Factors Limiting the Current Skill of Forecasts: Flaws in Model and InitializationCenter for Ocean-Land-Atmosphere studies (COLA)George Mason University (GMU)Emialia K. JinClimate Test Bed Seminar Series4 February 2009

Model Flaws mean error, phase shift, different amplitude, and wrong seasonal cycle, etc

Flaws in the way the data is used data assimilation and initialization, chaos within non-linear dynamics of the coupled system

Inherent limits to predictability some times are more predictable than others, amplitude of SST anomalies with respect to ENSO phase

Gaps in the observing systemWhat is limiting the ENSO predictability?*

Model Flaws mean error, phase shift, different amplitude, and wrong seasonal cycle, etc

Flaws in the way the data is used data assimilation and initialization, chaos within non-linear dynamics of the coupled system

Inherent limits to predictability some times are more predictable than others, amplitude of SST anomalies with respect to ENSO phase

Gaps in the observing systemWhat is limiting the ENSO predictability?*

Flaws in Model: Two Flavors of ENSOand Its PredictabilityAuthors: Emilia K. Jin1, J.-S. Kug2, F.-F. Jin2, J.-J. Luo3, and T. Yamagata31George Mason Univ./COLA, 2University of Hawaii, 3FRCGC/JAMSTEC

Background and Objective Conventional El Nio: as a phenomenon in the equatorial Pacific Ocean characterized by a positive sea surface temperature departure form normal in the NINO 3.4 region greater than or equal in magnitude to 0.5C averaged over three consecutive months (NOAA)

Different flavors of El Nio Trans- Nio (Trenberth and Stepaniak, 2001), Dateline El Nio (Larkin and Harrison 2005), El Nio Modoki (Ashok et al. 2007 ), Non-canonical ENSO (Guan and NIgam, 2008), Warm pool El Nio (Kug et al. 2008), etc.: Even though there are differences, the distinctive interannual SST variation over the central Pacific which becomes more active in recent year and significantly different global impact form conventional El Nio are common features.

The transition mechanisms and dynamical structure of two-types of El Nino are significantly different (Kug et al. 2008).

In this study, CGCMs ability to predict the distinctive characteristics of two types of El Nio is investigated using two state-of-the-art CGCMs retrospective forecasts.*

Observed Two Types of El NinoKug et al., 2008NINO4NINO3Composite of SST (Contour) and Rainfall (Shaded)(1982/83, 1986/87, 1997/98)(1990/91, 1994/95, 2002/03, 2004/05)Normalized NINO3 and NINO4 SSTWarm-pool Cold-tongue Either NINO3 SST or NINO4 SST is greater than their standard deviation Mixed *

Observed DJF SST AnomaliesWarm-poolCold-tongueMixed*

Retrospective Forecast Initial condition cases of 12 calendar months are analyzed. As observational counterparts, OISST, CMAP rainfall, and NCEP/NCAR reanalysis data are used. Model and DatasetCourtesy of J.-J. Luo, T. Yamagata, and NCEP EMC In this study, forecast data is reconstructed with respect to lead time (monthly forecast composite).*

ModelLeadmonthEnsemble MemberPeriodAGCMOGCMFRCGCSINTEX1291982-2006ECHAM 4T106 L19OPA 8.22x2 L31NCEPCFS9151981-2006GFST62 L64MOM 31/3x5/8 L27

Observed DJF SST AnomaliesWarm-poolCold-tongueMixed*

Simulated SODJFM SST AnomaliesForecast lead month 1CFSSINTEXShading is for model, and contour is for observationWarm-poolCold-tongueWarm-poolCold-tongue*

Simulated DJF SST AnomaliesForecast lead month 6Shading is for model, and contour is for observationWarm-poolCold-tongueWarm-poolCold-tongueCFSSINTEXNote: loss of predictability in the Warm Pool El Nino cases*

Composite of SST Anomalies along the EquatorForecast lead month 7CFSSINTEXWarm-poolCold-tongueMixedtimetimeShading is for model bias, Contour is for observed compositeNote: Positive anomaly and negative bias in the Warm Pool and Cold Tongue*

Interannual Variability of NINO3 and NINO4JanFebMarAprMayJunJulAugSepOctObservedNovDecCFSSINTEXNINO3NINO3NINO4NINO4*

Scatter Diagram of Normalized DJF NINO3 vs. NINO4NINO3 IndexNINO4 IndexCFSSINTEXLead month 1Lead month 7*

