Title: Recent Developments in Data Assimilation at NCAR MM5WRF 3DVAR
1Recent Developments in Data Assimilation at
NCAR(MM5/WRF 3DVAR)
- S. R. H. Rizvi
- National Center For Atmospheric Research
- NCAR/MMM, Bolder, CO-80307, USA
- Email rizvi_at_ucar.edu
2Outline of talk
- Overview of MM5/WRF 3DVAR
- Algorithm
- Control variables
- Balance
- Statistical parameters
- Observations
- Computational efficiency
- Latest development results
- Future plan
- Conclusion
3An overview of WRF/MM5 3DVAR
Namelist File
Xb
BE
Yo
3DVAR START
Setup Observations
Setup Background Errors
Read Namelist
Setup Background
Setup MPP
Compute Analysis
Calculate (O B)
Minimise Cost Function
Outer Loop
Output Analysis
Calculate Diagnostics
Tidy up
Diagnostic File
Xa
3DVAR END
43DVAR Algorithm
- Incremental 3DVAR approach
- Courtier et al. (1994) Veerse F. J.-N.
Thepaut (1998) -
- The cost function (J) is defined as
- J(X?) 1/2 X?T B1X? H?(X?) - dT
(O F)1 H?(X?) d - Cost 1/2 Background
Observations - X? Analysis increments (X - Xb )
- d Innovation vector (Yo- H(Xb) )
- Yo Observation vector
- H Forward (Non-linear) Observation
Operator (FOO) - H? Tangent linear operator of the forward
operator, H - B Background (previous forecast) errors
- O Observation (instrumental) errors
- F Representivity (observation operator)
errors
53DVAR Algorithm Contd.
- In terms of control variable ( V , X? UV
where B UUT ) the cost function (J) can
written as - J(V) 1/2 VVT H? (UV
)- dT (O F) 1 H? (UV)- d - Cost 1/2 Background
Observatios - Thus minimisation ( ?J/?v 0 ) of cost function
(J) leads to, - V - UT H?T (O F)1 (d H?UV) 0
or -
- AV R
Analysis equation - Where,
- A I UT H?T (O F)1 H? U
and - R UT H?T (O F)1 d
- Analysis equation is solved for V and thus X? is
determined -
63DVAR Algorithm Contd.
- Practical implementation of 3DVAR requires
simplifications - Simplified error covariances.
- Linearized observation operators, balance
equation. - Thinning of observations.
- Suitable choice of analysis control variables
- etc.
7Control variables
- In MM5/WRF there are three choices for the
control variable - cv_option 1
- U-component of wind
- V-component of wind
- Temperature
- Pressure
- Moisture variable as specific
humidity or relative humidity - cv_option 2
- Stream function (?)
- Velocity potential (?)
- Unbalanced part of pressure (Pu)
- Moisture variable as specific humidity
or relative humidity - cv_option 3
- Stream function (?)
- Unbalanced part of velocity potential (?u)
- Unbalanced part of temperature (Tu)
- Log of surface pressure
- Pseudo relative humidity
8Control variables Contd.
- Control variable (V) is defined as,
- X? UV , U Up
Uv Uh - Where, B E L ET UUT , U E L1/2
- E and L are the eigenvectors and eigenvalues of
B. - Uh Horizontal transform, using Recursive
filters -
Purser and Hyden (1998) - Uv Vertical transform, using eigenvectors (E
and L) - Up Physical transform, using Physical/Dynamical
laws
9Recursive Filter
Right moving Bi a Bi-1 (1- a) Ai
(Input A output B Left moving Ci a Ci1
(1- a) Bi (Input B output C ) a
smoothing factor (0 lt a lt 1) Characteristic
Scalelength (R)
R, N and d are fixed to get a
10Recursive Filter
Number of Passes
11Recursive Filter
Response of Scalelength
12Horizontal background errors
- Uh Isotropic/homogeneous recursive filter
algorithm.
