m. baldauf deutscher wetterdienst, offenbach

54
M. Baldauf Deutscher Wetterdienst, Offenbach COSMO-General Meeting 05.09.2011, Rome Development of a new fast waves solver for the Runge-Kutta scheme

Upload: fleur

Post on 07-Feb-2016

55 views

Category:

Documents


0 download

DESCRIPTION

Development of a new fast waves solver for the Runge-Kutta scheme. M. Baldauf Deutscher Wetterdienst, Offenbach. COSMO-General Meeting 05.09.2011, Rome. Motivation. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: M. Baldauf Deutscher Wetterdienst, Offenbach

M. BaldaufDeutscher Wetterdienst, Offenbach

COSMO-General Meeting05.09.2011, Rome

Development of a new fast waves solverfor the Runge-Kutta scheme

Page 2: M. Baldauf Deutscher Wetterdienst, Offenbach

Motivation

• Internal DWD project 'COSMO-DE L65' (increase # of vertical levels 50 65):(goal: better representation of convection and its initiation)But several numerical problems occur (lower BC for w, vertical advection scheme, tracer advection scheme, ...)Speculation: discretizations of fast processes not yet entirely consistent

• DWD-'Extramurale Forschung' project 'Numerische Raumdiskretisierungsverfahren hoher Ordnung für das COSMO Modell' (A. Will/J. Ogaja)Prerequisite: use of only centered finite difference and centered averaging operators

Page 3: M. Baldauf Deutscher Wetterdienst, Offenbach

D = div v

‚Fast waves‘ processes (p'T'-dynamics):

fu, fv, ... denote advection, Coriolis force and all physical parameterizations

sound buoyancy artificialdivergence damping

integration procedure: horizontally forward-backward, vertically implicit

Page 4: M. Baldauf Deutscher Wetterdienst, Offenbach

Integration procedure:

horizontally: forward-backward (Mesinger (1977) Contr. Phys. Atm.)vertically: implicit(Klemp, Wilhelmson (1978) MWR,Wicker, Skamarock (1998, 2002) MWR)

With the possibility of a 3D divergence damping, stability of the whole RK3-scheme (p'T'-dynamics) was shown in Baldauf (2010) MWR

Page 5: M. Baldauf Deutscher Wetterdienst, Offenbach

1. Improvement of the vertical discretization

Averages from half levels to main level:

Averages from main levels to half level with appropriate weightings (!):

centered differences (2nd order if used for half levels to main level)

G. Zängl could show the advantages of weighted averages in the explicit parts of the fast waves solver.New: application to all vertical operations (also the implicit ones)

Page 6: M. Baldauf Deutscher Wetterdienst, Offenbach

2. 'Strong conservation form' of the divergence operator

Divergence operator used up to now:

Strong conservation form:

Discretization of metric terms

more compact expressions in the strong conservation form

Doms, Schättler (2002) COSMO Sci. Doc. (I), Prusa, Smolarkiewicz (2003) JCP

~ d/dt proper BC

Page 7: M. Baldauf Deutscher Wetterdienst, Offenbach

3. Isotropic divergence damping

Gassmann, Herzog (2007) MWR recommend the use of the complete (=isotropic) 3D divergence damping more realistic dispersion relation for sound and gravity waves

The following tests comparethe current FW solver (fast_waves_rk.f90, version COSMO 4.18)with the new FW solver (fast_waves_sc.f90)

Page 8: M. Baldauf Deutscher Wetterdienst, Offenbach

3.a Linear flow over mountains

FW new

test case definition in: Schär et al. (2002) MWRLinear analytic solution (black contours) : Baldauf (2008) COSMO-Newsl.

current FW

U0=10 m/s, N=0.01 1/s, hmax=25 m

Page 9: M. Baldauf Deutscher Wetterdienst, Offenbach

3.b Nonlinear flow over mountains

FW new

U0=10 m/s, N=0.01 1/s, hmax=750 m

current FW

Page 10: M. Baldauf Deutscher Wetterdienst, Offenbach

3.b Nonlinear flow over mountains

FW new

U0=10 m/s, N=0.01 1/s, hmax=900 m

model abort after 6 h

current FW

Page 11: M. Baldauf Deutscher Wetterdienst, Offenbach

Results of idealised test cases see: COSMO-user Seminar, March 2011

SRNWP-workshop Bad Orb, May 2011

• Sound wave expansion• Linear Gravity wave in a channel (Skamarock, Klemp, 1994)• Linear flow over mountains (compare with analytic solution)• Non-linear flow over a mountain• mountain in a steady atmosphere• moist warm bubble test (Weisman, Klemp, 1982)• dry cold bubble (Straka et al.,1993)

All these idealised tests are simulated with either similar accuracyor slightly better.

