[Mitgcm-support] Note: December 14, 2000
mitgcm-support at dev.mitgcm.org
mitgcm-support at dev.mitgcm.org
Wed Jul 9 15:56:05 EDT 2003
HyperNews notification of new message. See:
http://escher.JPL.NASA.GOV:2000/HyperNews/get/forums/wkmtg/76.html
=======================================================================
NOTES FROM THE ASSIMILATION MEETING : 14 December 2000
1) K-filter
------------
- A new series of filters are being run that
a) are based on the horizontal mapping scheme used in the MOM
assimilation (correlation distance equaling coarse grid
distance, zero noise-to-signal ratio)
b) dynamic modes of time-mean TS field of c20000630
c) different process noise, some calibrated by covariance
matching
- A new series of barotropic filter runs are being conducted (kf021
and kf022). The new runs differ in horizontal coarse-to-fine
(c2f) mapping from the old runs (kf019). The old mapping used
correlation distance 1.5 times coarse grid distance (9-deg by
9-deg) and noise-to-signal ratio 0.1 (11/17/99 posting[1].)
The new mapping (my bt2) uses correlation distance equaling
coarse grid distance (6-deg by 6-deg) and zero noise-to-signal
ratio. (The old mapping resulted in cross correlations among
coarse state that didn't seem right.)
Runs kf021 and kf022 differ in assumed process noise, Q; kf021
uses NCEP wind covariance mapped to the coarse grid. kf022 is
based on the same Q but multiplied at each coarse grid point by
the ratio of the asymptotic sea level simulation error using
kf021's Q to the prior BT error estimate of 3/27/00[2] (i.e.,
covariance matching). Both assume 24-hour decorrelation of
process noise.
The skill of these two are compared with that of our prior BT
filter kf019 in this figure[3]; Sea level residual variance
reduction from simulation to dynamic update. Positive means
better skill. kf019 is similar to kf021 (NCEP covariance as
process noise) but used the old c2f mapper. Runs kf021 and kf022
also do not assimilate data in shallow regions that kf019 used.
There was also a bug in kf019; all data error used was offset by
one grid point along the T/P track.
The skill of kf021 is comparable with kf019. The skill of kf022
is superior to kf021 (and kf019) especially in low latitudes
(Indian Ocean and tropical Pacific) thanks to the larger process
noise in these regions. kf022 seems less skillful than kf021 in
the Argentine basin.
- New runs of the baroclinic filter are being conducted (kf023).
The new runs differ from old ones (kf017, kf020) in dynamic
modes. The old runs (kf017 and kf020) were based on dynamic
modes of the TS field of the run with large vertical mixing used
to remove the large shearing (9/14/00[4]). The new series of runs
is based on dynamic modes of the time-mean TS field of c20000630.
(The transition matrix for all three was linearized around the
same time-mean TS field of c20000630, even though that used for
kf017&kf020 was based on different dynamic modes.) All three
baroclinic filter runs use mapping of a) (my bc4).
All baroclinic runs so far assume NCEP wind covariance mapped to
the coarse grid as process noise with 3-day decorrelation. The
first run (kf017) displayed a degradation north of the K-filter
cell (see plot[5] from 11/09/00[6]). This is thought to be due to a
too large an assumed process at the cell boundary. Indeed a run
(kf020) with tapered process noise at the north-south boundary
leads to smaller degradation as seen in this skill comparison[7]
over 1993. Tapering was done by a Gaussian weight centered on
the equator with a 6-degree (twice coarse-grid width) e-folding.
kf023 is the same as kf020 but uses c20000630 dynamic modes and
also does not use data in shallow regions that were assimilated
in kf017 & kf020. The skill of kf023 is comparable with kf020
but the degradation off Taiwan is smaller than kf020. Maybe the
remaining degradation is due to excess process noise at the west
boundary?
Calibration of process noise used in kf023 is ongoing (covariance
matching).
2) Other
---------
- Last week's K-filter comparison (kf017) with TAO velocity was
incorrect. There was a time-offset (~60-days) in processed TAO
velocity. The analysis has been corrected[8]. K-filter does show
improvement over the simulation.
- Doubling algorithm on alhena is being packaged into a utility.
- Zhangfan is assessing the computational environment at ames.
Because of the max time-limit for the available queues are short,
we may need to fit the adjoint in their special queue.
- The 3-day adjoint is working OK with KPP but without GM. (Don't
know what will happen with GM yet.) The adjoint takes much
longer with KPP than without. Now the 30-day adjoint takes 2.2
hours and doesn't fit in the short64 queue. The 30-day will be
tested in long64 and if that works, a 90-day run will be tested
with test64.
=======================================================================
-----------------------------------------------------------------------
[1] http://escher.JPL.NASA.GOV:2000/HyperNews/get/forums/assim/70.html
[2] http://escher.JPL.NASA.GOV:2000/HyperNews/get/forums/assim/83.html
[3] http://escher.JPL.NASA.GOV:2000/hosts/escher/escher2/medea/if/forum/t73f3.ps
[4] http://escher.JPL.NASA.GOV:2000/HyperNews/get/forums/wkmtg/64.html
[5] http://escher.JPL.NASA.GOV:2000/hosts/escher/escher2/medea/if/forum/t71e_1993.ps
[6] http://escher.JPL.NASA.GOV:2000/HyperNews/get/forums/wkmtg/73.html
[7] http://escher.JPL.NASA.GOV:2000/hosts/escher/escher2/medea/if/forum/t73f4_1993.ps
[8] http://escher.JPL.NASA.GOV:2000/HyperNews/get/forums/assim/106/2/2/1.html
More information about the MITgcm-support
mailing list