Control of Gravity Waves Lars Isaksen Room 308, Data Assimilation, ECMWF Gravity waves and divergent flow in the atmosphere Two noise removal approaches: filtering and initialization Normal mode initialization Digital filter Control of gravity waves in the ECMWF assimilation system Lars Isaksen, ECMWF, March 2006 1 Data assimilation and use of satellite data Processes and waves in the atmosphere Sound waves, synoptic scale waves, gravity waves, turbulence, Brownian motions .. The atmospheric flow is quasi-geostrophic and largely rotational (non-divergent) – mass/wind balance at extratropical latitudes The energy in the atmosphere is mainly associated with fairly slow moving large-scale and synoptic scale waves (Rossby waves) Energy associated with gravity waves is quickly dissipated/dispersed to larger scale Rossby waves: the quasi-geostrophic balance is reinstated Lars Isaksen, ECMWF, March 2006 2 Data assimilation and use of satellite data 500 hPa Geopotential height and winds Approximate mass-wind balance Lars Isaksen, ECMWF, March 2006 3 Data assimilation and use of satellite data MSL pressure and 10 metre winds Approximate mass-wind balance Lars Isaksen, ECMWF, March 2006 4 Data assimilation and use of satellite data Which atmospheric processes/waves are important in data assimilation and NWP? Sound and gravity waves are generally NOT important, but can rather be considered a nuisance Fast waves in the NWP system require unnecessary short time steps – inefficient use of computer time Large amplitude gravity waves add high frequency noise to the assimilation system resulting in: – rejection of correct observations – noisy forecasts with e.g. unrealistic precipitation BUT certain gravity waves and divergent features should be retained in a realistic assimilation system. We will now present some examples. Lars Isaksen, ECMWF, March 2006 5 Data assimilation and use of satellite data Ageostrophic motion – Jet stream related An important unbalanced synoptic feature in the atmosphere Wind and height fields at 250 hPa Ageostrophic winds at 250 hPa Lars Isaksen, ECMWF, March 2006 6 Data assimilation and use of satellite data Mountain generated gravity waves should be retained Rocky Mountains Lars Isaksen, ECMWF, March 2006 7 Data assimilation and use of satellite data Temperature cross-section over Norway Gravity waves in the ECMWF analysis Norway Lars Isaksen, ECMWF, March 2006 8 Acknowledgements to Agathe Untch Data assimilation and use of satellite data Analysis temperatures at 30 hPa Acknowledgements to Agathe Untch Lars Isaksen, ECMWF, March 2006 9 Data assimilation and use of satellite data Equatorial Walker circulation Lars Isaksen, ECMWF, March 2006 10 Data assimilation and use of satellite data Divergent winds at 150hPa: ERA-40 average March 1989 Acknowledgements to Per Kållberg Lars Isaksen, ECMWF, March 2006 11 Data assimilation and use of satellite data Semi-diurnal tidal signal Lars Isaksen, ECMWF, March 2006 12 Data assimilation and use of satellite data Observed Mean Sea-Level pressure - Tropics Semi-diurnal tidal signal for Seychelles (5N 56E) Lars Isaksen, ECMWF, March 2006 13 Data assimilation and use of satellite data Filtering the governing equations Goal: Use filtered model equations that do not allow high frequency solutions (“noise”) – but still retain the “signal” Quasi-geostrophic equations/ omega equation Primitive equations with hydrostatic balance Primitive equations with damping time-step like Eulerian backward Primitive equations with digital filter Lars Isaksen, ECMWF, March 2006 14 Data assimilation and use of satellite data Initialization Goal: Remove the components of the initial field that are responsible for the “noise” – but retain the “signal” • Make the initial fields satisfy a balance equation, e.g. quasi-geostrophic balance or • Set tendencies of gravity waves to zero in initial fields – Non-linear Normal Mode Initialization Lars Isaksen, ECMWF, March 2006 15 Data assimilation and use of satellite data Normal-mode initialization Linearize forecast model about a statically-stable state of rest: dx iLx N(x) 0 dt L represents linear terms where N represents the nonlinear terms and diabatic forcing