Relationship between NINO3 and NINO4CFSSINTEXCOR=0.69*

Impact of Couplde Model Error on Predictability1st mode SEOF of SST (Low frequency mode)Obs.long run1st month9thmonth5thmonthNCEP CFSJJASINTEX-FMAMForecast lead monthCorrelationTemporal correlation of PC timeseries with observationPattern correlation of eigenvector with free long runSINTEX-FNCEP CFSSINTEX-FNCEP CFSCorrelation coefficients with respect to lead monthJin and Kinter, 2009Climate Dynamics With increase of the lead month, the forecast ENSO mode progressively approaches to the model intrinsic mode in free coupled run and departs from the observed. *

1st

0.950.833865

0.930.860195

0.910.85687

0.880.871776

0.860.88061

0.840.898792

0.810.912986

0.780.928475

0.760.939772

1st (3)

0.950.8338650.9889190.900412

0.930.8601950.9620380.90119

0.910.856870.940140.908415

0.880.8717760.9185530.9127

0.860.880610.878210.921347

0.840.8987920.8530820.927328

0.810.9129860.8151630.942434

0.780.9284750.7443490.957864

0.760.9397720.6616930.964079

1st (2)

0.833865

0.860195

0.85687

0.871776

0.88061

0.898792

0.912986

0.928475

0.939772

Sheet1

SEOF2NCEP

TS0.950.930.910.880.860.840.810.780.76

0.970.950.930.920.90.890.890.880.86

JJA0.8338650.8601950.856870.8717760.880610.8987920.9129860.9284750.939772

SINT

TS0.9889190.9620380.940140.9185530.878210.8530820.8151630.7443490.6616930.678910.689924

DJF

MAM0.9004120.901190.9084150.91270.9213470.9273280.9424340.9578640.9640790.9694560.969455

JJA0.8349130.8530380.8718570.8822310.8370280.7955190.810910.8166190.8514750.8930580.941543

SON

Sheet2

Sheet3

PRCGCSINTEXNCEPCFS Free long run forecast 202-year simulation Analyzing last 200 years(200-yr climatology) 52-year simulation Analyzing last 50 years(50-yr climatology) 1982-2004 period 9 members 12 calendar months ICs 12 months lead 1981-2003 period 15 members 12 calendar months ICs 9 months leadLuo et al. 2005Saha et al. 2006Model and DatasetCourtesy of J.-J. Luo, T. Yamagata, and K. Pegion*

Scatter Diagram of Normalized DJF NINO 3 vs. NINO 4From free long run of two CGCMsNINO3 IndexNINO4 IndexObs.CFSSINTEX1950-200550 years200 years0.690.820.86 Model Flaw: One Flavor of El NinoCOR=(NINO3, NINO4)Shading: Observed; models do not capture observed behavior*

Observed Composite of Precipitation AnomaliesWarm-poolCold-tongueForecast lead month 6CFSSINTEXObs.*

500 hPa GPH AnomaliesCFSSINTEXWarm-poolCold-tongueForecast lead month 6Obs.*

In two state-of-the-art CGCMs, the forecast skill of El Nio is investigated focusing on two flavors of El Nio: Warm-pool and cold-tongue.

As the lead month of forecast increases, the models fail to distinguish between two flavors of El Nio.

Both models have difficulties to reproduce the nonlinear relationship between NINO3 and NINO4 SST anomalies.

From the free long run, models tend to simulate the mixed mode of El Nino rather than warm-pool or cold-tongue El Nio.

Tropical precipitation and extratropical circulation anomalies associated with two flavors of El Nio are not captured by models.

Summary*

Model Flaws mean error, phase shift, different amplitude, and wrong seasonal cycle, etc

Flaws in the way the data is used data assimilation and initialization, chaos within non-linear dynamics of the coupled system

Inherent limits to predictability some times are more predictable than others, amplitude of SST anomalies with respect to ENSO phase

Gaps in the observing systemWhat is limiting the ENSO predictability?*

Flaws in the initialization: Impact of Ocean Initialization in CCSM3.0 Re-forecast ExperimentsAuthors: Emilia K. Jin12, B. Kirtman23, D.-H. Min3, K. Ashok4 and H-I. Jeong41George Mason Univ., 2COLA, 3Univ. of Miami/RSMAS, 4APEC Climate Center

Model and DatasetCourtesy of K. Ashok and H. Jeong (APEC Climate Center), and Ben Kirtman and Dug-Hong Min (Univ. of Miami)Retrospective Forecast CCSM 3.0: CAM3 T85L26 + POP 1.4 gx1v3 L40 Initial condition case of November are analyzed. As observational counterparts, OISST and CMAP rainfall are used. *

APCCCOLAInitializationAtmNCEP/NCAR Reanalysis 2(1, 3, 5th of November)Random conditions from CCSM 3.0 long run via AMIP (No observation)OceanSST-nudged scheme(Luo et al. 2005)GFDL MOM3 ODA (Rosati and Harrison, 2002)Member36Lead month712Period1982-20031982-1998ReferenceAshok et al. (2009)Kirtman and Min (2009)

