1. WRFDA 2012 Overview and Recent Adjoint-based Observation Impact Studies Hans Huang National Center for Atmospheric Research (NCAR is sponsored by the National Science Foundation)
Acknowledge: NCAR/NESL/MMM/DAS, NCAR/RAL/JNT/DAT, AFWA, USWRP, NSF-OPP, NASA, AirDat, PSU,
KMA, CWB, CAA, BMB, EUMETSAT

2. WRFDA Overview Goal: Community WRF DA system for regional/global,
research/operations, and
deterministic/probabilistic
applications.
Techniques:
3D-Var
4D-Var (regional)
Ensemble DA,
Hybrid Variational/Ensemble DA.
Model: WRF (ARW, NMM, Global)
Observations: Conv. + Sat. + Radar (+Bogus)
Support:
NCAR/NESL/MMM/DAS (Data Assimilation Section, also supporting WRF/DART)
NCAR/RAL/JNT/DAT (Data Assimilation Team, also supporting GSI)

3. Google WRFDA: www.mmm.ucar.edu/wrf/users/wrfda

4. Recent Tutorials at NCAR.
WRFDA Overview
Observation Pre-processing
WRFDA System
WRFDA Set-up, Run
WRFDA Background Error Estimations
Radar Data
Satellite Data
WRF 4D-Var
WRF Hybrid Data Assimilation System
WRFDA Tools and Verification
Observation Sensitivity
Practice
obsproc
wrfda (3D-Var)
Single-ob tests
Gen_be
Radar
Radiance
4D-Var
Hybrid
Advanced (optional)
The next: July 2012

5. WRFDA tutorials 21-22 July, 2008. NCAR. 2-4 Feb, 2009. NCAR.
18 April, South Korea.
20-22 July, NCAR.
15-31 Oct, Nanjing, China.
1-3 Feb, NCAR.
10 April, Seoul, South Korea.
3-5 August NCAR.
16 April. Busan, South Korea
20-22 July NCAR
10-20 October. Bangkok, Thailand.
WRFDA online tutorial and user guide

6. New features, v3.3
WRFPLUS3 – WRF TL/AD based on WRF3.3
4D-VAR redesigned/upgraded
RTM interface updated:
RTTOV (v10.0)
CRTM (v2.0.2)
NOAA-19 AMSUA and MHS added.
New background error option (cv6).
Add additional variables, e.g. unbalanced velocity, to balance computations;
Fully multivariate including relative humidity
Capability to generate forecast sensitivity to observations (FSO), compatible with WRF3.3

7. A recent adjoint-based observation impact study (or FSO) Thomas Auligné, Xin Zhang, Hongli Wang, Anwei Lai, Xiaoyan Zhang, Hui-Chuan Lin, Qingnong Xiao, Yansong Bao, Zhiquan Liu, Xiang-Yu Huang

8. FSO - Forecast Sensitivity to Observations
Analysis
(xa)
Forecast
(xf)
Observation
(y)
WRF-VAR
Data
Assimilation
WRF-ARW
Forecast
Model
Define
Forecast
Accuracy
Background
(xb)
Observation Impact
<y-H(xb)> (F/ y)
Forecast
Accuracy
(F)
Observation
Sensitivity
(F/ y)
Adjoint of
WRF-VAR
Adjoint of
WRF-ARW
Derive
Forecast
Accuracy
Background
Sensitivity
(F/ xb)
Analysis
Sensitivity
(F/ xa)
Gradient
of F
(F/ xf)
Obs Error
Sensitivity
(F/ eob)
Bias Correction
Sensitivity
(F/ k)
Figure adapted from Liang Xu (NRL)

9. xf is the forecast from analysis xa
From Langland and Baker (2004)
Forecast
Error
Time
xt is the true state, estimated by the analysis at the time of the forecast
xf is the forecast from analysis xa
xg is the forecast from first-guess at the time of the analysis xa

10. More details (for WRFDA implementation):
Reference state: Namelist ADJ_REF is defined as
1: xt = Own (WRFVar) analysis
2: xt = NCEP (global GSI) analysis
3: xt = Observations
Forecast Aspect: depends on reference state
1 and 2: Total Dry Energy
3: WRFDA Observation Cost Function: Jo
Geo. projection: Script option for box (default = whole domain)
ADJ_ISTART, ADJ_IEND,
ADJ_JSTART, ADJ_END,
ADJ_KSTART, ADJ_KEND
Forecast Accuracy Norm: e = (xf-xt)T C (xf-xt)

11. Limitations Approximation of “truth” Dependence of norm
Linear assumptions
Adjoint of the forecast model
Adjoint of the analysis (assimilation) …

