Simplified physics in Arpège/Aladin

Transcription

1 Simplified physics in Arpège/Aladin Olivier Rivière C.Loo, A.Dziedzic...

2 Introduction Simplified physics is the basis for developpement of linear-tangent and adjoint models used for: variational assimilation (Arpege/IFS 4D-VAR) singular vectors computations adjoint sensitivity studies (in Arpege/Aladin...) Now only dry simplified physics used operationally at Météo-France need for better representation of moist processes 2/20

3 Position of the problem Parametrization of moist processes highly nonlinear in the full physics: how to represent them with more regular simplified parametrizations? How close to the full physics must simplified parametrizations be? For practical reasons a strategy with simplified parametrizations kept independent from the full physics was made. Validity of LT hypothesis when going to higher resolutions (as in Aladin France or Arpege in the future)? 3/20

4 Existing simplified parametrizations (TL/AD) Param. Type of scheme Used in Arpege 4DVAR Vertical Diffusion K taken from trajectory Yes Large scale precip P (q qsat) No Gravity wave drag Close to old oper. version Yes (but LT hypothesis) Convection Simplified Mass-Flux Scheme No Radiation Simplified scheme No Mesospheric drag Yes but little impact Several tasks ongoing for simplified physics: New LSP scheme (done), fix of the old one (done) Improvement of GWD scheme by neglecting perturbations of some terms in the GWD formulation (done and currently in test) New convection scheme (in progress) 4/20

5 Outline of the talk Description of new simplified scheme for large-scale precipitation (LSP) Results in Arpege LT diagnostics 4D-VAR results Tests in Aladin France 5/20

6 Q v Q l Q i Q r Full and simplified moist physics in Arpege/Aladin Adjustment (ACNEBSM) Smith scheme Q v Q l Q i Q v Q l Q i Q r C FULL PHYSICS Falling Autoconversion Evaporation (ACMICRO) Melting Q l Q Q r i Collection (ADVPRC) 6/20

7 Q v Q l Q i Q r Full and simplified moist physics in Arpege/Aladin Adjustment (ACNEBSM) Smith scheme Q v Q l Q i Q v Q l Q i Q r C FULL PHYSICS Falling Autoconversion Evaporation (ACMICRO) Melting Q l Q Q r i Collection (ADVPRC) Q v Q l = Q i = 0 Q r = 0 Q P = max(0, v Q sat 1+L/C T Qsat ) Q v = Q v P Old. simp. param Q v Q l = Q i = 0 Q r = 0 C=0 or 1 6/20

8 Q v Q l Q i Q r Full and simplified moist physics in Arpege/Aladin Adjustment (ACNEBSM) Smith scheme Q v Q l Q i Q v Q l Q i Q r C FULL PHYSICS Falling Autoconversion Evaporation (ACMICRO) Melting Q l Q Q r i Collection (ADVPRC) Q v Adjustment Q v Autoconversion Q l = 0 Q i = 0 Q r = 0 Smith scheme New. simp. param Q l Q i Q r = 0 C Precipitation and removal of all condensed water Q v Q l = Q i = 0 Q P = max(0, v Q sat 1+L/C T Qsat ) Q v = Q v P Q v Q l = Q i = 0 Q r = 0 Q r = 0 Old. simp. param C=0 or 1 6/20

9 Qc condensed water Computation of condensed water using the Smith scheme Q c = Q v Q sat Inside a gridbox: q c = Q c + s with s the local variation of Q c. Statistics of s are given by a distribution function G(s) q c = R Q c (Q(c) + sg(s))ds and N = R Q c G(s)ds Probability distribution function Smith scheme Old scheme σ 1 0 -σ RHcri=f(σ) 0 σ 0 Qsat Qv 7/20

10 Overview of the new scheme Smith scheme used for computing Q C in a diagnostic way No condensed water and/or precipitation handled prognostically. Retrieval of a cloud fraction C taking any possible value between 0 and 1: scheme more suitable for linearization. Diabatic heating corresponding to precipitation of all Q c (to be improved?) Possible evolutions: Choice of a more regular probability distribution function Prognostic Q c Activation of autoconversion (coded) 8/20

11 Tuning of distribution s shape. LT diagnostics show large sensitivity to choice of σ After tuning of σ, reduction of ǫ moist in the tropics and in the midlatitudes. Old parametrization New param. RHcri oper New param tuned RHcri Error with LT model greater than without LSP Moist error / Dry error LT model with param for LSP performs better than dry LT Latitude ǫ moist /ǫ dry with ǫ = M(X 0 + δx 0 ) M(X 0 ) LδX 0 9/20