Ocean InitializationSST-nudged schemeMOM3 ODAAPCC re-forecasts: 2-dimensional ocean initializationCOLS re-forecasts:3-dimensional ocean initializationGFDL MOM3 Ocean Data AssimilationGrid interpolation to POP 1.4 gx1v3 L406 atm. ICs*

Root-mean-square error of NINO IndicesForecast lead monthNov Dec Jan Feb Mar Apr May Nov Dec Jan Feb Mar Apr May *

Anomaly Correlation Coefficients of NINO IndicesForecast lead monthNov Dec Jan Feb Mar Apr May Nov Dec Jan Feb Mar Apr May *

Temporal Correlation Coefficients of SST AnomaliesDJF (1st season)MAM (2nd season)*

Temporal Correlation Coefficients of SST Anomalies*Differences of Correlation (APCC minus COLA)

Pattern Correlation of SST Anomalies Pattern Correlation CoefficientsYear1st season (lead month 2-4)(160-280E, 30S-30N) (40-160E, 30S-30N) APCCmember

COLAmember*

1984 SST anomalies along the EquatorForecast lead month*

ACC of NINO 3.4 Index (November IC) Colored dots denote14 CGCMs re-forecasts from DEMETER and APCC/CliPAS (Jin et al. 2008). Tier-1 MME: 10 CGCM multi-model ensemble except NASA, UH, APCC and COLA. Dynamic-Statistical Model: MME of modified CZ model, two statistical models and persistence Black line denotes persistence.

RMSE of NINO 3.4 Index (Nov IC) Colored dots denote14 CGCMs re-forecasts from DEMETER and APCC/CliPAS (Jin et al. 2008). Tier-1 MME: 10 CGCM multi-model ensemble except NASA, UH, APCC and COLA. Dynamic-Statistical Model: MME of modified CZ model, two statistical models and persistence Black line denotes persistence.

NINO3: Warm minus Cold composite SST anomaliesInfluence of Systematic Error on CFS Forecast Skill Warm composite (82/83, 86/87, 91/92, 97/98) - Cold composite (84/85, 88/89, 98/99, 99/00) Dashed lines denote composite for Hindcasts at different lead timesObservationCFS long run(Hindcast composite)Forecast lead monthCorrelationCORR. with respect to lead monthbased on 1st SEOF mode of SSTCorrelation between 1st PCs based on observation and hindcasts at different lead timesCorrelation between 1st PCs based on long run and hindcasts at different lead timesJin and Kinter, Climate Dynamics 2009 Model Flaw: Slow coupled dynamics of CGCM*

1st

0.950.833865

0.930.860195

0.910.85687

0.880.871776

0.860.88061

0.840.898792

0.810.912986

0.780.928475

0.760.939772

1st (2)

0.833865

0.860195

0.85687

0.871776

0.88061

0.898792

0.912986

0.928475

0.939772

Sheet1

SEOF2

0.950.930.910.880.860.840.810.780.76

0.970.950.930.920.90.890.890.880.86

0.8338650.8601950.856870.8717760.880610.8987920.9129860.9284750.939772

Sheet2

Sheet3

DJF SST ClimatologyA 138-year long run of CCSM3.0 Retrospective forecasts1st season of Nov IC

YearSimulated Nino34 Index (CCSM3.0)*

Composite Analysis of Nino34 IndexWarm minus Cold compositeEl Nino compositeLa Nina composite For CCSM3.0 free long run, events more than one standard deviation of DJF NINO 3 index is selected and 32 El Nio and 24 La Nia is picked up. Warm composite (82/83, 86/87, 91/92, 97/98) - Cold composite (84/85, 88/89, 98/99, 99/00)

DJF Correlation of SST with PrecipitationLocal SSTNINO3.4Re-Forecasts (Nov IC) Re-Forecasts (Nov IC) Free long run*

DJF Precipitation ClimatologyA 138-year long run of CCSM3.0 Retrospective forecasts1st season of Nov IC*

Temporal Correlation Coefficients of Precipitation AnomaliesDJFMAM*

In this study, the intercomparison of long-lead coupled prediction experiments has conducted focusing on the multiple sets of retrospective forecasts with two types of initial conditions using same CCSM3.0 CGCM: 2-dimensional ocean initialization (SST nudged scheme) and 3-dimensional ocean initialization (MOM3 Ocean Data Assimilation).

Focusing on ENSO forecast, the ocean initialization of the COLA re-forecasts causes the remarkable improvement of forecast skill in spite of the large systematic errors of model. On the other hand, the APCC re-forecasts has little advantages of ocean initialization and the influence of models systematic errors are quite large.

These results emphasize the importance of initialization of forecast model, in particular ocean component. Summary*

Emilia K. Jinkjin@cola.iges.org

*******

***********************************