Page 12: M. Baldauf Deutscher Wetterdienst, Offenbach

Real simulations ...

... at the beginning caused several problems,

mostly in connection with the divergence damping near the bottom.

Page 13: M. Baldauf Deutscher Wetterdienst, Offenbach

Quasi-3D - divergence damping in terrain following coordinates

Stability criterium:

in particular near the bottom (x, y >> z) a strong reduction of div is necessary!

This violates the requirement of not too small div in the Runge-Kutta-time splitting scheme (xkd~0.1 (Wicker, Skamarock, 2002),in Baldauf (2010) MWR even xkd~0.3 is recommended).

Otherwise divergence damping is calculated as an additive tendency (no operator splitting) a certain 'weakening' of the above stability criterium is possible

Page 14: M. Baldauf Deutscher Wetterdienst, Offenbach

… additionally necessary for stability:

• don't apply divergence damping in the first small time step!(J. Dudhia, 1993 (?))

• boundary treatment in the term (d/dt) / of the divergence in ‚strong conservation‘ form:the obvious d/dt|ke+1=0 leads to a divergence which is (near the ground) about a factor 10 larger than a numerical approximation of this term.

Page 15: M. Baldauf Deutscher Wetterdienst, Offenbach

Boundary condition for the Euler equations

= free slip condition at the bottom (and at the top)

1.) Extrapolation of u, v to the ground (ke+1/2):this improved the pressure bias in COSMO-EU Options now: extrapolation 0th order (= 'true' free slip BC),1st or 2nd order.

2.) Discretisation of dh/dx, dh/dy:up to now: analogous to the order of the advection operator (upwind 5th order)now: centred diff. 4th order necessary

ke

ke+1/2

ke-1/2

w

w

u

u

Page 16: M. Baldauf Deutscher Wetterdienst, Offenbach

'COSMO-EU, 05.01.2011, 0 UTC', PMSL after 78 h (Exp. 8230)

during the simulation a negative pressure bias of about -0.5 hPa/3d develops(remark: the operational COSMO-EU was nearly bias free in Jan. 2011)

Exp. 8230 RoutineDiff.

Page 17: M. Baldauf Deutscher Wetterdienst, Offenbach

pressure bias in COSMO-EU (12.01.2011, 0 UTC run, stand alone)

FW_current

FW_new

Page 18: M. Baldauf Deutscher Wetterdienst, Offenbach

BC for w: extrapolation of u, v to the lower half level (=surface) does not have any influence to the mean surface pressure

FW_current

FW_new0., 1., 2. orderExtrapolation

pressure bias in COSMO-EU (12.01.2011, 0 UTC run, stand alone)

Page 19: M. Baldauf Deutscher Wetterdienst, Offenbach

'dynamical bottom BC for p' helps in producing the correct surface pressure (at least the mean value) in the current FW solver

FW_current

FW_new

FW_currentldyn_bbc=F

pressure bias in COSMO-EU (12.01.2011, 0 UTC run, stand alone)

Page 20: M. Baldauf Deutscher Wetterdienst, Offenbach

Boundary treatment of p'/ (or of p'/z) at the lower boundary:

1.) one sided finite difference (G. Zängl):

p'/ = 0 p'(ke) + 1 p'(ke-1) + 2 p'(ke-2)

2.) ‚dynamical bottom BC‘ (A. Gassmann, 2004, COSMO-Newsl.)From

and

derive a condition for p'/ .

Page 21: M. Baldauf Deutscher Wetterdienst, Offenbach

'dynamical bottom BC for p' also improves pressure bias in the new fast waves - solver

FW_new with ldyn_bbc=T

FW_newwith ldyn_bbc=F

pressure bias in COSMO-EU (12.01.2011, 0 UTC run, stand alone)

Page 22: M. Baldauf Deutscher Wetterdienst, Offenbach

influence of the FW-solver to 'indirect' variables like precipitation is quite smallexample: COSMO-DE, 12.01.11, 0 UTC

FW new FW currentDiff.