12. ASR (180km) Results January 2007

13. Design of experiments Cycling Data assimilation Start Time:2007081506
Final Time:
Boundary Conditions : FNL analysis
The initial condition for is the FNLdata, The WRF model 6h forecasting is the cycling data assimilation background.
ADJ_REF =1, Its own WRFDA analysis
ADJ_MEASURE =1

14. Impact of observations on 24h forecast (conventional observations)

15. Impact of observations on 24h forecast (satellite observations)

16. T8 (45km) Results One month:2007081506-2007091512
AWFA T8
WRFV3.2.1
MAP_PROJ=Mercator
WE =140
SN =94
DX =45km
DY =45km
E_VERT =57
MP_PHYSICS=WSM 5-class scheme
CU_PHYSICS= Grell 3D ensemble scheme
Target
region
WRFPLUS2.0
ADJ_ISTART=20 ADJ_IEND=60
ADJ_JSTART=20 ADJ_JEND=60
ADJ_KSTART=1 ADJ_KEND=27
WRFV3 Version 3.2.1
WRFDA Version 3.2.1
WRFPLUS2.0

17. Impact of observations on 6h forecast (grouped as observed variables)

18. Impact of observations on 6h forecasts (grouped according to observing systems)

19. SOUND
GeoAMV
GPSRO
hPa
hPa Km
Vertical Impact distribution, unit: J/Kg

20. Time series of impact of observations on 6h forecasts
SOUND
SYNOP
GeoAMV
X axis :time. 4 times per day during 30days ;
Y axis : Impact on the forecasting (Unit:J/kg)

21. Monitoring the observation impact over Taiwan CWB OP23 domain
Xin Zhang and Xiang-Yu Huang

24. In terms of the observational type
The largest error decrease is due to GeoAMV followed by SOUND, SYNOP and GPSREF
The impact from SATEM is marginal and neutral.
On 0000UTC and 1200UTC, the SOUND is the most important observation to decrease the forecast error, followed by GeoAMV, SYNOP, GPSREF/AIREP
on 0600UTC and 1800UTC, the GeoAMV is the most important observation to decrease the forecast error, followed by SYNOP/GPSREF, AIREP

29. In terms of the time series of impact of observational type
On 0000 and 1200UTC, the SOUND improve the forecasts in general.
On 0600UTC and 1800UTC, the SOUND almost degrades the 1/3 of the forecasts. Why ? Are the few SOUND obs. bad quality or something else is wrong ?
For other observation types, at 0600UTC and 1800UTC, there are many degraded cases due to observations. Need further investigation.

32. In terms of impact per observation for each type
The GPSREF is the most efficient observation to reduce the 24h forecast error per observation, followed by SYNOP, SHIPS and SOUND. It is consistent with the result from AFWA domains ( personal communication with Jason T. Martinelli of AFWA)

33. SOUND observations around 500hPa (206 records)

34. 500hPa SOUND impact

35. SYNOP observations

36. SYNOP impact

37. In terms of the geographical distribution of observation impact
Gives us an insight of the detail information about the observation impact for each observation.
Please note: On Taiwan, there are lots SYNOP observations. Due to the plotting technique, the blue dots are overwritten by the red dots.

38. Preliminary results
SOUND is the most important data to improve the forecast among the conventional data.
GPSREF is the most efficient data to improve the forecast and is the most important observation at 0600 and 1800UTC.
The impact of SATEM is marginal and not stable.
For OP23 system, at 0600UTC and 1800UTC, the observation impacts from all observational types are not stable. (Why?)
In two days ( and ), almost all the observations show the negative impact. Preliminary investigation shows that, in the first case ( ), the linear approximation of WRFPLUS is not valid, the nonlinear forecasts show the positive impact of observation, but the WRFPLUS derived the negative impact.

39. Future research topics
Identify observations with continuous negative impact.
Tune WRFDA based on observation impact results.
Improve the assimilation strategy of surface observation assimilation.
Investigate the differences of verifying forecasts to EC analysis, NCEP GFS analysis, WRFDA analysis and Observations.

40. Future research topics
For limited–area forecast models, the forecast errors not only come from initial condition (IC), but also come from lateral boundary condition (LBC). Need to include the initial forecast error gradient with respect to LBC in FSO.
Current WRF tangent linear adjoint models (WRFPLUS) only include 3 simplified physics packages. Need to study the accuracy of the linear approximation in FSO applications. We also plan to improve the physics packages in WRFPLUS.

41. Summary WRFDA 2012 Overview
Community System
Techniques: 3D-Var, 4D-Var, EnKF, Hybrid
New features in v3.3
A recent adjoint-based observation impact study
The concept and limitations
The WRFDA implementation
Applications
The next step