1. Italian Meteorological Service Pratica di Mare, Rome, Italy
MESOSCALE DATA ASSIMILATION FOR THE COSMO MODEL: STATUS AND PERSPECTIVES AT THE IMS
Massimo Bonavita, Lucio Torrisi, Antonio Vocino and Francesca Marcucci
CNMCA
Italian Meteorological Service
Pratica di Mare, Rome, Italy
29° EWGLAM, Dubrovnik, 8-11 October 2007

2. 29° EWGLAM, Dubrovnik, 8-11 October 2007
Summary
Current DAS configuration at the IMS
Recent changes in the DAS
Observation usage
Initialization for the COSMO Model: 3D-VAR vs Nudging
The quest for a ”flow-dependent” analysis
29° EWGLAM, Dubrovnik, 8-11 October 2007

3. 29° EWGLAM, Dubrovnik, 8-11 October 2007
NWP-DAS at the IMS
Domain size
769x513
Grid spacing
0.125° (14 Km)‏
Number of layers 40
Time step
150 sec
Forecast range
72 hrs
Initial time of model run
00/12 UTC
L.B.C.
IFS
L.B.C. update frequency
3 hrs
Initial state
CNMCA 3D-VAR
Initialization
Digital Filter
External analysis
None
Status
Operational
Hardware
IBM P690 (ECMWF)‏
N° of processors used
32 (Model), 90 (Analysis)
29° EWGLAM, Dubrovnik, 8-11 October 2007

4. 29° EWGLAM, Dubrovnik, 8-11 October 2007
NWP-DAS at the IMS
Data assimilation cycle:
3D-VAR FGAT analysis step;
3h refresh cycle;
Prognostic model: HRM hydrostatic model at 14 Km grid (0.125°)
29° EWGLAM, Dubrovnik, 8-11 October 2007

5. NWP-DAS at the IMS
3h vs 6h DAS cycle

6. NWP-DAS at the IMS FGAT vs OPE DAS cycle
E-SAT/ST Meeting, Reading 14-15/05/2007

7. NWP-DAS at the IMS 0.125° vs 0.25° DAS cycle
E-SAT/ST Meeting, Reading 14-15/05/2007

8. Daily observation usage stats.
Synoptic Asynoptic
RAOB ~ AIREP ~5500
PILOT ~ AMDAR ~38000
SYNOP ~ ACAR ~8500
SHIP,BUOY ~ WIND PROF ~1200
- QSCAT/ERS2 ~1800
- ASCAT ~4000
- AMV (MET9/MET7/MODIS)~14000
- AMSU-A Rad. (NOAA1X) ~14000
Synoptic Obs ~ Asynoptic Obs ~87000
Total ~ obs/day
29° EWGLAM, Dubrovnik, 8-11 October 2007

9. Observation usage 2. Use of hyper-spectral sounders data:
1. Use of METOP data:
a) AMSU-A rad.: extension of current NOAA1x
treatment, currently in passive monitoring
configuration
b) ASCAT winds: in place, impact and obs error
characteristics under investigation
c) GRAS Temperature/Spec. Hum. Profiles
2. Use of hyper-spectral sounders data:
a) IASI L2 NRT products when available
b) AIRS L2 products give positive NWP impact (Riishojgaard et al., 2007) but NRT availability is unclear

10. Initialization of COSMO Model
Initialization of COSMO Model at 7 Km resolution
29° EWGLAM, Dubrovnik, 8-11 October 2007

11. Initialization of COSMO Model
Initial Conditions: Interpolated 14 Km 3D-VAR analysis
Domain size
641 x 401
Grid spacing
(7 km)‏
Number of layers 40
Time step
40 s
Forecast range
72 hrs
Initial time of model run
00 UTC
Lateral bound. condit.
IFS
L.B.C. update freq.
3 hrs
Initial state
Interpolated 3D-VAR
Initialization
D.F.I.
External analysis
T,u,v, q, SP
Special features
Filtered topography
Status
Operational
Hardware
IBM P690 (ECMWF)‏
N° of processors
120

12. Initialization of COSMO Model
Initial Conditions: Nudging Data Assimilat. cycle

13. COSMO-ME/COSMO-I7(LAMI)(7km)
COSMO-ME vs COSMO-I7 Temp T+12 00run
MAM
COSMO-ME vs COSMO-I7 Temp T+24 00run
MAM -1
-0,5
0,5 1
1,5 2
2,5
100
150
200
250
300
400
500
700
850
925
1000
-1,5 -1
-0,5
0,5 1
1,5 2
100
150
200
250
300
400
500
700
850
925
1000
Mean Error (dot) and Mean Absolute Error (cont.)

14. COSMO-ME/COSMO-I7(LAMI)(7km)
COSMO-ME vs COSMO-I7 Temp T+36 00run
MAM
COSMO-ME vs COSMO-I7 Temp T+48 00run
MAM
-1,5 -1
-0,5
0,5 1
1,5 2
2,5
100
150
200
250
300
400
500
700
850
925
1000
-1,5 -1
-0,5
0,5 1
1,5 2
100
150
200
250
300
400
500
700
850
925
1000
Mean Error (dot) and Mean Absolute Error (cont.)

15. COSMO-ME/COSMO-I7(LAMI)(7km)
COSMO-ME vs COSMO-I7 Wmod T+12 00run
MAM
COSMO-ME vs COSMO-I7 Wmod T+24 00run
MAM -1
-0,5
0,5 1
1,5 2
2,5 3
3,5 4
100
150
200
250
300
400
500
700
850
925
1000 -1
-0,5
0,5 1
1,5 2
2,5 3
3,5 4
4,5
100
150
200
250
300
400
500
700
850
925
1000
Mean Error (dot) and Mean Absolute Error (cont.)