Page 23: M. Baldauf Deutscher Wetterdienst, Offenbach

Diff.

influence of the FW-solver to 'indirect' variables like precipitation is quite smallexample: COSMO-DE, 28.04.11, 0 UTC

FW new FW current

Page 24: M. Baldauf Deutscher Wetterdienst, Offenbach

COSMO-DE, 12.01.2011, 0 UTC, pmsl

Page 25: M. Baldauf Deutscher Wetterdienst, Offenbach

very strong inversion in an Alpine valley

COSMO-DE, 12.01.2011, 0 UTC run, ldyn_bbc=.TRUE.

current FW solver new FW solver

Page 26: M. Baldauf Deutscher Wetterdienst, Offenbach

COSMO-DE, 12.01.2011, 0 UTC run, ldyn_bbc=.FALSE.

current FW solver new FW solver

Page 27: M. Baldauf Deutscher Wetterdienst, Offenbach

Which benefits are currently visible ?

The original intention to develop the new fast wave solverwas to produce more consistent dynamic fields.Indeed in some situations the stability seems to be slightly higher, examples:

• a model crash of COSMO-2 at 16.06.2011, 0 UTC runs stable with the new FW solver

• a model crash of COSMO-DE at 12. July 2011, 6 UTC run couldbe repaired by the use of Bott2_Strang, but also alternatively by theuse of the new FW solver

• a simulation with high resolution of 0.01° only runs stable with thenew FW solver

• the new solver also runs without crash during 01.-31. Jan. 2011in both a COSMO-EU and a COSMO-DE setup

But: of course the fundamental difficulty of split explicit schemes with steep orography remain

Page 28: M. Baldauf Deutscher Wetterdienst, Offenbach

crash with the operational COSMO-DE at 12. July 2011, 6 UTC

unrealistic high value of qr in one grid box;strongly deformed wind field

model crash with thecurrent FW

stable simulation with the new FW

Page 29: M. Baldauf Deutscher Wetterdienst, Offenbach

from Axel Seifert (DWD)

simulated radar reflectivityCOSMO-run with a resolutionof 0.01° (~ 1.1km)1700 * 1700 grid points

model crash after 10 time steps with the current fast waves solver

stable simulation with the new FW

Page 30: M. Baldauf Deutscher Wetterdienst, Offenbach

Efficiency:

on NEC - SX9: new fast_waves_sc: reaches ~20 GFlops and needs about 30% more computation time than the current fast waves solver (~18 GFlops)

a COSMO-EU run needs about 5% more time

but it is not yet optimized for Intel processors (cache based): it takes about 80% more computing time (reason?) higher computation time can be expected due to- more vertical weightings - exchange of p' and div v

Page 31: M. Baldauf Deutscher Wetterdienst, Offenbach

Summary

New fast waves solver with• improved vertical discretizations• strong conservation form of divergence• (optional: 3D isotropic divergence damping)

• Idealised test cases (stationary/unstationary, linear/nonlinear, with/without orography) are simulated with either similar accuracyor slightly better

• runs stable in all inspected cases (COSMO-EU, COSMO-DE duringthe whole Jan 2011; and in selected cases)

• for both FW solvers holds: dynamical bottom boundary condition necessary to reduce pressure bias in COSMO-EU, but can have detrimental influence on COSMO-DE

• satisfying optimization for NEC SX9 achieved

Page 32: M. Baldauf Deutscher Wetterdienst, Offenbach

Outlook• extensive verification• efficiency:

• on NEC-SX9: further increase probably only with help of NEC specialists• on Intel/cache based: better inspection tools necessary (valgrind, ...)

• add currently available features:lateral radiation BC; p'-dynamics solver; lower BC for w via 'RK advection of height', ...

• application to the COSMO-DE L65 setup • Closer examination of the influence of (an-)isotropic divergence damping

Page 33: M. Baldauf Deutscher Wetterdienst, Offenbach

Ende

Page 34: M. Baldauf Deutscher Wetterdienst, Offenbach

Neu gegenüber dem bisherigen fast waves -Löser:

1. Verbesserung in der vertikalen Diskretisierung: abstandsgewichtete Mittelungsoperatoren

2. Divergenz in strong conservation form

3. optional volle 3D Divergenzdämpfung (d.h. in allen 3 Bewegungsgleichungen)

zusätzlich einige 'technische' Verbesserungen;Erhöhung der Lesbarkeit

neue Version fast_waves_sc.f90 (basierend auf COSMO 4.17)

Page 35: M. Baldauf Deutscher Wetterdienst, Offenbach

Eine bessere analytische Lösung für die lineare Schwerewelle im Kanal

Analytische Lösung von Skamarock, Klemp (1994) MWRist Boussinesq-approximiertkeine perfekte Übereinstimmungzwischen analytischer und simulierter Lösung möglich.