Diagonalize L by transforming to eigenvalue-mode - “Hough space”: dx iEET x N(x) 0 dt where Λ is the diagonal eigenvalue matrix Split eigenvalues into slow Rossby modes and fast Gravity modes. dx R iER Λ R ETR x R N R (x R , x G ) 0 dt dx G iEG Λ G ETG x G NG (x R , x G ) 0 dt Lars Isaksen, ECMWF, March 2006 16 x x R xG Data assimilation and use of satellite data Rossby modes and Gravity modes Frequency The ‘critical frequency’ separating fast modes from slow. Mixed Rossby-Gravity Wave Non-dimensional wavenumber Lars Isaksen, ECMWF, March 2006 17 Data assimilation and use of satellite data Non-linear Normal-Mode Initialization The fast Gravity modes generally represent “noise” to be eliminated. dx G iEG Λ G ETG x G NG (x R , x G ) 0 dt dxk i k xk N k 0 for one eigenvalue, k dt If Nk is assumed constant (i.e. slowly varying compared to gravity waves): x k ( t ) ( x k ( 0) dxk 0 At initial time set dt N k ik t N k )e ik ik Nk then xk (0) ik so Nk xk ( t ) 0 ik The high frequency component is removed and will NOT reappear. Assumes that the slow Nk forcing balances the oscillations at initial time. Lars Isaksen, ECMWF, March 2006 18 Data assimilation and use of satellite data Non-linear NMI: USA Great Planes Surface pressure evolution Non-linear NMI initialized field Temperton and Williamson (1981) Uninitialized field Lars Isaksen, ECMWF, March 2006 19 Data assimilation and use of satellite data Optimal and approximate low-pass filter Consider a infinite sequence of a ‘noisy’ function values: {x(i)} We want to remove the high frequency ‘noise’. One method: perform direct Fourier transform; remove high-frequency Fourier components; perform inverse Fourier transform. This is identical to multiplying {x(i)} by a weighting function: fk h x n ( nk ) n sin n c t ( n k ) x n n c is the cut-off frequency The finite approximation is: fk N hn x n N ( n k ) sin n c t ( n k ) x n n N N Lars Isaksen, ECMWF, March 2006 20 f0 N (n) h x n n N sin n c t ( n ) x n n N N Data assimilation and use of satellite data Digital filter Consider a sequence of model values {x(i)} at consecutive adiabatic time-steps starting from an uninitialized analysis A digital filter adjusts values to remove high frequency ‘noise’ Adiabatic, non-recursive filter: Perform forward adiabatic model integration {x(0),x(1),…,x(N)} Perform backward adiabatic model integration {x(0),x(-1),…,x(-N)} The filtered initial conditions are: N 1 Init ( x ) hn x ( n ) h n N ( 0) Lars Isaksen, ECMWF, March 2006 where h N h n N 21 n Data assimilation and use of satellite data Fourier filter and Lanczos filter c Damping factor for waves sin( n c t ) hn n Gibbs Phenomenon for Fourier filter sin( n c t ) sin[ n /( N 1)] hn n n /( N 1) Broader cut-off for Lanczos filter Wave frequency in hours Lars Isaksen, ECMWF, March 2006 22 Data assimilation and use of satellite data Transfer function for Lanczos filter 6 hour window Lars Isaksen, ECMWF, March 2006 23 Data assimilation and use of satellite data Transfer function for Lanczos filter 12 hour window Lars Isaksen, ECMWF, March 2006 24 Data assimilation and use of satellite data Transfer function for Lanczos filter 6 and 12 hour window Lars Isaksen, ECMWF, March 2006 25 Data assimilation and use of satellite data Response to Lanczos filter with 6h cut-off Lars Isaksen, ECMWF, March 2006 26 Data assimilation and use of satellite data Diabatic non-linear normal mode initialization Full-field initialization (ECMWF, 1982-1996) Incremental initialization (ECMWF, 1996-1999) Let xb denote background state, expected to be “noise free” xU the uninitialized analysis xI the initialized analysis and Init(x) the result of an adiabatic NMI initialization. Then xI = xb + Init(xU) – Init(xb) Lars Isaksen, ECMWF, March 2006 27 Data assimilation and use of satellite data Control of gravity waves within the variational assimilation Minimize: Jo + Jb + Jc • • • • Primary control provided by Jb (mass/wind balance) In 4D-Var Jo provides additional balance Digital filter or NMI based Jc contraint Diffusive properties of physics routines Lars Isaksen, ECMWF, March 2006 28 Data assimilation and use of satellite data Control of gravity waves within the variational assimilation