Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Comparison between ensembles generated with atmospheric andwith oceanic perturbations, using the MPI-ESM model
C. Marini 1. A. Köhl 1. D. Stammer 1
Vanya Romanova 2. Jürgen Kröger3
1Institut für Meereskunde - Universität Hamburg
2Universität Bonn
3Max Plank Institut für Meteorologie
International workshop on seasonal to decadal prediction, May 2013
DecadalClimate Prediction
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Ensemble generation for decadal predictions?
Our GOAL: representation of errors on OCEANIC initial conditions
WHY?Memory of the ocean = key hypothesis for predictability at decadal time scalesLarge errors on the oceanic state, especially at depth
? Zanna et al 2011, 2012 ?
Deep density perturbations in the Atlantic⇓
Faster (7.5 vs 18.5 yr) and larger AMOC ampli�cations than with upperoceanic perturbations
⇓Overestimation of oceanic predictability in decadal predictions exp where only
the atmospheric state is perturbed
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Ensemble generation for decadal predictions?
Our GOAL: representation of errors on OCEANIC initial conditions
WHY?Memory of the ocean = key hypothesis for predictability at decadal time scalesLarge errors on the oceanic state, especially at depth
? Zanna et al 2011, 2012 ?
Deep density perturbations in the Atlantic⇓
Faster (7.5 vs 18.5 yr) and larger AMOC ampli�cations than with upperoceanic perturbations
⇓Overestimation of oceanic predictability in decadal predictions exp where only
the atmospheric state is perturbed
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
A need to create more relevant ensembles for decadal predictions?
For now, ensemble generation : rather pragmatic
Perturbing the atmospheric horizontal di�usion coe�cient(e.g. Matei et al Sci 2012)
Lagged initialization : shifting by a few days→ the atmospheric initial conditions→ the oceanic initial conditions(e.g Smith et al Sci 2007, Msadek et al GRL 2010, Müller et al GRL 2012)
Use of ensemble of ocean reanalyses/assimilation runs- obtained by using di�erent initial conditial conditions of the historical run- obtained by perturbing the wind stress (+ SST perturbations at the hindcast's start) inWeisheimer et al GRL 2009, von Oldenborgh CD 2012
Get inspired from weather forecast and seasonal predictions :
Singular vectors (Hawkins and Sutton JCLI 2009)
Anomaly Transform
Breeding vectors
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
A need to create more relevant ensembles for decadal predictions?
Comparison between :
Perturbing the atmospheric horizontal di�usion coe�cient(e.g. Matei et al Sci 2012)
Lagged initialization : shifting by a few days→ the atmospheric initial conditions
→ the oceanic initial conditions(e.g Smith et al Sci 2007, Msadek et al GRL 2010, Müller et al GRL 2012)
Use of ensemble of ocean reanalyses/assimilation runs- obtained by using di�erent initial conditial conditions of the historical run- obtained by perturbing the wind stress (+ SST perturbations at the hindcast's start) inWeisheimer et al GRL 2009, von Oldenborgh CD 2012
Get inspired from weather forecast and seasonal predictions :
Singular vectors ... A cheap version of oceanic singular vectors ...
Anomaly Transform
Breeding vectors
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
An attempt to generate climatic relevant perturbations
Goal : Finding perturbations
• growing the most rapidly with a time scale relevant for decadal predictions
• scaled to represent the uncertainties contained in the initial conditions
→ Singular vectors (SVs) of the tangent propagator of the dynamical systemrepresenting the evolution of the ocean(Molteni 1996 for the atmosphere, Palmer and Zanna submitted for a review about SVs)
TOO EXPENSIVE : requires the tangent propagator and its adjoint of the oceanic model ...