Eine bessere analytische Lösung als Referenz wäre auch bei Modellvergleichen wünschenswert und wird z.Z. im Rahmen von

durchgeführt

Page 36: M. Baldauf Deutscher Wetterdienst, Offenbach

Eine bessere analytische Lösung für die lineare Schwerewelle im Kanal

Eine analytische Lösung für kompressible, nicht-hydrostatische Euler-Gleichungen kann für eine isotherme (T0=const) Hintergrund-Atmosphäre gefunden werden!

Anfangszustand: ruhend (bzw. u=const.) , p‘=0, vorgegebene T‘- (oder ‘-) Verteilung z.B. der Form:

Page 37: M. Baldauf Deutscher Wetterdienst, Offenbach

Die zeitliche Entwicklung der Fouriertransformierten Felder lautet

(analog für u, T‘, …) und sind gegebene Funktionen von kx, kz, T0, g, cp, cv

Page 38: M. Baldauf Deutscher Wetterdienst, Offenbach

Die zeitliche Entwicklung der Fouriertransformierten Felder lautet

(analog für u, T‘, …) und sind gegebene Funktionen von kx, kz, T0, g, cp, cv

Page 39: M. Baldauf Deutscher Wetterdienst, Offenbach

1. Sound wave expansion test

current FW FW new

Isothermal atmosphere (T=250 K)

Page 40: M. Baldauf Deutscher Wetterdienst, Offenbach

1. Sound wave expansion test

FW new

?

current FW

Page 41: M. Baldauf Deutscher Wetterdienst, Offenbach

1. Sound wave expansion test

FW new

?

current FW

Page 42: M. Baldauf Deutscher Wetterdienst, Offenbach

1. Sound wave expansion test

FW new

?

current FW

Artificial stationary solution (?)

Page 43: M. Baldauf Deutscher Wetterdienst, Offenbach

2. Linear gravity wave

test definition and (anelastic approx.) analytic solution: Skamarock, Klemp (1994) MWR

FW newcurrent FW

Page 44: M. Baldauf Deutscher Wetterdienst, Offenbach

2. Linear gravity wave

test definition and (anelastic approx.) analytic solution: Skamarock, Klemp (1994) MWR

FW newcurrent FW

Page 45: M. Baldauf Deutscher Wetterdienst, Offenbach

2. Linear gravity wave

test definition and (anelastic approx.) analytic solution: Skamarock, Klemp (1994) MWR

FW newcurrent FW

Page 46: M. Baldauf Deutscher Wetterdienst, Offenbach

2. Linear gravity wave

test definition and (anelastic approx.) analytic solution: Skamarock, Klemp (1994) MWR

FW newcurrent FW

Page 47: M. Baldauf Deutscher Wetterdienst, Offenbach

4. Mountain in a steady atmosphere with vertical grid stretching

FW newcurrent FW

Page 48: M. Baldauf Deutscher Wetterdienst, Offenbach

4. Mountain in a steady atmosphere(b) with vertical grid stretching

FW newcurrent FW

Page 49: M. Baldauf Deutscher Wetterdienst, Offenbach

4. Mountain in a steady atmosphere(b) with vertical grid stretching

FW newcurrent FW

Page 50: M. Baldauf Deutscher Wetterdienst, Offenbach

5. Testcase Weisman, Klemp (1982)

FW new

Umax=30 m/s, qv,0,max=13 g/kg

current FW

Page 51: M. Baldauf Deutscher Wetterdienst, Offenbach

FW new

5. Testcase Weisman, Klemp (1982)

current FW

Umax=30 m/s, qv,0,max=13 g/kg

Page 52: M. Baldauf Deutscher Wetterdienst, Offenbach

FW new

5. Testcase Weisman, Klemp (1982)

current FW

Umax=30 m/s, qv,0,max=13 g/kg

Page 53: M. Baldauf Deutscher Wetterdienst, Offenbach

FW new

5. Testcase Weisman, Klemp (1982)

current FW

Umax=30 m/s, qv,0,max=13 g/kg

Page 54: M. Baldauf Deutscher Wetterdienst, Offenbach

5. Testcase by Weisman, Klemp (1982) MWR

Max vertical velocity w:

New FW:small deviationsbetween the anisotropic (quasi-2D)and theisotropic (3D)divergence damping

Current FW

60 min.30 min.