Status and Challenges of Dynamic Seasonal Prediction Bin Wang University of Hawaii Acknowledge contributions from co-Authors: S.-I.An, R. Wu, I.-S. Kang, J. Shukla, X. Fu, Q. Ding, T. Li, F. Doblas-Reyes, J-Y Lee, CliPAS Team: C.-K Park, A. Kumar, B. Kirtman, J.Kinter, Emilia K. Jin, J. Schemm, A. Rosati, N.-C. Lau, W. Stern, T. N. Krishnamurti, S.Coke, Z. S. Schubert, K.-M. Lau, and M. Sawrez, S. Schubert, J.-S. Kug, P. Liu, and X. Fu Summer School BNU, Beijing 5 Aug. 2006

Current Status Review Multi-Model Ensemble (MME) Seasonal prediction of surface temperature and precipitation Can MME capture the dominant modes of interannual variability of A-AM?

Theoretical Basis for Seasonal prediction Conceptual break through on ENSO (Bjerkness 1966, 1999) Sources of seasonal prediction (Charney and Shukla 1981). Climate prediction as an initial value problem (Palmer 2000)

Dynamic Seasonal Prediction Prediction of ENSO with coupled models of Intermediate complexity (Cane and Zebiak 1985) NCEP and ECMWF started to produce operational ensemble forecast using AGCMs (Tracton and Kalnay 1993; Palmer et al. 1993) and two-tier system. Prediction of ENSO is essentially an initial value problem using CGCMs (Palmer 2000). The one-tier system. A number of meteorological centers worldwide have implemented routine dynamical seasonal predictions using one-tier system (Palmer et al. 2004; Saha et al. 2005). The multi-model ensemble (MME) prediction has been developed to alleviate uncertainties arising from model parameterizations and prediction errors. MME prediction performs better than any single-model component in general for both the two-tier systems (Palmer et al. 2000; Shukla et al. 2000; Barnston et al. 2003; Krishnamurti et al. 1999) and one-tier systems (Doblas-Reyes et al. 2005).

Mulit-Model Ensemble Prediction Development of a European Multi-Model Ensemble System for Seasonal to Inter-Annual Prediction (DEMETER) project (Palmer et al. 2004). Seven one-tier models. The Climate Prediction and its Application to Society (CliPAS) project is in support of Asian- Pacific Economic Cooperation (APEC) Climate Center (APCC). Five one-tier systems and Five two-tier systems.

 Development of European Multimodel Ensemble system for seasonal-to-interannual prediction  One-tier prediction system using CGCM  9 ensemble members of 7 models  forecast InstituteAGCM Resolutio n OGCMResolution Atmosphere initial conditions Ensemble generation CERFACE ARPEGE T63 31 Levels OPA x Levels ERA-40 Windstress and SST perturbations ECMWF IFS T95 40 Levels HOPE-E1.4x Levels ERA-40 Windstress and SST perturbations INGV ECHAM-4 T42 19 Levels OPA x Levels Coupled AMIP-type experiment Windstress and SST perturbations LODYC IFS T95 40 Levels OPA x Levels ERA-40 Windstress and SST perturbations Meteo-France ARPEGET63 31 Levels OPA GPx152 GP 31 Levels ERA-40 Windstress and SST perturbations MPI ECHAM-5T42 19 Levels MPI-OM1 2.5x Levels Coupled run relaxed to observed SSTs Atmospheric conditions from the coupled initialization run (lagged method) UK Met Office HadAM3 2.5x Levels GloSea OGCM based on HadCM3 1.25x Levels ERA-40 Windstress and SST perturbations DEMETER model Specification