16. COSMO-ME/COSMO-I7(LAMI)(7km)
COSMO-ME vs COSMO-I7 Wmod T+36 00run
MAM
COSMO-ME vs COSMO-I7 Wmod T+48 00run
MAM
-1, 5 -1
-0, 0, 1 1, 2 2, 3 3, 4 4, 5, 6
100
150
200
250
300
400
500
700
850
925
1000
-1,5 -1
-0,5
0,5 1
1,5 2
2,5 3
3,5 4
4,5 5
100
150
200
250
300
400
500
700
850
925
1000
Mean Error (dot) and Mean Absolute Error (cont.)

17. 29° EWGLAM, Dubrovnik, 8-11 October 2007
Initialization of COSMO Model
Interpolating a 3D-VAR analysis at 14 Km works well for the 7 Km COSMO Model
Significant improvements from 28 to 14 Km analysis: will investigate further resolution increase in 3DVAR DAS (hydrostatic limit)
Clear improvement over parallel COSMO implementation initialized with nudging Not clear cut comparison, lack of explicit balance constraints in nudging could be an issue at 7 Km scale
29° EWGLAM, Dubrovnik, 8-11 October 2007

18. 29° EWGLAM, Dubrovnik, 8-11 October 2007
Initialization of COSMO Model
Initialization of COSMO Model at Km resolution
29° EWGLAM, Dubrovnik, 8-11 October 2007

19. Initialization of COSMO Model
For the 2.8Km version (COSMO-IT) an experiment was run comparing two identical model configurations: one initialized from interpolated 14 Km 3D- VAR, the other with Nudging data assimil. cycle

20. Initialization of COSMO Model
Convegno FAI, Ischia 14/06/2007

21. Initialization of COSMO Model

22. Initialization of COSMO Model

23. Initialization of COSMO Model
Convegno FAI, Ischia 14/06/2007

24. Initialization of COSMO Model
Convegno FAI, Ischia 14/06/2007

25. 29° EWGLAM, Dubrovnik, 8-11 October 2007
Initialization of COSMO Model
Interpolating a 3D-VAR analysis at 14 Km does not provide balanced I.C. for 2.8 Km LM
Observation nudging is able to reduce/suppress precipitation spin-up present in the 3DVAR initialized version
After the first 6-9h skill scores of 3DVAR vs Nudging COSMO-IT are very similar: at 2-3 Km scale, nudging intrinsic balance constraints seems effective and the method looks competitive with 3D-Var
Currently, nudging initialization is employed in operational 2.8Km COSMO-IT
29° EWGLAM, Dubrovnik, 8-11 October 2007

26. The quest for a flow dependent analysis
Extensive discussion inside COSMO community
Lack of resources/expertise to develop 4DVAR
EnKF approach is simpler and seems to be ripe for trial in operational environment (shown to outperform 3DVAR in perfect model simulations and, more recently in real world experiments)
29° EWGLAM, Dubrovnik, 8-11 October 2007

27. The quest for a flow dependent analysis
Possible EnKF advantages for high resolution DAS:
Complex observation operators (i.e. precipitation) coped with automatically
Covariances are evolved indefinitely
Can be extended to assimilate asynchronous observations (4DEnKF)
Gives “optimal” initial perturbations for ensemble forecasting
29° EWGLAM, Dubrovnik, 8-11 October 2007

28. The quest for a flow dependent analysis
Which EnKF version to use? Agreement on LETKF (Hunt et al. 2005) because:
Version of Ensemble Square Root Filter (EnSRF), avoids additional sampling error of perturbed observations
Avoids inefficient sequential analysis of observations of other EnSRF
Computationally efficient (computations performed in ensemble subspace)
Very efficient parallel implementation
29° EWGLAM, Dubrovnik, 8-11 October 2007

29. The quest for a flow dependent analysis
On the other hand:
Rank deficiency of sampled B matrix can be detrimental for affordable ensemble size. Observation localization can help, possible need of hybrid analysis step with 3DVAR
Effective treatment of model error still an issue
29° EWGLAM, Dubrovnik, 8-11 October 2007

30. The quest for a flow dependent analysis
At which resolution should the filter be
used? Different ideas...
DWD has proposed the KEnDA project (COSMO project): Kilometer scale EnDA
Trying to tackle convection as an initial value problem too.
Running an ensemble DA and Forecast system at the Kilometer scale is very computationally expensive..
29° EWGLAM, Dubrovnik, 8-11 October 2007

31. The quest for a flow dependent analysis
At which resolution should the filter be
used? Different ideas...
At IMS we will not have the computing power of DWD for the foreseeable future!
Our forecasting target is the very short to extended short range, i.e. +3h->+72h
For non organized convection +3h is very long range forecasting
Convective systems whose life cycle and predictability extends beyond 3h can usually be modelled at the mesoscale (7-10 Km)
29° EWGLAM, Dubrovnik, 8-11 October 2007

32. The quest for a flow dependent analysis
MESO-DAPS project:
Data Assimilation and Prediction System at
the Mesoscale
Main advantage of unified approach:
“DA step samples initial uncertainties at
correct spatial scales for subsequent
ensemble forecast”
29th EWGLAM, Dubrovnik, 8-11 October 2007

33. The quest for a flow dependent analysis
MESO-DAPS project:
Based on LETKF (or LETKF-3DVAR hybrid approach)
10-14 Km resolution of ensemble members
Will provide lateral and initial conditions for nested Km scale COSMO model (+3-24h)
Currently national project. Will be proposed to COSMO community if proven

34. 29° EWGLAM, Dubrovnik, 8-11 October 2007
Thank you!
29° EWGLAM, Dubrovnik, 8-11 October 2007

35. 3rd SRNWP Workshop on short-range EPS
Rome, December 2007
Università Roma I “Sapienza”