1. Coupled Initialization Experiments in the COLA Anomaly Coupled Model
Ben Kirtman
George Mason University
and
Center for Ocean-Land-Atmosphere Studies
Missing Elements: (i) relating interactive ensemble (ic uncertainty), (ii) clarity regarding analogue approach, (iii) why would one want to do coupled initialization, (iv) more emphasis on why we need to nudge once the coupled modes are identified, (v)

2. Why Do Coupled Assimilation?
Better State Estimation (Air-Sea Fluxes For Example)
Assessing Observing Systems
Improve Coupled Model
Improve Coupled Predictions (Coupled Assimilation != Coupled Initialization)

3. COLA Perspective Emphasis Needs to be on Prediction
The Forecast Initialization Procedure Should Account for Two Problems:
Coupled Model Climate Does Not Agree with Observed Climate
Coupled Modes of the Coupled Model are Significantly Different From the Coupled Modes of Nature
Approaches: (i) Anomaly Initialization
(ii) Anomaly Coupling
(iii) Assimilate “Coupled Modes”
Maybe these two assertions are obvious. First assertion leads to an initialization shock which we will show. The way I view the second assertion is that the slow manifold of the coupled model is different from observations and the trick to initializing forecasts is to initialization the model slow manifold. Some progress is possible regarding both these issues without coupled initialization. However, seems easier to deal with these problems within the framework of coupled initialization.

4. Outline Motivation: Identifying Coupled Modes
Experimental Prediction: Where do we Stand?
Results From ARCS Project
Results From COLA Anomaly Coupled Model
Initialization Shock and Aliasing Coupled Modes
Identifying Coupled Modes
Mapping Observed Modes onto Model Modes
Coupled Initialization (Nudging) Model Coupled Modes

10. Is the Best Fit Best For Forecasting?
Give More Weight to the Model?
Ji et al. (1998) NCEP

11. Atmos Initialization
No Atmos Initialization
Persistence
Coupled forecast results suggest that some equilibrium between atmosphere and SST required.
Schneider et al., 1998; COLA
Coupled Model
Two points: (i) Anomaly Initialization and (ii) initialization of the atmosphere
Lead Time (months)

12. ODA
SST
Winds and SST
Rosati et al., 1997; GFDL

13. Chen et al., 1995, 1997; LDEO Wind Nudging within the Coupled System
Standard Zebiak-Cane Ocean: FSU Forcing
Wind Nudging within the Coupled System
Chen et al., 1995, 1997;
LDEO

16. COLA-IRI-GFDL-NCAR Collaboration on Global Coupled Prediction
OGCM: MOM3 with Agreed Upon Physics, High Res. 1.0x1.0x40 ÿ 1/3x1.0x40 Medium Res. 1.5x1.5x25 ÿ 1/2x1.5x25
AGCM: T42L18: ECHAM4.5 (IRI), COLA, CCM3
ICs: Ocean: GFDL ODA, 40 Cases: Jan. and Jul ; Atmosphere: Open but taken from long simulation with observed SST.
Coupling: No Flux Corrections or Anomaly Coupling

17. Problem: Coupled Model Climate Does Not Agree with Observed (ODA) Climate
Initialization Shock
May Degrade Forecast Skill

18. Analysis 80-99 Consensus COLA (High) COLA (Med) ECHAM (high)
ECHAM (Med)
CCM (Med)
Figure 13. Longitude-time sections of the evolution of the standard deviation of SST anomalies (C) in the equatorial Pacific as a function of month of year for analysis and retrospective forecasts: (a) Analysis , (b) Consensus, (c) COLA high, (d) COLA medium, (e) ECHAM high, (f) ECHAM medium, (g) CCM medium. Model results in left hand column are averaged over cases with January initial conditions, and in right hand column the averages are over the cases with July initial conditions. Time increases upwards in each panel

19. Analysis 80-99
Consensus
COLA (High)
COLA (Med)
ECHAM (High)
ECHAM (Med)
CCM (Med)

20. COLA
ECHAM
CCM
Figure 15. Longitude-time sections of the annual cycle of the standard deviation of SST anomalies (C; left column) and heat content anomalies (C; right column) in the equatorial Pacific from long simulations with the coupled models: (a) COLA medium, (b) ECHAM medium, (c) CCM medium. Time increases upwards in each panel

22. Initialization Shock
Problem: Coupled Model Climate Does Not Agree with ODA Climate
Fix the Coupled Model
Anomaly Initialization
Empirical Corrections (Anomaly Coupling)