→ Simpli�cation of the evolution of the oceanic state based onLinear Inverse Modeling (LIM)
dx
dt= Bx+ F
- x : 3d EOFs of T and Sfrom the 156-yr long historical run of the MPI-ESM model68.4% of the total variance
- B : Deterministic matrix, F : White noise
? Tzipermann and Zanna JCLI 2008 : 3d EOFs of T and S annual anomalies in the Atlantic
? Hawkins and Sutton JCLI 2009 : idem but only from the upper 2km of the Atlantic
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
An attempt to generate climatic relevant perturbations
Goal : Finding perturbations
• growing the most rapidly with a time scale relevant for decadal predictions
• scaled to represent the uncertainties contained in the initial conditions
→ Singular vectors (SVs) of the tangent propagator of the dynamical systemrepresenting the evolution of the ocean(Molteni 1996 for the atmosphere, Palmer and Zanna submitted for a review about SVs)
TOO EXPENSIVE : requires the tangent propagator and its adjoint of the oceanic model ...
→ Simpli�cation of the evolution of the oceanic state based onLinear Inverse Modeling (LIM)
dx
dt= Bx+ F
- x : 3d EOFs of T and Sfrom the 156-yr long historical run of the MPI-ESM model68.4% of the total variance
- B : Deterministic matrix, F : White noise
? Tzipermann and Zanna JCLI 2008 : 3d EOFs of T and S annual anomalies in the Atlantic
? Hawkins and Sutton JCLI 2009 : idem but only from the upper 2km of the Atlantic
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
An attempt to generate climatic relevant perturbations
Goal : Finding perturbations
• growing the most rapidly with a time scale relevant for decadal predictions
• scaled to represent the uncertainties contained in the initial conditions
→ Singular vectors (SVs) of the tangent propagator of the dynamical systemrepresenting the evolution of the ocean(Molteni 1996 for the atmosphere, Palmer and Zanna submitted for a review about SVs)
TOO EXPENSIVE : requires the tangent propagator and its adjoint of the oceanic model ...
→ Simpli�cation of the evolution of the oceanic state based onLinear Inverse Modeling (LIM)
dx
dt= Bx+ F
- x : 3d EOFs of T and Sfrom the 156-yr long historical run of the MPI-ESM model68.4% of the total variance
- B : Deterministic matrix, F : White noise
? Tzipermann and Zanna JCLI 2008 : 3d EOFs of T and S annual anomalies in the Atlantic
? Hawkins and Sutton JCLI 2009 : idem but only from the upper 2km of the Atlantic
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Linear Inverse Modeling (LIM)
Evolution of a perturbation of the system :
x(t + τ) = eτBx(t) = G(τ)x(t)
Ampli�cation over a time τ under the L2 norm :
µ(τ) =‖x(τ)‖2
‖x(0)‖2=
x(τ)T x(τ)
x(0)T x(0)=
x(0)TG(τ)TG(τ)x(0)
x(0)T x(0)
Singular vectors = eigenvectors of G(τ)TG(τ)
- τ = 5 yr- Selection of the �rst 4 SVs
+ Rotation and Scaling of the SVs similar to Molteni et al 1996
→ More uniform spatial distribution→ Match �on average� the estimates of the rms error1 of the oceanic reanalysisused to create the oceanic initial conditions (GECCO2)
1a posteriori errors
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Linear Inverse Modeling (LIM)
Evolution of a perturbation of the system :
x(t + τ) = eτBx(t) = G(τ)x(t)
Ampli�cation over a time τ under the L2 norm :
µ(τ) =‖x(τ)‖2
‖x(0)‖2=
x(τ)T x(τ)
x(0)T x(0)=
x(0)TG(τ)TG(τ)x(0)
x(0)T x(0)
Singular vectors = eigenvectors of G(τ)TG(τ)
- τ = 5 yr- Selection of the �rst 4 SVs
+ Rotation and Scaling of the SVs similar to Molteni et al 1996
→ More uniform spatial distribution→ Match �on average� the estimates of the rms error1 of the oceanic reanalysisused to create the oceanic initial conditions (GECCO2)
1a posteriori errors
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Pattern of the rotated and scaled SVs. Temperature at 150m.
SV1 T at 150m SV2 T at 150m
SV3 T at 150m SV4 T at 150m
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Pattern of the rotated and scaled SVs. Salinity at 150m.