InstituteAGCMResolution Ensemble Member Reference FSU FSUGCMT63 L2710 Cocke, S. and T.E. LaRow (2000) GFDL AM22 o lat x 2.5 o lon L2410 Anderson et al. (2004) SNU/KMA GCPST63 L216 Kang et al. (2004) UH CAM2T42 L2610 Liu et al. (2005) UH ECHAM4 T31 L1910 Roeckner et al. (1996) InstituteAGCMResolutionOGCMResolution Ensemble Member Reference *FRCGC ECHAM4T106 L19OPA 8.22 o cos(lat)x2 o lon L319Luo et al. (2005) NASA NSIPP1 2 o lat x 2.5 o lon L34 Poseidon V4 1/3 o lat x 5/8 o lon L273 Vintzileos et al. (2005) *NCEP GFST62 L64MOM31/3 o lat x 1 o lon L4015Saha et al. (2005) *SNU SNUT42 L21MOM2.21/3 o lat x 1 o lon L326Kug et al. (2005) UH ECHAM4T31 L19UH Ocean1 o lat x 2 o lon L210Fu and Wang (2001) APCC 1-Tier Models APCC 2-Tier Models APCC/CliPAS model Specifications

Superior to the persistence except JJA SE Africa and SWIO. Two MMEs similar. DEMETER better over the Asia, but CliPAS is better over the North America-North Atlantic. Tropical Pacific: JJA better in the region between 10N and 20N (north of the ITCZ) Tropical IO: the local summer surface temperature is better predicted than local winter season. Tropical Atlantic better during the local winter. Role fo ENSO. Poleward of 20N and 20S, higher in local winter than local summer: North Pacific and Atlantic, SPCZ. Role of westerly Jet. JJA and along 35N-50N Temporal Correlation / 2m Air Temperature

Temporal Correlation / Precipitation Two MMEs similar. DEMETER MME better in DJF over the East Asia and eastern tropical Africa. The high skills : the tropical Pacific between 10S and 20N, MC, and the northeast Brazil and the equatorial Atlantic Ocean Western Hemisphere tropics is better. Hoarse show low-skill region in the EH. Land regions are lacking skills. During DJF ENSO impacts extends to Land. DJF is better than JJA. Particularly in (a) the subtropical NP and NA (20-40N); (b) EIO; (c) SPCZ; (d) the northern South America, (e) Southern Africa; (f) Mexico and southern United States, and (g) Southeast Asia. However, over Southeast South America-southwest AO (20S-35S, 10W-50W) and east Australia, JJA (local winter) better. JJA: midlatitude wavelike patterns of moderate skill is notable.

Figure 11. Scatter diagram of anomaly pattern correlation between precipitation prediction over ENSO and A-AM regions using CliPAS (a, c) and DEMETER (b, d) system in summer and winter, respectively. Prediction Skill over ENSO versus A-AM region Precipitation

Figure 9. MME effective index and 2-D diagram of pattern correlation and normalized RMSE for precipitation prediction in summer (a) and winter (b). Black color indicates the skill over Monsoon region and grey color indicates that over ENSO region. Filled round (square) dot and contoured round (square) dot represent the averaged skill of models and MME skill in CliPAS (DEMETER) system, respectively. MME Effective Index / Precipitation

Figure 10. Range of pattern correlation coefficients between the observed and predicted precipitation (a, b) and temperature (c, d) over monsoon (dark color) and ENSO (light color) regions using different number of the model being composed in blended CliPAS and DEMETER Tier-1 predictions in JJA and DJF. Marks indicate the average value of the pattern correlation coefficients for various combinations of the models composed. Upper (lower) cross mark indicates the best (worst) model composite for each number of the model being composed.

Can Multi-Model Ensemble Capture the Dominant Modes of A-AM Variability?

Questions 1. What are the essential features of the leading modes of IAV of A-AM? 2. How well do the DEMETER and CliPASS 10 coupled climate models capture these leading modes?

Methodology: Seasona-Reliant EOF (Wang and An 2005, GRL) Physical consideration Anomalous climate (ENSO is only an example) is often regulated by the seasonal march of the solar radiation forcing and the resultant climatological annual cycles. We propose Season-reliant Empirical Orthogonal Function (S- EOF) analysis to detect seasonal evolving major modes of climate variability. Method of S-EOF Analysis Basic Idea: Search for seasonally evolving patterns from year to year. Keys: Examine a set of sequential seasonal anomalies for one year. Coding: Constructing a covariance/correlation matrix by treating the given set of seasonal sequence as one time step (year). Interpretation: The obtained spatial patterns for each S-EOF mode describes evolving seasonal anomalies in a given year. These patterns share the same yearly value in the corresponding PC.

