comparingcamfvandcamse ina'perfectmodel'data ...€¦ · comparingcamfvandcamse...

Comparing CAM-‐FV and CAM-‐SE in a 'perfect model' data assimila:on environment

Kevin Raeder, Jeff Anderson, Tim Hoar, Nancy Collins

1

NCAR/CISL/IMAGe/DAReS

Background

2

Ø  The Data Assimila:on Research Testbed has assimilated many kinds of observa:ons into many versions of CAM (eulerian and finite volume ‘FV’)

Ø  Results are comparable to opera:onal forecas:ng centers. Ø  Extensions into CESM:

Then compare CAM-‐SE against CAM-‐FV in a ‘perfect model’ context; the first valida:on step of a data assimila:on implementa:on.

1-‐slide tangent: WACCM+DART results

+  POP and CLM +  Assimila:on into a coupled model: CAM+POP+CLM +  WACCM (Pedatella & Liu in WAWG) +  CAM-‐Chem (Barré in CCWG)

Ø  Now extended to spectral element CAM (‘SE’)

WACCM(FV)+DART: SABER and AURA

3

Pedatella, et al. “Ensemble Data Assimila:on in the Whole Atmosphere Community Climate Model” JGR, in press

Compared to SABER Temps

Compared to AURA Temps

Compared to synthe:c (“perfect model”) SABER Temps

No obs assimilated Tropospheric obs only Trop+SABER T obs Trop+SABER+AURA T obs

0 1 2 3 4 5 6−1.5

−1

−0.5

0

0.5

1

1.5

Perfect Model

4

Free run of the model is used as the Truth. Observa:ons of the evolving model state are taken periodically. Observa:onal error is added to the observed values to make them realis:c.

Details: •  “1-‐degree” CAM5 from cesm1_1_1 using the HadOI data ocean •  Observa:ons of T, U, and V on 15 levels at 600 approximately uniformly

distributed loca:ons = 27000 every 12 hours •  Observa:on error: random draws from N(0,1 K) and N(0,2 m/s)

T, U, or V

single CAM run = Truth

obs errors

:me

= synthe:c observa:on = true value

DART

5

Ensemble Kalman filter algorithm requires running an ensemble of equally likely forecasts, then using sta:s:cs of the ensemble and observa:ons to allow the observa:ons to guide the ensemble to a bemer descrip:on of the atmosphere.

****

h

tktk+2

y y

hh

Ensemble spread = uncertainty of the model state. It grows during forecast, shrinks as informa:on is added during assimila:on.

Neglected errors -‐> ensemble spread is too small (over confident). This can be fixed with “adap:ve infla:on”, which increases ensemble spread but retains the mean.

Observa:ons more than 3×spread away from ensemble mean are ignored as being erroneous.

“Iden:cal Twin” Experiments

6

Will the ensemble track the observa:ons (Truth + obs error) if the Truth model = the DA model = CAM-‐SE? Ini:al ensemble has a :ny spread (O(round-‐off) in T) centered around the True State. No infla:on (shouldn’t be needed).


7

Will the ensemble track the observa:ons (Truth + obs error) if the Truth model = the DA model = CAM-‐SE? Ini:al ensemble has a :ny spread (O(round-‐off) in T) centered around the True State. No infla:on.

08/01 08/06 08/11 08/16 08/21 08/26 08/31 09/050

0.05

0.1

0.15

0.2

0.25

0.3

0.35

month/day − Aug.01,2005 00:00:00 start

rmse

and

tota

lspre

ad

Northern HemisphereRADIOSONDE_TEMPERATURE @ 500 hPa

rmse pr=0.19712, po=0.17714 totalspread pr=0.20781, po=0.18088

rmsetotalspread

data file: /glade/scratch/raeder/SE_RUD_itwin1/Diag_Truth_def_reg_8.1−31/obs_diag_output.nc

0

30

60

90

120

150

180

210

# of

obs

: o=

poss

, +=u

sed


8

08/01 08/06 08/11 08/16 08/21 08/26 08/31 09/050

0.05

0.1

0.15

0.2

0.25

0.3

0.35

01−Aug−2005 06:00:01 through 01−Sep−2005 18:00:00

fore

cast

rmse

189

190

191

192

193

194

195

196

# of

obs

(o=p

oss,

+=u

sed)

Southern HemisphereRADIOSONDE_TEMPERATURE @ 500 hPa

FV identical twinSE identical twin

data file: /glade/p/work/raeder/Exp/FV−SE_perfect/FV_RUD_itwin2_obs_diag.ncdata file: /glade/p/work/raeder/Exp/FV−SE_perfect/SE_RUD_itwin1_obs_diag.nc

How does the CAM-‐SE assimila:on compare to corresponding CAM-‐FV assimila:on?

“Fraternal Twin” Experiments

9

Will the ensemble of CAM-‐SE track the observa:ons of CAM-‐FV? And vice versa. A harder test: s:ll start with :ny spread, but wrong model state.


10

Will the ensemble of CAM-‐SE track the observa:ons of CAM-‐FV? And vice versa. A harder test: s:ll start with :ny spread, but wrong model state.

08/01 08/06 08/11 08/16 08/21 08/26 08/31 09/051.5

1.6

1.7

1.8

1.9

2

2.1

2.2

2.3

01−Aug−2005 06:00:01 through 01−Sep−2005 18:00:00

fore

cast

rmse

0

30

60

90

120

150

180

210

240

# of

obs

(o=p

oss,

+=u

sed)


FV ftwinSE ftwin

data file: /glade/p/work/raeder/Exp/FV−SE_perfect/FV_RUD_ftwin1_obs_diag.ncdata file: /glade/p/work/raeder/Exp/FV−SE_perfect/SE_RUD_ftwin1_obs_diag.nc


11

Will the ensemble of CAM-‐SE track the observa:ons of CAM-‐FV? And vice versa.