SV1 S at 150m SV2 S at 150m
SV3 S at 150m SV4 S at 150m
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Rms of the rotated and scaled SVs and rms error of GECCO2 at 150m.
rms of the SVs rms errors
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Rms of the rotated and scaled SVs and rms error of GECCO2 at 1220m.
rms of the SVs rms errors
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Hindcasts with the MPI-ESM model
Initial condition from an "assimilation run" :Nudging of T and S towards T and S anomalies from GECCO2 put onthe MPI-ESM climatology
10-yr long hindcasts starting every 1st January from 1991 to 2006
Two types of ensembles :
1 ensemble =
1 unperturbed member +
8 perturbed members
AtmosphericPerturbations
=Atm ICs shifted by 1 to 8 days
OceanicPerturbations
=Oce ICs
+SV1, +SV2, +SV3, +SV4,-SV1, -SV2, -SV3, -SV4
⇒ 16× 2 ensembles=
16× (2× 8+ 1)× 10yr = 2720 yr
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
A validation of the ensemble spread : Talagrand diagram
9 ensemble members sorted in ascending order⇓
10 bins for each time step : •| • | • | • | • | • | • | • | • |•At each time step the observation is located in one bin
⇓
Histogram :
Flat : relevant spreadL shape : bias
U shape : lack of spread
n shape : overestimated spread
Associated scores (Keller and Hense 2011) :• beta score (bS)bS=0 �at histogrambS>0 too large spreadbS<0 too narrow spread
• beta bias (bB)bB>0 bias towards higher values
bB<0 bias towards lower values
Mean bias removedfollowing the recommendation of the CMIP-WGCM-WGSIP Decadal Climate Prediction Panel
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Atlantic meridional overturning circulation around 26.5N
Monthly values, Veri�cation values : assimilation run
Total Overturning
1 2 3 4 5 6 7 8 9 100
100
200
300
400
500
1 2 3 4 5 6 7 8 9 100
100
200
300
400
500
1 2 3 4 5 6 7 8 9 100
100
200
300
400
500
1 2 3 4 5 6 7 8 9 100
100
200
300
400
500
1 2 3 4 5 6 7 8 9 100
100
200
300
400
1 2 3 4 5 6 7 8 9 100
100
200
300
400
1 2 3 4 5 6 7 8 9 100
100
200
300
400
1 2 3 4 5 6 7 8 9 100
100
200
300
400
Lag1
Lag2
Lag3
Lag4
Oceanic perturbations Atmospheric perturbationsbS=-0.16, bB=0.11 bS=-0.45, bB=0.09
bS=-0.23, bB=0.11 bS=-0.35, bB=0.08
bS=-0.07, bB=0.11 bS=-0.28, bB=0.10
bS=-0.22, bB=0.10 bS=-0.38, bB=0.08
The Ekman component has been removed
1 2 3 4 5 6 7 8 9 100
100
200
300
400
500
600
1 2 3 4 5 6 7 8 9 100
100
200
300
400
500
600
1 2 3 4 5 6 7 8 9 100
200
400
600
1 2 3 4 5 6 7 8 9 100
200
400
600
1 2 3 4 5 6 7 8 9 100
100
200
300
400
500
1 2 3 4 5 6 7 8 9 100
100
200
300
400
500
600
1 2 3 4 5 6 7 8 9 100
100
200
300
400
500
1 2 3 4 5 6 7 8 9 100
100
200
300
400
500
600
Oceanic perturbations Atmospheric perturbationsLag1 bS=-0.35, bB=0.15 bS=-0.95, bB=0.09
bS=-0.43, bB=0.13 bS=-0.98, bB=0.07
bS=-0.19, bB=0.11 bS=-0.63, bB=0.07
bS=-0.27, bB=0.11 bS=-0.81, bB=0.06
Lag2
Lag3
Lag4