Fractional Variance of the first 6 S-EOF modes Percentage variance (%) explained by the first six S-EOF modes of seasonal precipitation obtained from (a) CMAP, (b) GPCP, (c) ECMWF reanalysis, (d) NCEP reanalysis 2, (e) CliPAS MME prediction and (f) DEMETER MME prediction. The error bars represent one standard deviation of the sampling errors. OBS Reanalysis MME prediction

S-EOF 1 of CMAP: 30% DEMETER and CliPAS MMEs capture spatial patterns of QB &QQ with PCC of ~0.8 for SEOF1 and for SEOF2 S-EOF 2 of CMAP: 13% CMAPCliPASDEMETER CMAPCliPASDEMETER DEMETER and CliPAS one-tier MME hindcast capture leading IAV modes of precipitation

PC-1 PC-2 PC Spectra: MMEs underestimate QB Peak and total variances PC time series MMEs are highly correlated with CMAP Lead/lag correlation with Nino3.4 SSTA: MME capture ENSO-MNS relation SEOF2 leads ENSO 1 year S-EOF1 concurs with ENSO

MME hindcast VS. ER40 and NCEP-2 (IAV leading modes Taylor diagram: PCC vs. Normalized SD with Reference to CMAP and GPCP. DEMETER and CliPAS MME have higher PCC than ER 40 and NCEP-2 reanalysis, particularly for PCs. But MME PCs carry lower variances Spatial patterns PCs S-EOF1S-EOF2

Power spectrum of Time series of S-EOFs/ Precipitation Fig. 9. The power spectrum density (solid line) and red noise (dashed line) of the first (left-hand panels) and the second (right-hand panels) S-EOF principal component of seasonal precipitation anomaly obtained from GFS AMIP simulation (a, e), CFS 1-month (b, f), 5-month (c, g), and 9-month (d, h) lead time forecast, respectively.

Correlation Skill of EV and PC Predicted by CFS Fig. 10. Pattern correlation of eigenvector and temporal correlation of principle component predicted by CFS with forecast lead time up to 9 month for (a) the 1 st mode of SEOF and (b) the 2 nd mode of SEOF.

How well do the coupled models capture the leading modes of precipitation anomalies? ---The CMAP anomalies show a Quasi-Biennial (QB) and Low-frequency (LF) mode, which accounts for ~ 30% and 13% of the total variance. The QB mode concurs with but LF mode leads ENSO by ~one year. --The MMEs capture both the spatial patterns and temporal evolutions of the two modes with correlation coefficients higher than their counterparts in re-analyses (ER 40 and NCEP2). --However, the MMEs significantly underestimate the biennial tendency of first mode and the temporal variability of the two modes, as well as the “background” noise. These deficiencies limit the MME predictive skill.

Challenges: Monsoon prediction

SCIENTIFIC BASIS for climate prediction Charney and Shukla 1981 “Atmospheric climate variation in the tropics should be more predictable because it was largely due to slowly varying lower boundary (especially SST)”. Tier-2 strategies Palmer (2000) Prediction of seasonal climate fluctuations is essentially an initial value problem. Unlike the weather prediction, predictability arises from a memory of initial conditions in the ocean (as well as land surface). Tier-1 strategies

Two –tier climate prediction system Bengtsson, L., U. Schlese, E. Roeckner, M. Latif, T. Barnett, and N. Graham, 1993: A two- tiered approach to long-range climate forecasting. Science, 261, Graham and Barnett Under the influence of slowly varying boundary forcing, the atmospheric model should be able to capture the predictable portion of tropical rainfall.

Given observed SST forcing Can AGCMs simulate A-AM precipitation anomalies? 11 AGCMs AMIP type 10-member ensemble simulation Observed SST and sea ice as LB forcing 2-year period (9/1996-8/1998) Wang, Kang, Lee 2004 J. Climate

A question of fundamental importance to climate simulation and prediction: Why are all AGCMs, when given the observed lower-boundary forcing, unable to simulate summer monsoon precipitation anomalies? Even with the extremely strong 1997/98 forcing! Bad model climatology? Bad tropical teleconnection? Bad model physics?? Or something else!