Primary control provided by Jb (mass/wind balance) Lars Isaksen, ECMWF, March 2006 29 Data assimilation and use of satellite data NMI based Jc constraint Still used at ECMWF in 3D-Var and until 2002 in 4D-Var dx dx b Jc dt G dt G 2 I. Project “analysis” and background tendencies onto gravity modes. II. Minimize the difference. Noise is removed because background fields are balanced. Lars Isaksen, ECMWF, March 2006 30 Data assimilation and use of satellite data Weak constraint Jc based on digital filter Implemented by Gustafsson (1992) in HIRLAM and Gauthier+Thépaut (2000) in ARPEGE/IFS at Meteo-France Removes high frequency noise as part of 12h 4D-Var window integration Apply 12h digital filter to the departures from the reference trajectory A spectral space energy norm is used to measure distance. – At Meteo-France all prognostic variables are included in the norm – At ECMWF only divergence is now included in the norm, with larger weight Obtain filtered departures in the middle of the assimilation period (6h) Propagate filtered increments valid at t=6h by the adjoint of the tangentlinear model back to initial time, t=0. Get J c (x 0 ) and x J c (x 0 ) Jc calculation is a virtually cost-free addition to Jo calculations Lars Isaksen, ECMWF, March 2006 31 Data assimilation and use of satellite data Weak constraint Jc based on digital filter Apply 12h digital filter to the departures from the reference trajectory and obtain filtered values in the middle (6h): x N / 2 N N h x n 0 n n Use tangent-linear model, R, to get: x N / 2 hn R (t0,tn )x 0 n 0 Define penalty term using energy norm, E: J c (x0 ) 1 x N / 2 xN / 2 2 2 E The gradient of the penalty term is propagated by the adjoint, R*, of the tangent-linear model back to time, t=0: N x J c (x 0 ) n R (t0,t n )(x N / 2 x N / 2 ) * n 0 N R * (t0,tn ) n E(x N / 2 x N / 2 ) hn for n N 2 n 1 hn n N 2 n 0 Jc calculation is a virtually cost-free addition to Jo calculations Lars Isaksen, ECMWF, March 2006 32 Data assimilation and use of satellite data Hurricane Alma – impact of Jc formulation Jc on divergence only with weight=100 versus Jc on all prognostic fields with weight=10 MSL pressure and 850hPa wind analysis differences Lars Isaksen, ECMWF, March 2006 33 Data assimilation and use of satellite data Impact of Jc formulation Jc on divergence only with weight=100 versus Jc on all prognostic fields with weight=10 Impact near dynamic systems and near orography. Fit to wind data improved. In general a small impact. MSL pressure and 850hPa wind analysis differences Lars Isaksen, ECMWF, March 2006 34 Data assimilation and use of satellite data Minimization of cost function in 4D-Var Value of Jo, Jb and Jc terms Lars Isaksen, ECMWF, March 2006 35 Data assimilation and use of satellite data Minimization of cost function in 4D-Var Value of Jo, Jb and Jc terms – logarithmic scale Lars Isaksen, ECMWF, March 2006 36 Data assimilation and use of satellite data Himalaya grid point in 3D-Var - No Jc Lars Isaksen, ECMWF, March 2006 37 Data assimilation and use of satellite data Himalaya grid point in 3D-Var Lars Isaksen, ECMWF, March 2006 38 Data assimilation and use of satellite data Himalaya grid point in 4D-Var Lars Isaksen, ECMWF, March 2006 39 Data assimilation and use of satellite data Himalaya grid point in 4D-Var Lars Isaksen, ECMWF, March 2006 40 Data assimilation and use of satellite data Seychelles (5S 56E) MSL observations plus 3D-Var First Guess and Analysis Observed value First guess value 8 Feb 1997 14 Feb1997 Observed value Analysis value Lars Isaksen, ECMWF, March 2006 41 Data assimilation and use of satellite data Seychelles (5S 56E) MSL observations plus 4D-Var First Guess and Analysis Observations and first guess values Observations and analysis values 4D-Var handles tidal signal very well ! Lars Isaksen, ECMWF, March 2006 42 Data assimilation and use of satellite data We discussed these topics today Gravity waves and divergent flow in the atmosphere Two noise removal approaches: filtering and initialization Normal mode initialization Digital filter Control of gravity waves in the ECMWF assimilation system Lars Isaksen, ECMWF, March 2006 43 Data assimilation and use of satellite data