A large part of the ALADIN effort in Toulouse was devoted to phasing, once again, with the creation of cycles 12 and 13, with the help of Gabor Radnoti for the former, Jan Masek, Cornel Soci, Jozef Vivoda, Mark Zagar and Meriem Zitouni for the latter. A review of the corresponding changes, by Claude Fischer, is available in this Newsletter.
A new challenge : Malgorzata Szczech and Witold Owcarz managed to prepare a benchmark for ALADIN / POLAND in 4 days !
Three scientists from Tunisia : Abdelwahed Nmiri, Karim Bergaoui and Nihed Bouzouita, spent 3 months each in Toulouse to get more familiar with and work on ALADIN. Their contribution was quite useful.
30 visitors in Toulouse along these two quarters !
Karim Yessad examined thoroughly the ARPEGE and ALADIN codes to define which modifications are required to relax the thin layer hypothesis and simultaneously reintroduce some neglected Coriolis terms, following White and Bromley (1995). The hydrostatic and non-hydrostatic cases were examined, as well as the impact on DDH formulation. A detailed documentation is available.
Pierre Bénard coded and tested a new 2d formulation of the continuity equation for semi-lagrangian advection schemes, in 2d academic then real cases. This new version has a lower computational cost, is more precise and mass conserving, but less diffusive which brings some more orographic noise at long time-steps.
A wide diagnostic study of ALADIN-NH problems has been undertaken by Tamas Szabo in the simplified framework of the 2d vertical-plane version used with academic conditions. This study led to two major conclusions :
This work also allowed a clean coding of the "sponge" layer in the 2d model.
Petra Smolikova applied to real 3d situations a new attempt to damp orographic resonance problems in semi-lagrangian advection, based on the EDSA method and a new formulation of the continuity equation designed by Pierre Bénard. Though it proved quite successful in academic 2d cases, it failed on real cases, producing even more noisy fields. Quite different solutions must be found now.
Malgorzata Szczech began to code the numerous modifications required for handling of ozone as a thermodynamic spectral variable in ARPEGE and ALADIN, for use in assimilation and with the corresponding physics. Roughly speaking, only the corresponding key was available before. This work will be continued in Cracow and Toulouse along next summer and autumn.
Claude Fischer, Karim Yessad, Radmila Bubnova and Jan Masek examined the impact on non-hydrostatic dynamics of the modifications introduced at ECMWF by Deborah Salmond and Mike Cullen for the implementation predictor-corrector scheme. A preliminary phasing was realized for cycle 13 (CY22T2), but major and obviously temporary modifications were introduced just after in CY23 ...
Steluta Alexandru compared two global databases for orography : GLOB95 (operational) and GTOPT030 (more recent), at the global and the local scale, using both ARPEGE and ALADIN / ROMANIA. The second database has an effective resolution of 2'30, everywhere, and provides a better description of height and urbanisation over Europe. Changes are larger in the Southern Hemisphere or over Greenland.
Afterwards she tried to optimize the spectral orography of ALADIN / ROMANIA, choosing between the two formulations of the cost function and tuning the corresponding parameters. She proposed two alternative choices, providing a more realistic orography than the operational one, based on an old LACE tuning as for many other domains.
Andrey Bogatchev and Valery Spiridonov tried to improve the parameterization of snow cover in ARPEGE / ALADIN. They tested once again the Douville scheme, with the full soil freezing parameterization this time. Results were as deceiving as previously. They developed afterwards a new formulation of the snow cover fraction, taking into account several characteristics of vegetation: type (forest versus crops or grassland), vegetation fraction, leaf area index. Tested on ALADIN / BULGARIA on a several situations, this was shown to decrease forecast errors and their correlation with the vegetation cover. This was done for the Douville scheme then for the operational one. They will after try to further improve the albedo and introduce the impact of rain on the snow cover melting. These modifications are expected to be ready for a parallel suite next winter.
Taking into account the several attempts in ALADIN and the experience of other NWP teams who tried to implement this scheme operationally (Canada, HIRLAM), the Douville scheme is likely to be abandoned for NWP in ARPEGE / ALADIN. An alternative could be a closer cooperation with HIRLAM, especially the Spanish team, to interface the Fernandez scheme with ISBA.
Eric Bazile and Mehdi El Abed coded and tested a new formulation of evaporation over sea, proposed by J.L. Redelsperger and taking into account precipitations, on FETCH and TOGA cases. Some instabilities arise from the interaction with the previous scheme, requiring some further improvements.
