Evaluation of GCM convection schemes via data assimilation: e'g' to study the MaddenJulian Oscillati

1
Evaluation of GCM convection schemes via data
assimilation e.g. to study the Madden-Julian
Oscillation in a model that doesnt have one

• Brian Mapes
• RSMAS, University of Miami
• with
• Julio Bacmeister
• (then NASA, now NCAR)

2
Why assimilation-based science?

• I. MJO is low frequency
• small Eulerian (local) rate of change
• ? many small processes (tendency terms), or small
  imbalances among bigger terms, could cause the
  observed changes
• many simple/toy single-effect demonstration
  models exist, but
• Physically comprehensive modeling is needed at
  this stage

3
Why assimilation-based science?

• II. Slow speed of motion
• even wrt weak tropical flows
• resting/uniform basic states questionable
• III. MJO large scale, yet confined...
• zonally, seasonally
• Geographically realistic modeling is needed at
  this stage

4
Why assimilation-based science?

• IV. But GCMs dont simulate it well...
• or would be solved long ago
• MJOs well-observed, well-resolved large scale
  structure needs to be brought into a models
  quantitative framework empirically by
  assimilation

5
New! MERRA reanalysis

• Modern Era (from 1979) Reanalysis
• for Research and Applications
• Budget datasets incl. analysis tendencies
• Uses GEOS-5 GCM (formerly NSIPP)

OBS precip, u850 GEOS5
Kim et al. 2009
6
use piecewise constant Zana(t) to make above
equations exactly true in each 6h time
interval while visiting analyzed states
exactly Replay analyzed wx
some analyzed state variable Z at some point
7
?Z/?t Zmodel Zana ?Z/?t
(Zdyn Zphys) Zana
Poor mans version ( interpretive aid)
Zana (Ztarget Z) /trelax
any analyzed variable Z at 6h intervals
time
8
Misses analysis (in direction toward model
attractor) by a skinch, but analysis is already
biased that way
(analyzed MJO a bit weak)
analyzed vs. Observed Z
time
9
Poor mans data assimilation nudge to analyses
?Z/?t Zmodel Zana ?Z/?t
(Zdyn Zphys) Zana

• Zana (Ztarget Z) /trelax
• Need to choose trelax
• Any small value will converge to same results
• Strong forcing (incl. q div) forces rainfall
  (M. Suarez), but can blow up model (B. Kirtman)
• Dodge trouble, and do science discriminate
  mechanisms, by using different trelax values for
  different variables (e.g. winds div vs. rot T,
  q)

10
Learning from analysis tendencies

• (?Z/?t)obs (Zdyn Zphys) Zana
• If state is kept accurate (LS flow gradients),
  then (?Z/?t)obs and advective terms Zdyn will be
  accurate
• and thus
• Zana ? -(error in Zphys)

11
Example 1 mean heating rate errorsdT/dtmoist
dT/dtana
100 500 mb 1000
(magnitudes much smaller)
15-30 December, 1992 (COARE)
High wavenumber in model T(p) profile disagrees
w/obs. so is fought by data assim WRONG
Strange stripe of moist-physics cooling at
700mb (melting at 10C, re-evap)
12
Example 2 MJO-related physics errorsjust do
more sophisticated Zana averaging (MJO phase
composites)

• Case studies (JFMA90, DJFM92)
• of 3D (height-dependent) fields (dT/dtana ,
  dq/dtana , etc)
• averaging Indian-Pacific sector longitudes
  together
• 27-year composite
• of various 2D (single level or vertical integral)
  datasets
• as a function of longitude

13

• Error lesson model convection scheme acts too
  deep (drying instead of moistening) in the
  leading edge of the MJO.

14
When MJO rain is over Indian Ocean, W. Pac.
atmosphere is observed to be moistening, but GCM
doesnt, so analysis tendency has to do it
15
Equatorial section of MJO phase 2 dqdt_ana
anomalies
Precip
16
9 8 7 6 5 4 3
2 1 0 back (W)
front (E)
Objective, unbiased-sample MJO mosaic of CloudSat
radar echo objects Riley and Mapes, in prep.
17
Physics lack of convective organization ? (a
whole nuther talk)
org 0.1
org 0.5
New plume ensemble approach (in prep)
18
OK, a better scheme (candidates)

• For schemes as mission-central as convection,
  evaluation has to be comprehensive
• Zana is a powerful guide to errors!
• Mean, MJO... but also diurnal, seasonal, ENSO,...
• simply save d()dt_ana, as well as state vars ()
• send into existing diagnostic plotting codes
• similar to (obs-model) analyses, but automatic
• (all data on same grid, etc.)

19
How to get Zana datasets? Nudge GCMs to worlds
great analyses

• Full blown raw-data assimilation is expenive,
  really...are we gonna beat EC, JMA, NCEP?
• Multiple GCMs nudged to multiple reanalyses
• Bracket/ estimate/ remove 2-model (anal. model
  eval. GCM) error interactions
• Commonalities teach us about nature, since all
  exercises share global obs. intensive assim.
• Differences play valuable secondary role of
  informing individual model improvement efforts
• (Shameless CPT proposal in communitys hands
  now...)

20
help wanted (avail. now)