Prediction skill for JJA rainfall (2 years) 11-model ensemble mean

A major finding All models tend to yield positive SST-rainfall correlations in the summer monsoon that are at odds with observations. A HYPOTHESIS was raised Treating monsoon as a slave to prescribed SST results in the models’ failure. Or AMIP strategy is inadequate for simulating monsoon rainfall.

How well are the Asian summer monsoon rainfall hindcasted in the Two-Tier MME prediction? NCEP, NSIPP, JMA, KMA, SNU MME: Multi-Model Ensemble AMIP type (Observed SST forcing) ensemble hindcast 6 to 10-member ensemble 21-year period ( )

5-AGCM EM hindcast skill (21Yr) OBS SST-rainfall correlationModel SST-rainfall correlation Two-tier system was unable to predict ASM rainfall. TTS tends to yield positive SST-rainfall correlations in SM region that are at odds with observation (negative). Treating monsoon as a slave to prescribed SST results in the failure. PPC is nearly zero in ASM region Wang et al Two-tier MME hindcast of summer MNS rainfall (Wang et al. 2005)

Area averaged correlation coefficients (skills) El Nino region (10 o S-5 o N, 80 o W-180 o W) WNP (5-30 o N, o E) Asian-Pacific MNS (5-30 o N, o E)

Fig.3 Observed Rainfall-SST correlation ( ) Rainfall Leads SST by 1-month SST leads Rainfall by 1-month Simultatious

Coupled model Experiments IPRC Hybrid Coupled Atmosphere-Ocean Model Atmosphere: ECHAM4 T30 AGCM Ocean:WLF intermediate ocean model (0.5x0.5) (Wang et al. 95, Fu and Wang 2001) 50 year integration Contour indicates 95% significant level

Lagged cross-correlations between SST and precipitation anomalies. Negative correlation with SST lagging the atmosphere suggests that o-to-a feedbacks are weak. SST lags SST leads

Summary SSTA in heavy precipitating summer monsoon region is largely a result of atmospheric forcing. Treating monsoon as a slave to prescribed SST results in the models’ failure. An AGCM, coupled with an ocean model, simulated realistic SST-rainfall relationships; however, the same AGCM fails when forced by the same SSTs that are generated in its coupled run. Neglect of atmospheric feedback makes the forced solution depart from the coupled solution in the presence of initial noises or tiny errors in the lower boundary. Coupled ocean-atmosphere processes are extremely important in the heavy precipitating monsoon convergence zones where the atmospheric feedback to SST is essential.

Issues How predictable is the AAM IAV? What factors limit climate models’ predictability? Other potential sources of predictability? Importance of land processes Can we improve seasonal prediction besides getting better models? Is monsoon ISO a contributor or a trouble maker to seasonal prediction? How to improve the coupled models?

Shinoda et al Land surface process provide memories up to a season?

Koster et al Hot places of land surface feedback

AGCMs simulate climatology poorly over the WNP heat source region ISM (5-30N, E) WNPSM (5-25N, E)

Diurnal cycle biases (Yang and Slingo 2001) UKMO Unified Model Satellite Local time of peak precipitation Satellite shows early evening peak over land, early morning peak over ocean ITCZ. Models show late morning peak over land, midnight peak over ocean.

ISV Variance is too small MJO variance does not come from pronounced spectral peak but from over reddened spectrum: too strong persistence of equatorial precipitation (13/14)

Prediction Needs Design metrics for objective, quantitative assessing predictability and prediction skill. Improve MME prediction system and study combined EM and probabilistic forecast. Improve initialization scheme and initial conditions and representation of slow coupled physics. Develop new strategy and methodology for sub- seasonal monsoon prediction. Better understand physical basis for seasonal prediction and ways to predict uncertainties of the prediction (Palmer 2004).

Thanks Diamond Head