HIRLAM 34 DVar developments Nils Gustafsson SMHI Parallel

HIRLAM 3/4 D-Var developments Nils Gustafsson, SMHI

Parallel data assimilation work along 2 lines in HIRLAM for the synoptic scales: HARMONIE for the mesoscale: • 3 D-Var and 4 D-Var • Further developments during 2008 -2009 • To be phased out operationally 2010 -2012 • • Based on ALADIN (IFS) 3 D-Var mid 2008 4 D-Var early 2009 To replace the synoptic scale HIRLAM (20102012)

HIRLAM 4 D-Var components: • Tangent linear and adjoint of the semi. Lagrangian (SETTLS) spectral HIRLAM. • Simplified physics packages: Buizza vertical diffusion and Meteo France (Janiskova) package(vertical diffusion, large-scale condensation and convection). • Multi-incremental minimization (spectral or gridpoint HIRLAM in outer loops). • Weak digital filter constraint. • Control of lateral boundary conditions.

Noise in assimilation cycles with the gridpoint model

Comparison tests 3 D-Var – 4 D-Var • SMHI area, HIRLAM 7. 1. 1, KF/RK, SMHI area, statistical balance background constraint, reference system background error statistics (scaling 0. 9), no ”large-scale mix”, LINUX cluster, 4. 5 months, operational SMHI observations and boundaries • 3 D-Var with FGAT, incremental digital filter initialization • 4 D-Var, 6 h assimilation window, weak digital filter constraint, no explicit initialization

Summary of forecast scores Period Surface pressure Upper air April 2004 Neutral Positive impact of 4 D-Var Jan 2005 Positive impact of 4 D-Var June 2005 Neutral Positive impact of 4 D-Var Jan 2006 (11 days) Positive impact of 4 D-Var Small positive impact of 4 D-Var Jan 2007 Positive impact of 4 D-Var Small negative impact of 4 D -Var on 300 and 200 h. Pa heights

Operationalization of 4 D-Var • SMHI tests show positive impact of 4 D-Var in comparison with 3 D-Var • SMHI results need to be confirmed with the reference system (and new NL physics) • Improved parallel scaling is needed: (a) open. MP within nodes & MPI between nodes; (b) Message passing for SL advection ”on demand” • To be included in HIRLAM 7. 2 (late 2007)

Pre-operational tests of 4 D-Var at SMHI Cop - SMHI op. 22 km, Hirlam-6. 3. 5, KF/RK, 3 DVAR FGAT Cnn - Hirlam-7. 1. 2, KF/RK, 4 DVAR

Illustration structure functions Impact of one single surface pressure observation 5 h. Pa less than the corresponding background equivalent (red: surface pressure, black: winds at lowest mod level) Analytical NMC (48 -24) Statistical NMC (36 -12) Statistical Ensemble

Flow dependent background covariances through non-linear balance equations Non-linear balance equation on pressure levels: Tangent-linear version of omega equation on pressure levels: where

Vertical crossection of T increments Statistical balance Weak constraints balance eq.

A new moisture control variable and a new moisture balance Within the analytical balance formulation we follow Holm and use relative humidity as control variable and the TL RH definition for the balance: In addition, the background error variance depends on the background relative humidity (makes it more Gaussian). Within the statistical balance formulation (with q as control variable), we already have a statistical balance relation:

In order to avoid double-counting of the temperaturemoisture balance, we could try to improve the statistical balance relation by using coefficients from the analytical balance relation, for example: So far we have tried: In this case, we also used a background error variance depending on the background relative humidity

New assimilation control variable for humidity (analytical balance version) Old formulation q New formulation: RH*= RH/σ (RH +0. 5 RH) b b

New assimilation control variable for humidity (statistical balance with multivariate humidity) Assimilation increments due 5 simulated specific humidity observations, 10 g/kg smaller than corresponding background equivalent (sigmao: 1 g/kg) q at 850 h. Pa (g/kg times 10) ps (h. Pa times 10)

SEVIRI data coverage (At SMHI, we don’t store the raw-data for the full SEVIRI disc operationally)

Example of impact of SEVIRI data on 3 D-Var analysis • • Difference of analysed 500 h. Pa relative humidity (SEVIRI experiment minus Control) Impact can be seen mainly in the southern part of the domain

3 D-Var 4 D-Var

3 D-Var/4 D-Var impact study • • • 3 D-Var: Positive impact on upper-troposhperic water vapour is found Positive impact on MSLP forecast is found • • • 4 D-Var: Positive impact on upper-tropospheric water vapour is found Also Temperature and Geopotential fields show some response (small positive impact) • Another impact study for December 2005 shows neutral impact of SEVIRI data in terms of forecast scores. Work is now continuing with a much more difficult problem, assimilation of cloudy SEVIRI radiances!

What can we expect to achieve with the HIRLAM data assimilation before it will be phased out? • 4 D-Var with several outer loops and improved moist physics • Control of lateral boundary conditions in 4 D-Var • A new moisture control variable • Large scale mix vi a Jk cost function term • Background and large scale error statistics based on Ens. Ass • Tuning of screening and Var. QC • Use of several new types of observations. (IASI? ) Most development efforts should be finished during 2008! A synoptic scale HARMONIE should be comparable!!