Decadal Prediction and Predictability

Presenter:Edwin K. Schneider

COLA SAC MeetingSeptember 2011

Research Areas

1) Decadal prediction and predictability using the NCEP CFS (version 2) CGCM. Extension of NWP, seasonal, decadal forecasting to decadal time scales.

2) Understanding the mechanisms and predictability of Atlantic decadal variability, currently in a CGCM (CCSM/CESM).

• Long-standing project (Schneider and Kinter 1994; discovered the AO and AAO, but didn’t comment on AO).

3) Issues in climate change, including role of ocean dynamics, aerosol effects, and the expansion of the tropics.

Major Results1) Decadal prediction

• We have completed CMIP5 core decadal forecasts with CFSv2/NEMOVAR, which are now being evaluated.

• We are wrestling with the model bias issue and making some progress.• We are pushing beyond CMIP5, in collaboration with other groups (e.g.

NCEP parallel runs, IRI, GSFC).

2) Mechanisms of decadal variability1) We are developing a better understanding of mechanisms of interannual

to decadal SST variability and the interactions between internal and externally forced variability in “perfect model” simulations.

2) We are settling a controversy: Do AMIP simulation have a fundamental flaw in simulating the SST forced response of the atmosphere? Answer: no.

• Issues in climate change (please see Jian Lu’s poster):• Role of Wind-Evaporation-SST (WES) feedback: Wind stress “overriding”

experiments with CCSM3.• WES feedback is of major importance to explain the response to doubled

CO2 SST and precipitation patterns in the tropics.• WES is important for the Southern Annular mode internal variability.

Research Area 1Decadal Prediction

• Decadal prediction and predictability using the NCEP CFS (version 2) CGCM.

– COLA Team: Cash, DelSole, Huang, Kinter, Klinger, Krishnamurthy, Lu, Marx, Schneider, Stan, Zhu

– Collaborations: NCEP EMC and CPC; IRI; GSFC(Thanks to NCEP for providing the model, data sets and some

technical assistance)

Project Plan

1) Complete and analyze the CMIP5 “core” hindcast/forecast cases as a baseline.

1) Produce additional runs to address problems with the experimental design.

1) Conduct additional experiments to address issues of model bias, improve the hindcasts.

Technical Description• Model

– CFS version 2 provided by NCEP EMC (identical to model used by NCEP for operational S-I prediction and CMIP5)

• Initial data– Atmosphere, land, sea ice: CFSR reanalysis (1980-present)– Ocean: NEMOVAR (ECMWF) interpolated to CFS (1960-present)

• 4-member ensembles– 10 year predictions from Nov. 1960, 1965, 1970, 1975, 1980, 1985, 1990, 1995, 2000,

2005– Extend to 30 years for 1960, 1980, 2005 cases

• Computer resources– NASA Pleiades (Thanks to NAS)

CFS v2

1. An atmosphere of T126L64

2. An interactive ocean with 40 levels in the vertical, to a depth of 4737 m, and horizontal resolution of 0.25 degree at the tropics, tapering to a global resolution of 0.5 degree northwards and southwards of 10N and 10S respectively

3. An interactive 3 layer sea-ice model

4. An interactive land model with 4 soil levels

Results: Biases

• Bias, defined as forecast minus verification as a function of lead time.– Average of all lead times– Evolution high latitudes– Bias development year 1

• Verification data is NCEP/NCAR reanalysis 2m air temperature (NB: too cold in high latitudes vs. CFSRR).

Mean Bias T2mAveraged Years 0-10

Small bias in tropics

Large bias in high

latitudes

Bias Evolution T2m

65°N to 75°N

65°S to 75°S

Sea Ice/AMOC/Salinity Biases: Bohua Huang’s talk

Results: Ensemble Mean Hindcast Anomalies

• NINO3.4 SST index• North Atlantic AMV SST index

• Ensemble mean for each hindcast• Annual cycle removed

NINO3.4 Decadal Hindcasts

NINO3.4 SSTA HindcastsIndividual CasesModel

Obs

NINO 3.4 SSTA Forecast Scores

Rebound of skill in year 2:

Tim DelSole will return to this issue.

Correlation RMS Error

North Atlantic AMV Index(SST Averaged 0N-60N, 300E-360E)

HindcastsObservations

North Atlantic AMV Index(SST Averaged 0N-60N, 300E-360E)

Bias Corrected HindcastsObservations

AMV SSTA Forecast Scores

Correlation RMS Error

Possible Issue

• Initial state choice – CMIP5 protocol: Limited number of initial

years may lead to biased evaluations.– Suggest that a better experimental design

for the same cost might be to sample the initial states more thoroughly with fewer ensemble members.

Summary Decadal Prediction• Multiyear predictability of ENSO has been found.

– Is this real or does it reflect inadequate sampling of initial conditions?

• Have noticed some problems and have had some success in addressing biases.– Inadequate AMOC strength and variability is a

serious model bias.• Decadal predictability in the North Atlantic has

been identified, associated with the long term warming trend.– What is the source of this trend?– What is the source of the skill?

(probably not AMOC)

Research Area 2:Mechanisms of Decadal Variability

• Understanding the mechanisms and predictability of Atlantic decadal variability in a CGCM (CCSM/CESM).

– Research group • COLA/GMU: Schneider, Huang, Chen,

Colfescu• Collaboration: RSMAS – Kirtman group

Motivation

• Diagnose and understand the mechanisms of simulated low frequency Atlantic SST variability.

• In particular, what were the roles of weather noise forcing, coupled feedbacks, and ocean dynamics?

• What are the implications for “decadal” predictability?

