Research directions in ALADIN data assimilation: Research directions in ALADIN data assimilation Claude Fischer,
With less stolen material than in Medullin …
Slide2: Background error statistics « the Jb »
Assimilation cycles
Observations
TL/AD computations: the ALATNET study of C. Soci
Technical state of the art
General conclusions
Background error statistics: the « Jb »: Background error statistics: the « Jb » -plane (El Ouaraini and Berre)
Off-diagonal terms in B (Stefanescu and Berre)
Isotropic assumtion revisited (Bölöni)
Compactly supported correlations (Guidard)
Wavelets (Deckmyn – paper submitted)
Relocatable B (Bouttier) – Arome
Goals: more anisotropy in C+I, less wrap-around problems, better portability
Background error statistics: the « Jb »: Background error statistics: the « Jb » Mesoscale short range structure functions: lagged NMC method (Široka etal., 2003)
Sensitivity of the NMC samples with respect to forecast range and difference (Bölöni)
Ensemble of forecasts derived from Arpège (Stefanescu and Berre)
Tuning of b by a posteriori validation (Sadiki and Fischer, paper in preparation)
Lönnberg-Hollingsworth methods
Goals: find better tunings for error variances and correlation lengthscales
Blending and BlendVar assimilation cycles (Brožkova etal., 2001, Bölöni and Široka, 2001) : Blending and BlendVar assimilation cycles (Brožkova etal., 2001, Bölöni and Široka, 2001) Blending: digital filter blend for
3D fields + linear combination of
Surface fields, in order to combine
Arpège analyzed large scales with
Aladin forecast small scales Aladin 3D-VAR analysis:
J = Jb + Jo
Several Jb formulations exist:
Standard Jb (large scale)
Lagged Jb (mesoscale) LBC0 LBC1 Arp Analysis Ald 6h fct
« Large scale » cost-function Jk (Guidard): « Large scale » cost-function Jk (Guidard) J(x) = Jb(x) + Jo(x) + Jk(x), where
H1 : global LAM low resolution H2 : LAM high resolution LAM low res. V : « large scale » error covariances xAA : global analysis
Large scale update - evaluation: Large scale update - evaluation BO versus BOK: observation over all the domain LAM background BOK analysis BO analysis global analysis truth Statistically:
No difference between BO and BOK + observation
3D-VAR/Aladin in a doubly nested model, Aladin/Hungary (S. Alexandru & A. Horanyi / ALATNET PhD) : 3D-VAR/Aladin in a doubly nested model, Aladin/Hungary (S. Alexandru & A. Horanyi / ALATNET PhD) Comparisons between several 3D-VAR assimilation suites, with different coupling data (Ald/Hun, Ald/LACE, Arp), different initialization, different coupling strategies
The « Budapest solution » of a nested LAM assimilation cycle: standard NMC Jb, no blending, DFI inside DA cycle, coupling with Arpège
Case studies and monitoring scores
Observations : Observations Screen level (Ps, U10, RH2m) (Moll, Jurašek, Horvath)
High resolution « mesonet » in 3D-VAR (Auger) in 2004
Satellite radiances: ATOVS/AMSU-B (Randriamampianina, Sahlaoui), Meteosat SEVIRI (Montmerle)
Humidity boguses (Nuret, Hdiddou)
Radar reflectivities: start in 2004
TL/AD computations: sensitivity computations using the adjoint model (Cornel Soci, Alatnet PhD): TL/AD computations: sensitivity computations using the adjoint model (Cornel Soci, Alatnet PhD) Goal – improvement of 6 h precipitation forecast – in this case study: diminish CAPE to trigger the second convective nucleus ! 27 18
Sensitivity computations and modification of the initial conditions (Cornel Soci, Alatnet PhD) : Sensitivity computations and modification of the initial conditions (Cornel Soci, Alatnet PhD) Lessons:
Many problems were encountered with simplified physics at 10km resolution:
Tuning of parameters
Numerical instabilities driven by inadequate parametrizations (Kessler scheme)
Computational cost
Not all cases showed sensitivity to the initial conditions: LBC or model formulation
Convective case study: no impact of verifying analysis, though a « useful » signal was obtained from the adjoint model
Do these results bring us closer or further away from 4D-VAR ?
Technical state of the art: Technical state of the art LAMFLAG geographical preprocessing to C+I
3D and 4D screening run
3D-VAR minimization stable
Eulerian hydrostatic TL/AD models safe (?)
Desired evolutions of the system: 3D-FGAT, SL TL/AD, CONGRAD, new obs operators (radar)
Is it worth to spend time and energy on exotic configurations like NHS/TL+AD ?
General conclusions: General conclusions The push for more observations must be maintained (denser data, experience on data analysis + ODB, link with verifications)
Surface analysis: OI (CANARI), 2D-VAR
Scientific program 2002/04: Gourdon, Toulouse, Medullin discussions
The three pillars of variational assimilation: scientific goals, local 3D-VAR consolidation, maintenance of the code
Screening and 3D-VAR installed in Toulouse, Budapest, Casablanca and Prague
Aladin -> Arome and vice versa in DA
Large scale update – evaluation (2): Large scale update – evaluation (2) BO versus BOK: obs. over a part of the domain LAM background BOK analysis BO analysis global analysis truth Statistically:
BOK better than BO + observation