Correlation lengthscale
Single T obs (O-B1K, p500hPa)
13Vertical background errors
- Uv Vertical EOFs transform
Eigenvectors of PSI
Single U obs (O-B1m/s, p200hPa)
14Wind and Mass Balance
Linearized geostrophic, cyclostrophic balance
equation
Unbalanced pressure
C Regression coefficient
(Statistically determined)
15Statistical parameters
- Observation error
- Background error
- Balance rgeression coefficients
- Characteristic lengthscale
16Observations
- Conventional
- Upper air (TEMP, PIBAL, AIREP, ACARS,PROFILER).
- Surface (SYNOP, METAR, SHIP,BUOY)
- Remotely sensed retrievals
- Cloud-track winds (SATOBS).
- ATOVS thicknesses (SATEMs).
- Ground-based GPS TPW/ZTD.
- SSM/I oceanic surface wind speed and TPW.
- SSM/T1 temperature retrievals.
- SSM/T2 relative humidity retrievals.
- Scatterometer (Quikscat) oceanic surface winds.
- Radiances
- SSM/I brightness temperatures.
17Statistical approximations
- Climatological background errors
- Estimated via tuned NMC-method
statistics - Simplified horizontal background error
covariances represented by simple recursive
filters - Uncorrelated observation errors
- Neglect error correlations between analysis
variables (streamfunction, potential,
unbalanced pressure and humidity variable (q or
RH) - Approximate balance relationship
- Geostrophic, Cyclostrophic, Hydrostatic
increments.
18Computational efficiency
- 3DVAR MPP Domain Decompositions
1
Recursive Filters and FFTs
2
1
2
3
4
3
4
Minimisation
2
1
1
3
Obs. Operators
4
3
4
3
19Computational efficiency Contd.
IBM-SP, Domain size 140x150x41
20Computational Efficiency Contd.
- Data Compression Via Truncation Of Vertical EOFS
- Cost for 100x100x31 (CAA domain) 45km 3DVAR
- with conventional observations
Conclusion Halve 3DVAR cost with data, with
negligible lose of accuracy
213DVAR/MM5 AFWA Global Theaters
223DVAR/MM5 AFWA Tropical Theaters
23Recent Developments
- Conjugate Gradient Minimisation
- Implementation of outer loop
- Background error computation
- Surface data assimilation
- Improved vertical interpolation
- Assimilation of Radar data
24Conjugate Gradient (CG) Method
- Conjugate gradient method is an efficient way of
solving simultaneous system of linear equations,
- A X B
- It is an iterative method and converges very fast
if A is positive definite matrix -
- Convergence is accelerated by selecting the
search direction to make sure that the search is
always made in a direction perpendicular to the
direction already searched. - Following Golub and Van Loan (1990) and Chandra
(1978), CG algorithm have been coded in MM5/WRF
3DVAR as an additional minimisation option. -
25CG Algorithm
- Equation to be solved (Analysis equation) AV
R , where - A I UTH?T (O F)1 H?U
and - R UTH?T (O F )1 d
- Set, r0 - R and P1 - r0
- where R UTH?T(O F)1d
- For k ? 1
- fk A Pk where A I UTH?T (O
F)1 H?U - S ?rk-1 , rk-1? / ?Pk , fk?
- Vk Vk-1 S ? Pk
- rk rk-1 S ? fk
- Pk1 - rk ?rk , rk? / ?rk-1 , rk-1? ?
Pk -
- Iterate the above sequence till the desired
convergence is achieved
26CG Performance
27CG Efficiency
28Implementation of outer loop
- The analysis equation (AV B) is solved using
double - iteration loop as follows
-
- Set X Xb , so that V 0
- Start of outer iteration
- X Xb U V
- Start of inner iteration
- Compute R UTH?T(O F)1d
- Solve A d R for d
- Update V V d
- End of inner loop
- End of outer loop
29Outer loop Performance
30Outer loop advantages
- Use of additional data
- Non-linearity of the forward observation operator
- Multiple background Quality control
- Efficient utilisation of PBL information for
assimilation of surface data - Effective utilisation of meso-scale data
31Background errors
- The background errors are computed using
NMC-method (Parrish and Derber,1992). - This is the most popular method used at various
operational NWP centers - Generally one month forecast data is used for
computing the various statistical parameters
including the background errors, used in 3DVAR - It is highly compute intensive and requires huge
amount of computing resources. - Currently MM5/WRF 3DVAR uses interpolated
statistics generated from AVN forecast. - The original statistics is at 101x181x21 with 210
Km. resolution.