Sandor Kertesz focused on the problem of lakes : test and debugging of the corresponding part of configuration E923, impact studies and design of a new parameterization, advised by Eric Bazile. A more extensive description is available in this Newsletter.
Mohamed Jidane, with the help from Eric Escalière to format initial data, evaluated over Morocco a new global database built by J.L. Champeaux et al. with a resolution of 1 km over most land surfaces. In cooperation with other scientists in Morocco, he compared the initial land-cover map to local informations. The new dataset was tested afterwards along a 2-weeks assimilation suite and the subsequent 48h-forecasts in a sunny period (April 2000). The impact on surface variables (surface and mean temperatures, superficial moisture, soil wetness index) and 2m-temperature and relative humidity was examined. The diurnal cycle is improved, with higher maximum temperatures. When averaged on the southern part of the domain, where most changes lie, differences in surface temperature may reach 2 K. A further evaluation shall be performed in Morocco, using more observations and finer diagnostics.
Roger Randriamampianina brought a significant contribution to the improvement of the radiation scheme. In 1996, Neva Pristov made some modifications in the radiation scheme in order to correct the underestimation in the prediction of the morning 2m temperature. The source / sink term (corresponding to the radiative cooling and heating) in the prognostic equation of the temperature is calculated in the physical parametrisation part of the model as a vertical radiative flux divergence. To calculate the flux divergence, the vertical fluxes have to be determined by integrating the radiative transfer equation. Four different integrations are to be done, namely the angular, vertical, spectral and optical ones.
In the source-code the short-wave and long-wave (LW) fluxes
are computed separately. The work focussed on the computation of the LW
fluxes. For this, the radiative exchanges are of special interest :
The cooling to space term is computed exactly, while the two other exchange terms are introduced as an approximate in the operational scheme. The optical depths for the exchange between a given level and the space (d¯) / surface (d) are calculated, and their minimum (min(d,d¯ )) is used when estimating the exchange between the layers in order to avoid instabilities.
The improvement of the radiation scheme in 1996 concerns the exact computation of the exchange with the surface. This modification made the scheme more expensive, and had relatively small impact on the forecast. For this reason the modification was not introduced in the operational scheme.
In order to improve the radiation scheme, the exchange
between model layers was addressed here, in two steps :
1 -To create a reference scheme, using the "emissivity method", that
takes into account
all the possible exchanges on an exact way.
2 -To modify the scheme introduced by Neva Pristov (NP scheme), to get a
better fit to the
reference one. For the "tuning" of the NP scheme we have tried both :
a- to change the gaseous absorption width to compute the best optical
depth, and
b- to find the best function for the computation of the optical depths
instead of
taking their minimum.
Solution "b" was clearly the most appropriate one.
Preliminary tests on the impact of all the modifications were performed with a 1d model. They show a real improvement, especially at the surface. Some further 1d (on the MUREX data) and 3d experiments performed afterwards confirm this, with maybe some more refinements required for the Tropics. This work will be pursued, to enter one of the next ARPEGE parallel suites.
Youssef Moudden add some refinements to the thermodynamics of ARPEGE / ALADIN. To obtain a more exact computation of the saturating vapour pressure, the hypothesis of constant specific heat for liquid water and ice was relaxed. They would depend on temperature as a second- and a first- order polynomials respectively. All the rest of thermodynamics and parts of the physics were rewritten accordingly in test mode.
Abdelwahed Nmiri tested a preliminary ALADIN / TUNISIE on a situation of severe floods over Tunisia, with encouraging results. He re-examined afterwards the physics / dynamics interface and the dependency on the time-step, using a new type of diagnostics. The previous results were confirmed and some problems in the computation of fluxes in ALADIN exhibited.
Apart from a kind and useful but absorbing help to the many visitors, the small Toulouse team in charge of physical parameterizations focused on the problems of low-level cloudiness and of the warm bias in the boundary layer noticed last winter (Jean-Marcel Piriou) and on the principle of functional boxes for the parameterization of liquid and ice water (Eric Bazile).
Radi Ajjaji, together with Elisabeth Gérard and Françoise Taillefer, debugged the last version of CANARI, ensuring that the same results are obtained whatever the value of LMESSP and the number of processors. He stressed that the size of "halo" is to be adapted to each domain.
