j. l. anderson, g. romine, k. raeder, h. liu, n. collins ... · the data assimilation research...

The Data Assimilation Research Testbed: New Algorithms and ApplicationsJ. L. Anderson, G. Romine, K. Raeder, H. Liu, N. Collins, T. Hoar

National Center for Atmospheric Research – Institute for Mathematics Applied to GeosciencesBoulder, Colorado, USA

DART: The Data Assimilation Research Testbed

NCAR’s Data Assimilation Research Testbed (DART) provides a cutting-edge facil-ity to perform ensemble data assimilation with support for a wide range of models,conventional and novel observations, algorithms, and diagnostic tools. Severalresearch projects are presented that demonstrate the effectiveness and flexibil-ity of the DART system to produce high-quality products. Our email address isdart@ucar.edu, the DART site is www.image.ucar.edu/DAReS/DART

Rank Histogram filtering algorithm

DART continues to add tools to improve ensemble filter performance.• An inexpensive sampling error correction algorithm that automatically local-

izes observation impacts and can reduce error in large geophysical applica-tions [JLABergen]Localization ! as function of N and sample correlation .

This is precomputed for table look-up.

19

!

ˆ r Test II: Low-order dry dynamical core

60

Localization for mid-level V ob. on U state variables (N=80).

Local four maxima, but mostly still gaussian. Figure 1: Left: Localization α based solely on ensemble size (N) and samplecorrelation. Right: the resulting α for mid-level V obs on the U state variable of a’perfect model’ experiment using 80 members of a dry dynamical core.

• A non-gaussian observation space update that works with arbitrary prior andobservational error distributions [JLA2010]

2.5 2 1.5 1 0.5 0 0.5 1 1.5 2

0

0.5

1

1.5

PriorProb

abilit

y D

ensi

ty

Posterior

EAKF Posterior

Figure 2: Non-gaussian ensemble update algorithm: A continuous prior distri-bution (green histogram) and an observation likelihood (red curve) are combinedto give a continuous posterior (blue histogram). Green asterisks mark the priorensemble while blue asterisks mark the posterior from the new filter (above prior)and an ensemble adjustment filter (EAKF; below). The new filter eliminates thespurious ensemble members near -0.5 in the EAKF.

Severe Convective Storms on the High Plains

Considerable challenges remain in understanding and predicting the initiation andevolution of high impact convective weather events, particularly in the vicinity ofcomplex terrain. Significant opportunities for improving guidance from storm-scaleensemble forecasts should exist where atmospheric conditions are rapidly evolvingsuch as during convective initiation. A DART/WRF (Weather Research and Fore-casting model) ensemble assimilation system has been developed with both meso-and storm-scale probabilistic analyses and forecasts to demonstrate current andfuture capabilities for storm-scale prediction.

The period of 4-17 June 2009 is examined on a CONUS domain which providesinitial and boundary conditions for regional storm-scale analyses centered near theColorado Front Range (Figure 3).

Figure 3: Surface elevation (shading; m MSL) in the WRF model in the (a) CONUSdomain (4x = 15 km) and (b) Front Range domain (4x = 3 km) along with loca-tions of radars where observations were drawn for radar assimilation experiments.

Details about the meso-scale assimilation:• 3-hourly analysis• 50 ensemble members• error statistics comparable to other state-of-the-art mesoscale forecast systems

Details about the storm-scale assimilation:• 6-hour forecasts every three hours from 15-00 UTC• continuous assimilation of conventional Doppler radar observations for one hour

precedes forecasts

Right: Prior and posteriorobservation-space diagnostics[m/s; root-mean square of theinnovations, ensemble spread,and mean of forecast/analysisminus observations shown in red,blue, and green, respectively]for Doppler velocity during a 1-hassimilation window starting at 23UTC 11 June, 2009. Times (frac-tional day) are indicated on thebottom. Assimilated radial velocityobservations fit the bred, down-scaled mesoscale backgroundwith low RMS and bias.

Control deterministic forecasts, drawn from the meso-scale analysis member with a‘best fit’ to the ensemble mean at the start of the forecast period, are also spawnedduring each ensemble forecast period. In initial tests, radar assimilation leads toimproved ensemble storm-scale forecasts (Figure 4) of weather hazards and betterquantitative precipitation forecasts.

Figure 4: Ensemble probability (%; shading) of updraft helicity exceeding 75 m2/s2

during the forecast period 00→06 UTC 12 June, 2009 for a portion of the storm-scale domain. SPC preliminary storm reports (hail in green, strong winds in blue)are also shown. Left - control forecast (no radar assimilation), Right - forecastfollowing 21 cycles of Doppler velocity and ‘clear air’ reflectivity observation assim-ilation.

Atmosphere/Ocean Data Assimilation

Decadal forecasts of the ocean are being explored with a loosely coupled data as-similation system in which there is no communication from the ocean assimilationto the atmosphere. A fully-coupled system is planned.

