYNOP data. SATEM, SATOB and all other observation types showed just a little impact.
A geographic zoom over Pyrenean region showed absence of TEMP observation type in all the circle of 300 km radius centred on (43°N, -3°W). This means no humidity information since all other observation types don't inform on humidity. This could have certainly an impact on humidity analysis in sense that humidity analysis will be influenced by 4d-var multi-variate characteristic implied at least by the adjoint of tangent linear integration and observation operators. When forcing Rv to Rd during the integration of TL/AD models and observation operators, the multi-variate effect especially on humidity field become negligible as it can be seen on these approximate equations.
When putting Rv=Rd only before executing observation operators with their direct TL and tangent formulations (TL/AD of forecast model being integrated without this forcing), the impact steel unchanged. This means that observation operators are not responsible of the bad analysis increments. But as Rv=Rd is applied for all the components of a simulation, some other negative impacts could appear at least caused by assuming Rv=Rd into observation operators.
Other experiments performed with slightly modified TESTLI (a procedure testing the tangent linear hypothesis) showed a neutral impact of horizontal diffusion both in Eulerian and semi-lagrangian type integration. TESTLI showed also no difference between finite differences and Tangent Linear model integration when transporting in time a humidity increment.
This indicates that the anomaly resides in the adjoint integration which involves multi-variate effect as all great magnitude terms computed during the direct TL influence humidity during Adjoint reverse integration when there is no other information (observation) to compensate this fact.
An experiment done in the same environment but with an artificial TEMP observation added in the vicinity of the affected Pyrenean region (the TEMP is elaborated from interpolation of guess humidity on a vertical profile) corrected entirely the bad humidity analysis. This enforced the doubts on 4d-var performances over regions with poor observation network.
A diagnostic on the gradient of cost function with respect to specific humidity showed that the bad humidity increments are not produced at the first iteration as it was expected (communication with F. Bouttier) , but they are a result of an accumulation along the iterations : each forward and backward (simulation) execution adds a contribution to the increments that increase them in the same sense.
So, putting Rv=Rd as a first (non scientific) solution reduces the multi-variate effect, which is among the strong good characteristics of a 4d-var system, on humidity ; but its non seen effects could be more "dangerous". One can test other solutions as :
As a further step, points 4 and 5 will be tested, and their results will be communicated in the next newsletter.
2. Jean-Marcel PIRIOU : "Correction of compensating errors in physical packages; validation with special emphasis on cloudiness representation"
EUROCS Stratocumulus Case: (i) it has been shown on this case that the top of PBL entrainment is too big if the turbulence scheme uses saturation humidity functions inside the shallow convection functions (option GCCSV=1.), this led to an operational change (now GCCSV=0.); (ii) The Single Column Model (SCM) interface to this EUROCS case, initially coded by Blazenka Vukelic, has been introduced in the main SCM library.
Turbulence: (i) reading articles ; (ii) introducing an interactive mixing length in the ALADIN turbulence scheme. This reduces the temperatures biases of the mid and high latitudes. The impact on winter cyclogenesis is under study.
Deep convection: the EUROCS idealized humidity case, which compares Single Column Models to Cloud Resolving models with respect to deep convective clouds entrainment, has been run with the ALADIN SCM. The results of the present scheme are pretty good, which tends to show that the job done in ACCVIMP about "undilute plumes", following an idea from Steve Derbyshire, has improved the sensitivity of the model with respect to free atmosphere relative humidity.
CLOUDNET project: this European project (CLOUD NETwork) intends to validate operational models versus observed radar and lidar data. The aim is to improve the representation of clouds, cloudiness and microphysics inside operational LAMs and GCMs. The Reading meeting gave an opportunity to discuss "which fields" and "at which time scales" to provide and inter-validate. ALADIN predictions should be sent to an European common data base by the end of 2002.
3. Wafaa SADIKI : "A posteriori verification of analysis and assimilation algorithms and study of the statistical properties of the adjoint solutions"
Nothing new (implementation of the new operational suite and 3d-var in Casablanca).
4. Filip VANA : "The dynamical and physical control of kinetic energy spectra in a NWP spectral semi-lagrangian model"
Nothing new (LACE duties).
| Newsletter 21 & 4 |  |
Home | ALATNET PhD and Post-Doc Studies |