23. +2 -9 Atmos Initialization No Atmos Initialization Persistence
Two points: (i) Anomaly Initialization and (ii) initialization of the atmosphere
Lead Time (months)

24. Empirically Correcting the Model System Error
3-D OGCM Flux Correction
Coupled Forecasts

28. OGCM Systematic error corrections

30. COLA Anomaly Coupled Model
OGCM GFDL MOM3
1.5x1.5x x1.5x25
AGCM COLA T42L18
ICs:
Ocean: GFDL ODA Interpolated to Above Resolution. 80 Cases: Jan, Apr, Jul, Oct
Atmosphere: Generated by Running with Observed SSTs. Six Atmospheric Perturbations Generated by Resetting the Calendar and Running for One Week
ÿ 6 x 80 Cases = 480 Hindcasts

44. Even With Anomaly Coupling the
Forecast Do Not Have the Right
Variability
Problem: Coupled Modes of the Coupled Model are Significantly Different From the Coupled Modes of Nature

45. Initializing the Model Coupled Modes
Identify Model Coupled Modes
Mapping Between Model Modes and Observed Modes
Inserting Mapped Modes into Coupled Model (Coupled Nudging)
Experimental Forecasts

46. Sub-Surface Temperature Nudging Procedure
Where time scale of the damping (a) is 10 days-1 at the surface and 50 days-1 at 250 meters. Linearly decreasing from the surface to 250 meters and constant below 250 meters.
Does Nudging Work? - Compare Different Data Assimilation Systems.

47. This shows that the “assimilation of the ODA” works, but that there are differences.

48. Forecasts from nudged data same as assimilation.

49. Note that this is the January 1997 Forecast
Note that this is the January 1997 Forecast. This result shows that the nudged assimilation produces similar forecasts as the original data.

51. Initializing the Model Coupled Modes
Identify Model Coupled Modes
Mapping Between Model Modes and Observed Modes
Inserting Mapped Modes into Coupled Model (Coupled Nudging)
Experimental Forecasts

52. Identifying Coupled Modes
Analogue Approach
Heat Content Analogues Don’t Work
Coupled Modes of Coupled Model are Not the Same as the Observations
Use Long Simulation to Identify “Best Fit” SST Evolution (12 Months)
Relate Analogue “Initial State” to Observed Initial State

53. Simulated Heat Content Does Not Match Observations Well
GFDL ODA
Anomaly Coupled Simulation RMSE Best Fit
Show that heat content analogues directly are not good-more evidence that model coupled modes are different from observed coupled models
Simulated Heat Content Does Not Match Observations Well

54. And Then Leads to a Relatively Poor Forecast
This figure demonstrates that a heat content best fit does not give good forecasts.
Forecast Initialized with GFDL ODA Jan1982

55. Best Fit Initial Heat Content Does Not Make A Good Analogue.
Model Modes are Different From Obs
What Do The Modes That Fit The Forecast Look Like?
Fit SST Evolution to Observations
What is Initial Heat Content?

56. However, show that we can find “best fit” coupled modes from the SST evolution

57. Show that these “best fit” modes are robust, but also provide some estimate of the uncertainty in the initial conditions.

58. Simulated Heat Content Does Not Match Observations Well
GFDL ODA
Anomaly Coupled Simulation RMSE Best Fit
Show that heat content analogues directly are not good-more evidence that model coupled modes are different from observed coupled models
Simulated Heat Content Does Not Match Observations Well

59. Initializing the Model Coupled Modes
Identify Model Coupled Modes
Mapping Between Model Modes and Observed Modes
Inserting Mapped Modes into Coupled Model (Coupled Nudging)
Experimental Forecasts

60. What do the Coupled Modes Look like.

63. Initializing the Model Coupled Modes
Identify Model Coupled Modes
Mapping Between Model Modes and Observed Modes
Inserting Mapped Modes into Coupled Model (Coupled Nudging)
Experimental Forecasts

64. Do actual coupled initialization/forecasts: need to talk about why.

70. Initializing the Model Coupled Modes
Identify Model Coupled Modes
Mapping Between Model Modes and Observed Modes
Inserting Mapped Modes into Coupled Model (Coupled Nudging)
Experimental Forecasts

73. Concluding Remarks Emphasized Coupled Initialization for Prediction
Recognize that Best Fit State May Not be Best for Prediction
Initialization Shock
Ad-Hoc Procedures for Identifying Model Coupled Modes, Mapping Observations Inserting them into Coupled Model
Coupled Modes Different From Observed Modes