!"#$#

%&'(')*+&+#,-./#'#0.+1)*#2.3!)&4#56&'(781/.+!9&-&#:*+1&/#

;'4)&*#<#

=2>?75@$##

%&')#::A+#

5B+#3+&4#B*#

=28%7=2>?#

-&'(')*+&+#C8%A7280#

8++D/D)'E.(#

:*+1&/#

2>:0F#6.3!)&-#

9D+1.-*#G)&+H#

'1/.+!9&-D6#,.-6D("#?5?7C8%A#

'++D/D)'E.(#

+*+1&/#I.-)4#56&'(#

C'1'B'+&#$JJK##

5B+&-L'E.(+#

280K#D(DE')#G)&+M#!.+1&-D.-#&(+&/B)&#/&'(#.,#+1'1&#L'-D'B)&+#

################################!-D.-#&(+&/B)&#/&'(#.,#'))#.19&-#L'-D'B)&+#

2N0#-&+1'-1#G)&+M#!-D.-#&(+&/B)&#/&'(#.,#'))#L'-D'B)&+#

2@2>#-&+1'-1#G)&+M#!-D.-#&(+&/B)&#/&'(#.,#'))#L'-D'B)&+#

280#+1'1&#L'-D'B)&+#O#?:P#AP#QP#RP#SP#2NCN@SP#2NC@2>#

?.+1&-D.-#O#L')3&+#'T&-#19&#'++D/D)'E.(#.,#.B+&-L'E.(+#'1#19'1#E/&#

?-D.-#O#L')3&+#B&,.-&#'++D/D)'E.(#UB31#'T&-#'#+9.-1#,.-&6'+1V#

?5?#'(')*+&+H#

1&/!&-'13-&P#

+')D(D1*P##

L&).6DE&+P#

+3-,'6&#9&D"91#

$W#X3(&#$JFJ#Figure 5: A schematic describing the coupling between the atmospheric andoceanic data assimilations. DART/CAM is run independently of the ocean assimi-lation and provides an ensemble of equally likely atmospheric forcings for POP.

A pair of coupled (ocean-atmosphere-land) forecasts are started at 2000 and willbe integrated to 2030 for the IPCC AR5 using the Community Climate SystemModel version 4 (CCSM4). Initial conditions come from the ends of a forced ocean-ice hindcast, and a two year ocean reanalysis from DART/POP. The initial condi-tions from DART substantially reduce model error relative to the true state of theocean, as represented by DART/POP analyses, for up to 2 years, and provideinformation about the rate of relaxation to the model’s climatology.

!"

#"

$%&'!"()*)+$,'-."/012324")50361789":96724"$8936789;"

80 oW

70 o

W 60

oW 50

oW 40

oW 30

oW 20

oW

10o W

0o

20 oN

30 oN

40 oN

50 oN

60 oN

Ocean State Estimate w/ assimilation

Jan 2000

oC

0

5

10

15

20

25

80 oW

70 o

W 60

oW 50

oW 40

oW 30

oW 20

oW

10o W

0o

20 oN

30 oN

40 oN

50 oN

60 oN

CCSM4 forecast from ocean state w/o assimilation

Jan 2000

oC

0

5

10

15

20

25

!!"# $%!

!!&# $%!!!'#

$%! !!(#

$%! !!)#

$%! !!*#

$%!

!!+#$%!

!!,#$%!

!!!#$ !!

!!+# $-!

!!*# $-!

!!)# $-!

!!(# $-!

!!'# $-!

!

!

Difference from ocean state:

CCSM4 forecast from ocean state w/o assimilation

Jan 2000

oC

!,#

!"

!'

!)

!+

#

+

)

'

"

,#

①! ,924<;6;"(2=>"?"69"@8A"#BCD."E58D"#"<025"

2;;6D642789"8E"F*/"CG"?"293"&"8H;I"

E85103"H<"/,J?K$,'"2@D8;L"2924<;6;"

09;0DH40L"

②! :$;"E58D"2"M69312;@"E85103"H<"$*J%#"

M6;@856124";N5E210"ONP0;L"

③! /6Q050910"8E"#"E58D"!L"

!"

#" R"

Figure 6: Average upper ocean temperature (0-250m) in the North Atlantic (1).These represent the initial conditions for the DART/POP forecast. The hindcast ini-tial conditions (2) exhibit a large cold bias, relative to the analyses, where the GulfStream fails to turn northward near 45W and a moderate warm bias in the easternand northern parts of the domain.

The forecast from hindcast ICs maintains the errors seen in Figure 6, while the fore-cast from DART/POP ICs (Figure 7; upper right) has much smaller errors (lowerpanels). Note that the DART/POP forecast is starting to degrade in the area ofhighest temperature gradient near 50◦W and 40◦N .

!"#$%&%'($")#*+*%,+(%-"%)(./.0"123%

80 oW

70 o

W 60

oW 50

oW 40

oW 30

oW 20

oW

10o W

0o

20 oN

30 oN

40 oN

50 oN

60 oN

Ocean State Estimate w/ assimilation

Dec 2001

oC

0

5

10

15

20

25

80 oW

70 o

W 60

oW 50

oW 40

oW 30

oW 20

oW

10o W

0o

20 oN

30 oN

40 oN

50 oN

60 oN

CCSM4 forecast from ocean state w/ assimilation

Dec 2001

oC

0

5

10

15

20

25

!!"# $%!