He investigated afterwards the impact on the ALBACHIR (ALADIN / MAROC) assimilation suite of the modifications introduced for Diagpack : separation between upperair and surface analyses, definition of a statistical model for the latter, new formulation of the impact of the mapping factor, more consistent analysis of upperair relative humidity following the previous study of Mehdi El Abed. Here the aim is to minimize not only the distance between analysis and observations as for Diagpack but also the distance between short-range forecasts and observations. The skill of the new assimilation (based on AL12 + bugfix) was evaluated against the old one (AL08), dynamical adaptation from ARPEGE analysis and ARPEGE forecasts on April 1999 situations. The horizontal characteristic lengths of observations had to be increased since data are quite sparse in the southern part of the domain. For 2m-temperature, 2m-relative humidity and 10m-wind the retained tunings are (50 km, 50 km, 50 km) for Diagpack defaults, (200 km, 100 km, 200 km) for ALBACHIR assimilation, (350 km, 300 km, 50 km) for ARPEGE assimilation. The background errors were derived from operational verification statistics. The namelist parameter controlling the impact of the mapping factor (RCALPH) was tuned on 10 days long assimilation experiments with subsequent 48h-forecasts. It now lies in the range 0.10-0.14, instead of 0.28 previously. The full modset will be tested again with a parallel suite in Morocco.
Lora Gaytandjieva and Elisabeth Gérard updated and tested the snow cover analysis in CANARI in the framework of single-obs experiments. They will now try to improve the interpolation of first guess fields at observation points, optimize the use of climatological surface fields, introduce a dependency on height in the statistical model and define new tunings (defaults are for an unstretched T63 with the old climatologies). These improvements will first be tested in the ARPEGE assimilation suite (to provide better initial fields to ALADIN).
Cornel Soci pursued the investigation of the impact of lateral boundary conditions in sensitivity studies. The adjoint of ALADIN model has been used to study the sensitivity of forecast errors to initial and lateral boundary conditions. He has tried to find out how the boundary data gridpoint perturbations act, which type of data are more responsible for the forecast failure and if ALADIN is able to produce a posteriori good forecast as ARPEGE did (Hello et. al, 1998).
The gradients of the forecast errors with respect to the lateral boundary data have been used to modify the original lateral boundary conditions by adding a quantity dx = -a * grad(x), a being a scaling factor. A similar technique for deriving modified initial conditions have been used.
It was revealed that the impact of the gridpoint lateral boundary data perturbation in the forecast of a rapid cyclogenesis was very marginal. In addition, artificial noisy gravity waves in the coupling area were created. Their amplitude and existing time period during the sensitivity run was closely related to the technique used for the injection of perturbations. It was shown that if the perturbation is added only in the very first coupling file, the noise is quickly disappeared while it is very persistent if each coupling file is perturbed.
In this study the role of the initial conditions was clearly shown while the impact of the lateral boundary conditions could not be clearly stated. However, we should notice that to create a good forecast, ALADIN needs at least better initial conditions supplied by the global model ARPEGE.
The work on 3d-var concentrated on a better specification of the general framework for variational analysis in ALADIN. Mostly, 4 points were addressed in Toulouse:
1. Scripts :
It is very important to work with a stable and coherent set
of scripts. Thanks to Elisabeth Gérard and Wafaa Sadiki, we have now 2
basic scripts to test and calibrate 3d-var experiments:
- a single-obs script, which contains some short-cuts to avoid stupid
behaviours and adapt the 3d-var to this very particular type of experiments.
- a full-obs script, including the selection of a domain-dependent window for
observation handling, screening and cmafoc operations, minimization and
surface analysis. This script mimics an ARPEGE-type of operational script.
2. Geometry :
Several tests were carried out by Claude Fischer and Adam Dziedzic to deal with the purely mathematical E-zone. Tests included a weak-constraint formulation, a strong-constraint formulation (with s b's projected on an interior domain) and the rejection of the observation in the I-zone. The strong constraint formulation seems to be well suited, and Adam has shown that a fairly broad rim zone is needed to damp the background error variances (typically : 200 km width with mesoscale error statistics). The projection damps efficiently the analysis increments throughout the E-zone, at the expense of a small Gibbs wave production.
3. Impact of initialization :
A preliminary work was started by Adam to evaluate the impact of digital filter initialization (DFI) on the analysis increments. Adam showed that the DFI indeed need quite an in-depth recalibration in order not to spoil the analysis increments. This recalibration will probably also depend on the nature of the first-guess and of the background error statistics used in the 3d-var. This problem raises questions that will give a fair amount of work for the next months.