No, because the ini:al ensemble spread is so small that the assimila:on ignores the large number of ‘outlier’ observa:ons taken from the other model.


12

Will the ensemble of CAM-‐SE track the observa:ons of CAM-‐FV? And vice versa.

No, because the ini:al ensemble spread is so small that the assimila:on ignores the large number of ‘outlier’ observa:ons taken from the other model.

What if we give the assimila:on a figh:ng chance by turning on the adap:ve infla:on? That may allow the spread to increase, which will allow more observa:ons to be assimilated.

Fraternal Twins with Infla:on

13

Obs from CAM-‐FV. Assimila:ng model is CAM-‐SE.

08/01 08/06 08/11 08/16 08/21 08/26 08/31 09/050

0.5

1

1.5

2

2.5

3


rmse

and

tota

lspr

ead



rmsetotalspread

data file: /glade/scratch/raeder/SE_RUD_ftwin_ainf1/Diag_Truth_def_reg_8.1−31/obs_diag_output.nc

0

40

80

120

160

200

240

# of

obs

: o=

poss

, +=u

sed


14


08/01 08/06 08/11 08/16 08/21 08/26 08/31 09/050

0.5

1

1.5

2

2.5

3


rmse

and

tota

lspr

ead



rmsetotalspread


0

40

80

120

160

200

240

# of

obs

: o=

poss

, +=u

sed


15


08/01 08/06 08/11 08/16 08/21 08/26 08/31 09/050

0.5

1

1.5

2

2.5

3


rmse

and

tota

lspr

ead



rmsetotalspread


0

40

80

120

160

200

240

# of

obs

: o=

poss

, +=u

sed


16

“SE twin infl” = Obs from CAM-‐FV. Assimila:ng model is CAM-‐SE.

08/01 08/06 08/11 08/16 08/21 08/26 08/31 09/050

0.5

1

1.5

2

2.5

3

01−Aug−2005 06:00:01 through 01−Sep−2005 06:00:00

fore

cast

rmse

0

40

80

120

160

200

240

# of

obs

(o=p

oss,

+=u

sed)


FV ftwin inflSE ftwin infl

data file: /glade/p/work/raeder/Exp/FV−SE_perfect/FV_RUD_ftwin_ainf2_obs_diag.ncdata file: /glade/p/work/raeder/Exp/FV−SE_perfect/SE_RUD_ftwin_ainf1_obs_diag.nc

“FV twin infl” = Obs from CAM-‐SE. Assimila:ng model is CAM-‐FV.

Fraternal Twins with Infla:on; Bias

17

“SE twin infl” = Obs from CAM-‐FV. Assimila:ng model is CAM-‐SE.

“FV twin infl” = Obs from CAM-‐SE. Assimila:ng model is CAM-‐FV.

08/01 08/06 08/11 08/16 08/21 08/26 08/31 09/05−2

−1.5

−1

−0.5

0

0.5

1

1.5

01−Aug−2005 06:00:01 through 01−Sep−2005 06:00:00

fore

cast

bia

s (m

odel

− o

bser

vatio

n)

0

6

12

18

24

30

36

42

# of

obs

(o=p

oss,

+=u

sed)

AntarcticRADIOSONDE_TEMPERATURE @ 500 hPa

FV ftwin inflSE ftwin infl

data file: /glade/p/work/raeder/Exp/FV−SE_perfect/FV_RUD_ftwin_ainf2_obs_diag.ncdata file: /glade/p/work/raeder/Exp/FV−SE_perfect/SE_RUD_ftwin_ainf1_obs_diag.nc

State Space Confirma:on

18

Compare 2 assimila:ons which use the same set of observa:ons: 1)  Iden:cal twin using CAM-‐FV for obs and assimila:on 2)  Fraternal twin using CAM-‐FV for obs, but CAM-‐SE for assimila:on

Specific humidity (kg/kg)

longitude (degrees_east)

latit

ude

(deg

rees

_nor

th)

raed

er S

at F

eb 8

16:

13:0

0 20

14

Q is not an observed variable, but has been brought to the observed model state.

1) 2) Ensemble mean

Ensemble mean

Conclusions

19

Ø  Data Assimila:on with DART and CAM-‐SE passes the ‘perfect model’ tests. Ø  The adap:ve infla:on algorithm effec:vely permits even ensembles with

small spread and wrong mean to be shited to the correct model state. Ø  Assimila:ons with CAM-‐SE appear comparable to those with CAM-‐FV,

although much more analysis could be done. Collabora:on? Ø  This tool can iden:fy some biases in a new model by comparison with known

biases in an old model, without a long run of the new model. Ø  A new model can be searched for problem areas by direct comparison to

observa:ons. Ø  We’re ready for a real observa:on test of CAM-‐SE+DART, including

oberva:ons that are not model variables.

For more informa:on: hmp://www.image.ucar.edu/DAReS/DART/ (no facebook, no twimer)

Resources

20

5 nodes/instance (ensemble member) (usually 3, but sporadic memory problems) ~2200 core hours/ 12 hour advance and assimilate 27000 obs. Wall clock; < 30 min/(forecast+assim cycle) + wai:ng between jobs ~200 Gb/ensemble restart set. Save infrequently Up to 10 Gb/assimila:on of DART output. So up to 0.5 Tb/month, user selectable.

comparing*camfv*and*camse* in*a'perfectmodel'*data ...€¦ · comparing*camfv*and*camse*...

Documents

comparingcamfvandcamse ina'perfectmodel'data ...€¦ · comparingcamfvandcamse...