12 Comparison of linear and nonlinear evolution of increments. M(δX 0 + δx 0 ) M(X 0 ) LδX 0 old scheme A lot of lacking structures in the tropics with the LT model using the old param. for resolved precipitation 10/20

13 Comparison of linear and nonlinear evolution of increments. M(δX 0 + δx 0 ) M(X 0 ) LδX 0 new scheme Structures well positionned but with amplitude underestimated with the new LSP parametrization. 11/20

14 Tangent-linear diagnostics: impact of precipitation 1/2 RMS of ǫ = M(X 0 + δx 0 ) M(X 0 ) LδX 0 averaged over the globe (3h) 0 T107 0 T Vertical Diffusion,VD+ old LSP scheme, VD+ new LSP scheme 12/20

15 Tangent-linear diagnostics: impact of precipitation 2/2 RMS of ǫ = M(X 0 + δx 0 ) M(X 0 ) LδX 0 averaged over the globe (3h) T T399 s12 s123 MJ s123 OR Vertical Diffusion,VD+ old LSP scheme, VD+ new LSP scheme 13/20

16 Application to Arpege s 4DVAR E-suite: 20 days experiment (4/02/09->25/02/09) Same resolution than operational suite (forecast:t538,minim1:t107,minim2:t224) New simplified parametrization for LSP included in both minimizations. 14/20

17 Application to Arpege s 4DVAR: first scores EURATL Min= 0.83 Max=1.3 Moy= 0.04 NORD Min= 0.83 Max=1.3 Moy= 0.04 AMNORD Min= 1.08 Max=1.53 Moy= 0.09 ASIE TROPIQ ASIE Min= 0.8 Max=1.25 Moy= 0.19 Min= 0.8 Max=0.54 Moy= 0.13 AUS NZ SUD AUS NZ Min= 0.77 Max=1.87 Moy= 0.22 NORD Min= 0.13 Max=1.38 Moy= 0.23 TROPIQ Min= 0.8 Max=0.54 Moy= Min= 0.53 Max=1.3 Moy= 0.32 Reference: ECMWF analysis Blue: introduction of LSP improves forecast compared to operational suite Scores neutral to slightly positive New moist parametrization performs better than old one (not shown) Tests currently done at higher resolution (T105/T224/T399) 15/20

18 J When to introduce moist processes in 4D-VAR? Evolution of J during 3d minim T105/T224/T399 dry physics T105/T224/T399 3d minim with LSP T105/T224/T399 2nd and 3d minims with LSP Number of iterations Largest positive impact of inclusion of moist processes especially if introduced during two last minimizations. 16/20

19 Tests of the linearized physics in Aladin France Adjoint and LT models are also mantained in Aladin. Basic configuration of Aladin 4D VAR (not validated) is existing under Olive Tangent-linear diagnostics (Conf 501li): M(X + δx) M(X) compared to LδX δx: guess-analysis taken from Aladin-FR oper. suite. Tests performed over 3 and 6h Domain: Aladin France 17/20

20 Linearization error in Aladin over 3h with different simplified physics e501li, M(G)+L(A G) M(A), RMS T franx01, le , 00+3h e501li, M(G)+L(A G) M(A), RMS U franx01, le , 00+3h GW DV MJ SM GW DV MJ SM Pression [hpa] T Pression [hpa] U Erreur e501li, M(G)+L(A G) M(A), RMS V franx01, le , 00+3h Erreur e501li, M(G)+L(A G) M(A), RMS Q franx01, le , 00+3h Pression [hpa] V GW DV MJ SM Pression [hpa] Q GW DV MJ SM Erreur Vertical diffusion (VD),VD+old LSP,VD+new LSP,VD+GWD Important error around 300 hpa: dynamics? Erreur 18/20

21 Limits of LT hypothesis DV02.36.err.3m2.U.P , 00+6h min= max= moy= ect=1.18 DV02.36.err.2p5.0.5.U.P , 00+6h min= max= moy= ect= N 50 N 50 N 50 N N 40 N N 40 N M (X0 +δx)+m (X0 δx0 ) 2M (X0 ) 2 M (X0 + δx) M (X0 ) LδX DV02.36.valeur.1.U.P , 00+6h min= max= moy= ect= N 50 N at 300 hpa, LT approximation is no longer valid even after 3h at 10km resolution ( role of dynamics?) N 40 N U0 (300hPa) important limitation for methods based on use of LT/AD models at those resolutions 19/20

22 Conclusion New LSP simplified parametrization performs well in terms of LT approximation and slightly increases scores in 4D-VAR but more added value is expected at higher resolutions New GWD parametrization cheaper in terms of CPU and improving LT model Ongoing work on convection with a simplified Betts-Miller scheme coded and LT version being tested. At very high resolution, role of the nonlinearities to be assessed. 20/20