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Atlantic meridional overturning circulation around 47.5N
Monthly values, Veri�cation values : assimilation run
1 2 3 4 5 6 7 8 9 100
100
200
300
1 2 3 4 5 6 7 8 9 100
100
200
300
1 2 3 4 5 6 7 8 9 100
100
200
300
1 2 3 4 5 6 7 8 9 100
100
200
300
1 2 3 4 5 6 7 8 9 100
50
100
150
200
250
1 2 3 4 5 6 7 8 9 100
50
100
150
200
250
300
1 2 3 4 5 6 7 8 9 100
50
100
150
200
250
1 2 3 4 5 6 7 8 9 100
50
100
150
200
250
300
Oceanic perturbationsAtmospheric perturbationsLag1 bS=0.17, bB=0.17 bS=-0.59, bB=0.07
bS=-0.15, bB=0.10 bS=-0.60, bB=0.06
bS=-0.17, bB=0.10 bS=-0.51, bB=0.06
bS=-0.29, bB=0.12 bS=-0.35, bB=0.08
Lag2
Lag3
Lag4
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Oceanic heat content upper 700m and SST in the Atlantic (yearly values)
Atlantic basin
Veri�cation values :NODC dataset (Levitus et al 2009)
1 2 3 4 5 6 7 8 9 100
0.5
1
1.5
2
2.5
1 2 3 4 5 6 7 8 9 100
0.5
1
1.5
2
2.5
1 2 3 4 5 6 7 8 9 100
0.5
1
1.5
2
1 2 3 4 5 6 7 8 9 100
0.5
1
1.5
2
1 2 3 4 5 6 7 8 9 100
5000
10000
15000
1 2 3 4 5 6 7 8 9 100
5000
10000
15000
1 2 3 4 5 6 7 8 9 100
5000
10000
15000
1 2 3 4 5 6 7 8 9 100
5000
10000
15000
Oceanic perturbations Atmospheric perturbations
Lag1
bS=0.14, bB=0.21 bS=-1.76, bB=0.05
bS=0.15, bB=0.21 bS=-0.91, bB=0.07
Lag2
Lag4
Lag3
bS=-0.57, bB=0.09bS=0.22, bB=0.22
bS=0.25, bB=0.23 bS=-0.43, bB=0.10
x 10⁴
x 10⁴
x 10⁴
x 10⁴
North Atlantic (80W-0, 0-60N)
Veri�cation values :HadISST dataset (Rayner et al 2003)
1 2 3 4 5 6 7 8 9 100
2000
4000
6000
8000
1 2 3 4 5 6 7 8 9 100
2000
4000
6000
8000
1 2 3 4 5 6 7 8 9 100
2000
4000
6000
8000
1 2 3 4 5 6 7 8 9 100
2000
4000
6000
8000
1 2 3 4 5 6 7 8 9 100
2000
4000
6000
8000
1 2 3 4 5 6 7 8 9 100
2000
4000
6000
8000
1 2 3 4 5 6 7 8 9 100
2000
4000
6000
8000
1 2 3 4 5 6 7 8 9 100
2000
4000
6000
8000
bS=0.39, bS=0.34 bS=0.23, bS=0.15
bS=0.43, bS=0.35 bS=0.02, bS=0.18
bS=0.38, bS=0.31 bS=0.09, bS=0.21
bS=0.38, bS=0.29 bS=0.17, bS=0.23
Oceanic perturbations Atmospheric perturbations
Lag1
Lag2
Lag3
Lag4
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Ensemble spread of SST averaged in the North Atlantic
1 2 3 4 5 6 7 8 9 100
0.05
0.1
0.15
0.2
0.25
1 2 3 4 5 6 7 8 9 100
0.05
0.1
0.15
0.2
0.25
years
years
Ensemble spread from every starting date
oceanic perturbations
atmospheric perturbations
→ Ensemble spread of SST averaged in theNorth Atlantic decreases with time foroceanic perturbations
Idem for temperature averaged in the upper200m of the Atlantic basin.
Why?Perturbations averaged over these areas are larger than the std of the signal→ The model gets back to its climatology
Trade-o� between :? strong perturbations that actually represent the errors on initial conditions? when too strong perturbations, model gets back to its climatology
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Does the ampli�cation predicted by the LIM occur in the hindcasts withoceanic perturbation?