4. Calibration inside the 3d-var :
Wafaa Sadiki has assessed the impact of the mesoscale statistics, as defined and computed first by Maria Siroka in Prague. Her results show that the mesoscale Jb is a good candidate for small scale analysis (in the sense of the model resolution, not in a Diagpack type of thinking !). In addition, Wafaa did a thorough calibration of the observation to background cost function ratio, by 2 independent methods. She also showed that one of these methods might be useful in future to assess the internal coherence of the ALADIN 3d-var system.
Karim Bergaoui and Jean-Daniel Gril worked on the blending of spectral fields. First Jean-Daniel transformed the Prague script into a very clear, portable and modular one, available on demand, while Karim coded the required modifications to combine spectral fields using the ALADIN library. Some experiments were performed afterwards with ALADIN / FRANCE, with parameters abruptly derived from the LACE experiments (intermediate resolution, filter, ...). Results are still being examined because of unexpected noise in the longer ranges of forecast.
Some improvements were brought by Dominique Giard to digital filtering, more precisely to the launching or finalisation procedure, both to enable its use for ALADIN (date handling) and introduce in 4d-var an additional, semi-internal, incremental filtering, at highest resolution, with a negligible cost.
Jaouad Boutahar introduced on demand bilinear (4-points) and nearest point interpolations in Full-Pos. The first one may be more suitable for surface fields, the second one will enable an easier debugging.
He addressed also the problem of lakes / islands created in the target grid. It is now possible to initialize the corresponding surface variables with the climatology instead of using the surrounding land / sea values as previously. This may be extended to all identified lake points and assumes that clim files have been corrected using realistic / observed values.
While waiting for the corresponding cycle, an external tool for this latter task is available by Jean-Daniel Gril.
Witold Owcarz introduced a new method for post-processing boundary-layer fields, following ideas presented by other NWP teams at the last EWGLAM meeting. Fields are recomputed on the target grid from the climatological surface characteristics and the interpolated surface and upperair variables, rather than "directly" interpolated. The impact of thsi new procedure, as well as that of 4-points interpolation, will be tested this autumn.
Nihed Bouzouita coded the computation of moisture convergence in Full-Pos, with several options available. Results are consistent with those obtained with the computation within the physics package.
To combine these developments and the introduction of the large set of new fields (documented in this Newsletter), Ryad El Khatib had to deeply modify Full-Pos, once again. This was an opportunity to remind that the interpolation of the snow cover changed since AL12, when a control of consistency with the surface temperature was introduced.
Andrey Bogatchev examined the impact of moving from the shared memory (SM) to the distributed memory (DM) code on the operational version of ALADIN / BULGARIA, running on a SUN workstation. ECMWF intends to suppress the SM code as soon as possible. He tested 3 issues :
Jean-Marc Audoin performed similar portability tests on the French workstation, focusing on Full-Pos. His results are described in a previous section. He also tested the last cycles for CANARI with Lora Gaytandjieva and for 923 with Steluta Alexandru, together with the corresponding tools.
Jean-Marcel Piriou developed tools to create and handle nice forecasted cloudiness pictures, taking into account SST and land-cover types, and very similar to real satellite images. They were built around ARPEGE but can easily be adapted to ALADIN. An illustration is provided here.
The new definition of the ALADIN geometry, after the rewriting of EGGX, was tested on a large range of domains by Jean-Daniel Gril. A short description of the changes and a presentation of the new interactive domain maker (MAKDO) can be found in this Newsletter.
Eric Bazile and Peter Bechtold (Laboratoire d'Aérologie, Toulouse) focused on the dramatic 12-13 November 1999 flash flood in Southern France (Lézignan case). They used respectively ALADIN and MESO-NH with the same resolution (10 km, 31 vertical levels) and coupling. They realized a quite exhaustive study, testing the impact on the forecast of several features, such as : coupling files, update of SST, convection scheme, advection scheme, orography description, ... taken separately or not. Here MESO-NH proved better than ALADIN, with a better location of heavy precipitations. This corresponds to a better description of wind fields over the Western Mediterranean sea. Let us recall that, operationally, ARPEGE gave good forecasts, far better than IFS, but that ALADIN / FRANCE did not bring any further clear improvement.
Ryad El Khatib made some experiments with ALADIN / FRANCE on the first Christmas storm, testing several aspects of dynamics : time-step, 2-time-level versus 3-time-level, hydrostatic versus non-hydrostatic, and of the coupling frequency (1h instead of 3h). Only the latter change led to a real improvement of forecasts.
He also adapted E923 to import the orography of a quadratic grid onto a linear grid (with the same gridpoints) and performed preliminary sensitivity studies.
|
Home |