!!&# $%!!!'#

$%! !!(#

$%! !!)#

$%! !!*#

$%!

!!+#$%!

!!,#$%!

!!!#$ !!

!!+# $-!

!!*# $-!

!!)# $-!

!!(# $-!

!!'# $-!

!

!

Difference from ocean state:

CCSM4 forecast from ocean state w/ assimilation

Dec 2001

oC

!,#

!"

!'

!)

!+

#

+

)

'

"

,#

!"#$%&&& $ $ $ $ $ !"#$%&&' $ $ $ $ $&

&(%)

&()

&(*)

'

'(%)

'()

+,-$.//0/$1#$20/.3"45$026178"51597.!:58"#513$9;;./$03."#$5.6;./"59/.

0<

=#151"81>.7$?@$"441618"510#

=#151"81>.7$?@0$"441618"510#

4"."5+%('%678$(9"1%:;*%<#*+*%#+%<"#*+%&%="#$*>%

?+@"$%$"*0<+*%678<=%-"."5+%(0+%+(%A%="#$*>%

Figure 7: The ocean state as represented by DART/POP analysis for December2001 (upper left) is the benchmark for the 2 forecasts. The time series (lower left)show the area weighted RMS error of the average upper 250m temperature for theocean-ice hindcast simulation (black curve) and the DART/POP IC simulation (bluecurve).

Satellite Soundings and Tropical Storm Forecasts

The DART/WRF (Weather Research and Forecasting model) system has beenused to quantify the impact of AIRS water vapor observations on the forecast of in-tensity and track of Super Typhoon Sinlaku. Sinlaku formed at 06 UTC 8 Sep 2008over the western Pacific and became a Super typhoon-4 at 18 UTC 10 Sep. Weexamine the hypothesis that assimilating AIRS Q data for 2 days prior to the gene-sis can improve the analsyes and forecasts of the initial intensification. We use theAIRS Q data as processed by CIMSS – the Cooperative Institute for MeteorologicalSatellite Studies at the University of Wisconsin-Madison.

These simulations use a standard set of WRF parameterizations and settings:• 45 km resolution, 45 levels• 32 ensemble members•DART adaptive inflation during assimilation• Initial ensemble mean conditions from NCEP 1 degree global analysis; initial

ensemble generated with 3DVar perturbations•Cycling analysis every 2 hours from 00 UTC 6 Sep to 12UTC 9 Sep• FCST run: Ensemble forecasts from the initial conditions; no observations•Only-Q run: Assimilation of only AIRS Q soundings

Figure 8: Left: daily AIRS Q coverage for Sep 6-9, 2008 (a-d, respectively). Theneed is to spread the information to the unobserved position of the cyclone (the’X’). Right: The daily analysis increments for 7 Sep 2008. The AIRS Q observa-tions clearly provide information about Q, temperature, and winds.

Figure 9: Left: Locations of the radiosondes withheld for verification. Right: 2-hourforecast fits to the radiosonde observations for 6–9, Sep 2008. The solid lines arefor the RMSE of the ”no assimilation” experiment, the dashed lines demonstratethe effect of assimilating the AIRS Q soundings .

A more rigorous test is to compare to an assimilation that uses other observations.

•CTL run: Assimilate radiosonde, cloud winds, aircraft data, surface pressuredata

• AIRS-Q run: Same as CTL plus AIRS Q soundings

•No artificial cyclone vortex bogus data is used.

• Each run started at 00 UTC 6 Sep

• cycle every two hours till 12 UTC Sep 12

• launch a forecast and evaluate every 6 hours

Left: Mean of the ensembleforecasts from 12 UTC 9 Sep.Top: The intensity of Sinlaku.The dashed line represents theeffect of using the AIRS Q, animprovement over the experi-ment that did not use the AIRSdata. Bottom: The AIRS Q ob-servations clearly improved thetrack forecast.

References

[DART 09] J. Anderson, T. Hoar, K. Raeder, H. Liu, N. Collins, R. Torn, andA. Arellano, 2009: The Data Assimilation Research Testbed: A Com-munity Facility. BAMS 90 No.9 pp. 1283–1296

[JLA2010] Anderson, J. L., 2010: A non-gaussian ensemble filter update for dataassimilation. Mon. Wea. Rev. 138, pp. 4186-4198

[JLABergen] Anderson, J. L., 2010: Reducing the Impact of Sampling Er-rors in Ensemble Filters. Presented at The 5th International Work-shop on Ensemble Kalman Filter for Model Updating, Bergen, Norwaywww.image.ucar.edu/pub/DART/2010/2010 Bergen JLA talk.pdf

[WOD98] Levitus, S., T. Boyer, M. Conkwright, D. Johnson, T. OBrien, J. Antonov,C. Stephens, and R. Gelfeld, 1998: Introduction, vol. 1. World OceanDatabase 1998, NOAA Atlas NESDIS 18, 346 pp.

The National Center for Atmospheric Research is sponsoredby the National Science Foundation.