Ampli�cation over a time τ = 5 yr under the L2 norm :
µ(τ) =‖x(τ)‖2
‖x(0)‖2 =x(τ)Tx(τ)
x(0)Tx(0)=
x(0)TG(τ)TG(τ)x(0)
x(0)Tx(0)with x =
(Tunpert − Tpert
Sunpert − Spert
)
0 1 2 3 4 51
1.5
2
2.5
0 1 2 3 4 50.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
0 1 2 3 4 50.8
1
1.2
1.4
1.6
1.8
2
2.2
0 1 2 3 4 50.8
1
1.2
1.4
1.6
1.8
2
2.2
SV3 SV4
SV1 SV2
year year
Ampli
Ampli
- Red : Ampli�cation predicted by the
LIM
- Mean of ampli�cation over the
hindcasts :
Black : +SVBlue : -SVSolid thin line : contribution of temperature
Dashed thin line : contribution of salinity
Ampli�cations of the perturbationsdo occur under the L2 norm
BUT the average over a given areaof the perturbations decreases
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Comparison with other methods
Comparison between ensemble generated with:
Cheap Oceanic Singular vectors
Atmospheric laggged initialization
Atmospheric and Oceanic laggged initialization Jürgen Kröger
Ensemble of assimilation runsobtained by using di�erent initial conditions of the historical run Jürgen Kröger
Anomaly Transform Poster 31 of Vanya Romanova
Set up :
Initial condition from the "assimilation run" towards GECCO2 anomalies
10-yr long hindcasts starting the 1st January 1996
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Atlantic Meridional Overturning Circulation at 26.5N
1996 1997 1998 1999 2000 2001 2002 2003 2004 200515
16
17
18
19
20
21
22
23
24
25
1996 1997 1998 1999 2000 2001 2002 2003 2004 200515
16
17
18
19
20
21
22
23
24
25
1996 1997 1998 1999 2000 2001 2002 2003 2004 200515
16
17
18
19
20
21
22
23
24
25
1996 1997 1998 1999 2000 2001 2002 2003 2004 200515
16
17
18
19
20
21
22
23
24
25
Atmospheric lagged initialization
Oceanic lagged initialization
Cheap oceanic Singular Vectors
Anomaly transform
1996 1997 1998 1999 2000 2001 2002 2003 2004 200515
16
17
18
19
20
21
22
23
24
25
From ensemble of assimilation run 1996 1997 1998 1999 2000 2001 2002 2003 2004 20051516171819202122232425
1996 1997 1998 1999 2000 2001 2002 2003 2004 20050
0.5
1
1.5
2
assimilation runatmospheric lagged initializationanomaly transformcheap oceanic singular vectorsfrom ensemble of assimilation run
Ensemble Mean (dashed), +/-ensemble std (solid), and reference run (red)
Ensemble std and std of the reference run
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Temperature averaged in the upper 200m
North Atlantic basin
1996 1997 1998 1999 2000 2001 2002 2003 2004 2005268.7
268.8
268.9
269
269.1
269.2
269.3
269.4
1996 1997 1998 1999 2000 2001 2002 2003 2004 20050
0.05
0.1
0.15
0.2
0.25
Ensemble Mean (dashed), +/-ensemble std (solid),
and reference run (red)
Ensemble std
assimilation runatmospheric lagged initializationanomaly transformcheap oceanic singular vectorsfrom ensemble of assimilation run
Paci�c basin
1996 1997 1998 1999 2000 2001 2002 2003 2004 2005284
284.1
284.2
284.3
284.4
284.5
284.6
284.7
1996 1997 1998 1999 2000 2001 2002 2003 2004 20050
0.05
0.1
0.15
0.2
assimilation runatmospheric lagged initializationanomaly transformcheap oceanic singular vectorsfrom ensemble of assimilation run
Ensemble Mean (dashed), +/-ensemble std (solid),
and reference run (red)
Ensemble std
WORK IN PROGRESS part of Miklip...Hard to evaluate the bene�ts of one method with only one starting date...
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions
Introduction Poor man's oceanic SVs Oceanic vs Atmospheric perturb Comparison with other methods Conclusion
Conclusion
F Comparison between a cheap version of oceanic singular vectors andatmospheric lagged initialisation
Oceanic perturbationsglobally better spread, at least not worse
Too large perturbations ⇒ model gets back to its climatology?
F How to improve the cheap oceanic singular vectors?
Get rid of the linear approximation
Choice of a better norm to compute the singular vectors
C. Marini. A. Köhl. D. Stammer Oceanic perturbations - decadal predictions