32Global Domain For Calculation Of AVN
Background Errors101 x 181 x 21 at 210 Km.
resolution
33New Background errors
- For understanding the issues involved with using
the interpolated statistics new Background errors
were generated for the following three regions - Indian 75 x 101 x 23 at
90 Km. - AMPS 181 x 101 x 29 at 90
Km. - T4B 226 x 289 x 41 at
15 Km.
34Lengthscale (psi)
35CPU requirements
36Single-obs (Temperature) test (T4B)
Old BE
37Wind response (T4B)
O L D BE
38Single-obs (Temperature) test (AMPS)
New BE
Old BE
39Wind response (AMPS)
O L D BE
40New approach for Surface data assimilation
- Why new approach?
- It was observed that too many surface observation
reports, specially over the complex terrain, were
getting rejected - While using surface obs by reducing it to the
lowest model sigma level some background error
information also gets blended into obs
41Surface data assimilation - Contd.
- Forward observation operator and its adjoint are
developed to compute 10-m wind and 2-m
temperature, moisture based on Monin-Obukhov
similarity theory Cardinali et al. (1994),
Courtier et al. (1998) and - Guo et al. (2002)
- Suitable correction is applied to Surface
Pressure corresponding to the difference between
the actual and model terrain height - Conclusion 50-60 additional surface data
utilisation
42Improved vertical interpolation
h, p, T
k1
Vertical Interpolation Old 1) If ho not
observed, derive ho from po . 2) Interpolate in
h. New 1) Interpolate in h or p (depends on
observation).
To, po
k
h, p, T
k-1
h, p, T
43Radar Data Assimilation
- Additional control variables for assimilating
- Reflectivity Radial velocity
- - vertical velocity (w)
- - cloud rain water (qr)
- - cloud liquid water/ice (qc)
- - cloud water vapor (qv)
- Full Micro-Physics have been added as a part of
FOO and its tangent linear code is developed and
integrated with the minimisation scheme.
44Radar data assimilation Contd.
- Observation operator (Sun and Crook, 1998)
- ri - distance between radar and the observation
- qr - rainwater
45Radar data assimilation Contd.
- Richardson equation which is the combination of
continuity, thermodynamic and hydrostatic
relationship is used to diagnose vertical
velocity (W) - Linearized Adiabatic Richardson equation for W'
46Future plan
- a) Community WRF 3DVAR
- Testing WRF/WRF 3DVAR cycling.
- Release new version (V2.0) in June 2004 and
support. - b) New observations
- IR/MW radiances (ATOVS, etc).
- Radar reflectivity.
- GPS refractivity.
- c) Ground temperature as new control variable
- d) NCAR real-time applications
- Antarctica Mesoscale Prediction System (AMPS).
- CONUS (impact of GPS data).
- e) Variational algorithm
- Synoptically-dependent background errors.
- First-guess at analysis time (FGAT), incremental
4DVAR?
47Conclusion
- For 3DVAR, CG method is ideally suited. The
efficiency comes from faster convergence and
memory saving. - Implementation of Outer loop is very useful
for effective data utilization and quality
control point of view. - In 3DVAR, it is not good idea to use interpolated
statistical parameters from the statistics which
is computed at much coarse resolution. However,
there is no problem when it is used the other
way. - For computation of 3DVAR input statistics, it
will always be good to have an initial estimate
of the lengthscale parameter with limited
dataset. It helps a lot in saving the
computational cost. - It will always be desirable to use flow dependent
BE.
48Acknowledgments
- Y.-H Kuo
- Dale M Barker
- Yong -R Guo
- Wei Huang
- Xiao Qingnong
- Mi-Seon Lee
- etc. at NCAR
49Thank you !