Approach

• Perfect model/perfect data framework

• Isolate the weather noise in a CGCM simulation.

• Force an Interactive Ensemble CGCM (IE-CGCM) with the weather noise in controlled experiments.

Data and Models

• NCAR CCSM– CCSM3 (T85 x 1 and T42 x 1)

• 300 year control simulation (T42 x 1)• 20C3M 1980-2000 simulation (T85 x 1)

– CAM3 (AGCM component of CCSM3) + CLM3 (LSM component of CCSM3)

– Interactive ensemble CCSM3– Ensemble mean of 6 copies of CAM3 coupled to

ocean, land, sea ice– T42 x 1

NCAR CCSM3

• Is an appropriate model for this study.

• It produces reasonable simulations of decadal/multidecadal modes of SST variability (i.e., you might guess that this was a model of Earth).– NAO, AMV, TAV

Fig.5 Upper panel: regression of SST anomalies against its normalized AMV index. Lower panel: AMV index. The red curves are the annual mean and the black curves indicate the 11-year running mean with high frequency removed. Data are detrended.

Atlantic Multidecadal Variability“Obs” (130 years) “Ctrl” (300 years)

1880 1900 1920 1940 1960 1980 2000 720 750 780 810 840 870 900 930 960 990

0.5

0

-0.5

0.5

-0.5

0

Weather Noise Forced Interactive Ensemble

AGCM

1AGCM … AGCM

2 N

SST

Response 1 Response 2 Response N…

Ensemble Mean Surface

Fluxes

OGCM

Weather Noise Surface Fluxes

Response to Observed External Forcing CCSM3 and IE-CCSM3 1870-2000 (20C3M)

Colors = ensem

ble mem

bersB

lack= ensemble m

ean

Red

= C

CS

M3

ense

mbl

e m

ean

Blu

e =

CC

SM

3 IE

CCSM3 CMIP4 Global Means

Ts

Ts

Precip

Z 500

1870 1980 1870 1980

.4

-.4

.04

-.04

15

-15

.4

-.4

Experiment Description Purpose1. Control (“Ctrl”)

300-yr CGCM run Control run

2. AGCM ensemble (“signal”)

6 member AGCM ensemble, forced by the same SST from “Ctrl”

To remove the atmospheric noise and get the forced response to the “Ctrl” SST

Weather Noise

Total (from “Ctrl”) “Signal”

Experimental Design

Controversy!Important for Our Method!

• No SST forced AGCM produces a realistic response to observed SST.

• Is this due to fundamental flaw in uncoupled vs. coupled simulations? – This is the result described in publications: Wu and Kirtman

(2004), Wang et al. (2005).• Is this due to model bias?

• Test in a perfect model framework to eliminate model bias (see Hua Chen’s poster-includes comparison with results from above papers).

Answer!Important for Our Method!

• No SST forced AGCM produces a realistic response to observed SST.

• Is this due to fundamental flaw in uncoupled vs. coupled simulations? No

• Is this due to model bias?

Answer!Important for Our Method!

• No SST forced AGCM produces a realistic response to observed SST.

• Is this due to fundamental flaw in uncoupled vs. coupled simulations? No

• Is this due to model bias? Yes: there is no other possibility.

Linear Trends in the Tropics

4. Copsey, D., R. Sutton, and J. R. Knight, 2006: Recent trends in sea level pressure in the Indian Ocean region. GRL, 33.

Trends 1950-1996Copsey et al. CCSM3 20C3M

SST

SLP Obs

CGCM

SLP AGCM

On the Limitations of Prescribing Sea Surface Temperatures in Atmospheric General Circulation Model Experiments

 James W. Hurrell1*, Simon Brown2, Adam Phillips1, and Christophe Cassou1

 

1National Center for Atmospheric Research, Boulder, CO, USA.2United Kingdom Meteorological Office, Bracknell, U.K.*Corresponding author email: jhurrell@ucar.edu Climate Dynamics April 2004 (unpublished, figures lost)

Hurrell et al’s Comparison• Examined ratio of standard deviations CGCM:AGCM.• Their results appear to agree with predictions of

Barsugli and Battisti (1998) model. • Our results appear to agree with Hurrell et al’s., and

additionally provide support to Barsugli and Battisti’s assumptions.

• Our determination of weather noise is in fact based on Hurrell et al’s “limitations.”

Point Correlation of SST IE-CCSM3 with Global Noise Forcing

vs.CCSM3 Control (“Truth”)

Monthly Anomalies Annual Anomalies

=> Most of the SST variability is weather noise forced.

Experiment Description Purpose3. IE_G_all IE-CGCM forced by 3

types of weather noise surface fluxes globally (100 years)

To isolate the role of weather noise in the low frequency SST variability from other mechanisms

Experiment Description Purpose4. IE_ATL_all

Noise surface fluxes restricted to the whole Atlantic only; zero elsewhere (100 years)

To isolate the role of local weather noise from other mechanisms

IE ExperimentsGlobal weather noise

Weather noise in Whole Atlantic only

…

Summary Decadal Prediction

• Our decadal prediction project is well under way, using CFS_v2, and is generating questions:– ENSO skill rebound?– Sampling of initial states?– Biases North Atlantic and polar regions?– Source of skill in hindcasting AMV?

Summary - Decadal Mechanisms

• Contrary to some findings in the literature, we find (in CCSM3) that AMIP simulations are able to capture the SST-forced response of the atmosphere.

• Weather noise and external forcing are the major sources of SST variability in CCSM3.– Weather noise is the cause of variability about the

ensemble mean for the CCSM3 20C3M ensemble.– Using our method, we will arrive at a reasonably

complete understanding of the mechanisms of decadal variability in CGCMs.