MOM4 Tangentlinear and Adjoint Project

1
MOM4 Tangent-linear and Adjoint Project
GFDL Tony Rosati, Shaoqing Zhang, Matt
Harrison, Andrew Wittenberg
JPL Benny Cheng, Tony Lee, Ichiro
Fukumori
Harvard Jake Gebbie, Eli Tziperman
2
ECMWF and 4D-VAR
4D-VAR with Adjoint Model is ECMWF's preferred
method for operational NWP
3
ECCO Consortium
Given 8 million observations Ocean GCM 6
million uncertain parameters
Use Adjoint method to solve
Result Ocean product with a complete,
dynamically-consistent physical picture of
circulation 1992-2000.
4
Data Assimilation and ENSO Prediction
Lead time
0 mon.
3 mon.
6 mon.
NINO3
9 mon.
12 mon.
time

• An improved data assimilation scheme to a simple
  nudging resulted in significantly
  improved ENSO forecasts.
• With a more-realistic model and 4D-VAR, expect
  potential for even greater forecast improvements.

Chen, Zebiak, Cane, et al., 1995, 2005
5
An Efficient Sensitivity Calculation
Galanti and Tziperman, 2002 Adjoint-calculated
Sensitivity w/ MOM
Adjoint sensitivities show a novel teleconnection
between ENSO and the extratropics via
baroclinically-unstable Rossby waves.
6
Progress to date

• Tracer eqns. , advection/diffusion
• External-mode (Harvard) and surface forcing (JPL)
• Parallelized adjoint (JPL)
• Tangent-linear (GFDL) and
• Adjoint (Harvard) for all of the above
• Cost and control packages (Harvard)
• CVS repository for
  collaboration (Harv)

7
JPL Test Results
16 processors, tripolar grid

• 1-month adjoint
• gradient check
• Running at JPL
• Correct to 110,000,000
• Cost function is global sum (no hidden errors)

Sensitivity to SST
8
JPL Test Results
16 processors, tripolar grid

• 1-month adjoint
• gradient check
• Running at JPL
• Correct to 110,000,000
• Cost function is global sum (no hidden errors)

Sensitivity to SST
9
Harvard Test Results
FWD CASE
ADJ CASE
Tracer perturbation
Sensitivity to tracer pert.
30-day adjoint integration of an advective/diffusi
ve version of MOM4 at Harvard Sensitivity
is transported upstream! The MOM4 adjoint
gives physically-sensible results.
T(0)
T(F)
ConstantVelocity
T(FINAL)
T(0)
10
Next steps and anticipated timetable

• Internal mode (baroclinic) dynamics (2 mons.)
  (Harv,JPL)
• Optimization, descent-method search (4 mons.)
  (JPL,GFDL)
• Physics packages, isopycnal mixing, KPP/GM

Goals

• ENSO sensitivity (within year) (Harvard)
• Adjoint of statistical atmosphere (within year)
  (GFDL, Hrv.)
• Adjoint code distributed with MOM4
• ENSO prediction (started within next year)

11
The MOM4 Tangent-linear and Adjoint

• Allows the MOM4 community to use the experience
  gained by ECMWF and ECCO, esp. 4D-VAR, climate
  prediction and re-analysis
• Opens up new avenues of research through linear
  algebra and control theory techniques, e.g.,
  sensitivity studies, SVD, stochastic optimals

All the pieces are in place, only the baroclinic
dynamics need to be adjointed before an ENSO
sensitivity study can be done.