On representing model uncertainty in climate predictions T.N.Palmer ECMWF with thanks to Francisco Doblas-Reyes, Thomas Jung and Antje Weisheimer, ECMWF
Jan 15, 2016
On representing model uncertainty in climate
predictions
T.N.Palmer
ECMWF
with thanks to Francisco Doblas-Reyes, Thomas Jung and Antje Weisheimer,
ECMWF
Model uncertainty
Model uncertainty
Scenario uncertainty
Scenario uncertainty
Initial uncertainty
Hawkins and Sutton, 2009
Standard Numerical Ansatz for Climate Model
Deterministic local bulk-formula parametrisation
Increasing scale
;nP X
Eg momentum“transport” by:
•Turbulent eddies in boundary layer
•Orographic gravity wave drag.
•Convective clouds
1X 2X 3X nX... ...
2. pt
u u g uEg
Atmosphere Atmosphere Atmosphere Atmosphere Atmosphere Atmosphere
Land surfaceLand surfaceLand surfaceLand surfaceLand surface
Ocean & sea-ice Ocean & sea-ice Ocean & sea-ice Ocean & sea-ice
Sulphateaerosol
Sulphateaerosol
Sulphateaerosol
Non-sulphateaerosol
Non-sulphateaerosol
Carbon cycle Carbon cycle
Atmosphericchemistry
Ocean & sea-icemodel
Sulphurcycle model
Non-sulphateaerosols
Carboncycle model
Land carboncycle model
Ocean carboncycle model
Atmosphericchemistry
Atmosphericchemistry
Off-linemodeldevelopment
Strengthening coloursdenote improvementsin models
1975 1985 1992 1997
The
Met
.Offi
ce H
adle
y C
entr
e
Towards Comprehensive Earth System Models
1970 2000
Atmosphere Atmosphere Atmosphere Atmosphere Atmosphere Atmosphere
Land surfaceLand surfaceLand surfaceLand surfaceLand surface
Ocean & sea-ice Ocean & sea-ice Ocean & sea-ice Ocean & sea-ice
Sulphateaerosol
Sulphateaerosol
Sulphateaerosol
Non-sulphateaerosol
Non-sulphateaerosol
Carbon cycle Carbon cycle
Atmosphericchemistry
Ocean & sea-icemodel
Sulphurcycle model
Non-sulphateaerosols
Carboncycle model
Land carboncycle model
Ocean carboncycle model
Atmosphericchemistry
Atmosphericchemistry
Off-linemodeldevelopment
Strengthening coloursdenote improvementsin models
1975 1985 1992 1997
The
Met
.Offi
ce H
adle
y C
entr
e
1970 2000
Uncertainty
A Missing Box
How can uncertainty be represented in ESMs?
• Multi-model ensembles
• Perturbed parameters
• Stochastic parametrisation
Seasonal multi-model ensemble
Seasonal Reforecasts (months 2-4) of El Niño with a comprehensive coupled model
observations
predictions
Multi-model seasonal reforecasts of El Niño
precipitation in DJFstart dates: Nov hindcast period: 1991-2005
lower tercile
Amazon Central America Northern Europe
Multi-model Seasonal Forecast Reliability
Failure of multi-model ensemlble
Slide 13
Surface Pressure
Potential Vorticity on 315K
Blocking Anticyclone
As recognised in AR4, the current
generation of climate models has difficulty simulating a number of internal
modes of climate variability such as
the persistent blocking
anticyclone.
Blocking Index. DJFM 1960-2003
ERA-40
T159
T1259
T1259 run on NSF Cray XT4 “Athena” (two months of dedicated usage) Similar results found by M.Matsueda MRI Japan
For all their pragmatic value, multi-model ensembles are ad hoc “ensembles of opportunity”.
Component models have common shortcomings, eg due to
limited resolution.
How can uncertainty be represented in ESMs?
• Multi-model ensembles
• Perturbed parameters
• Stochastic parametrisation
Deterministic local bulk-formula parametrisation
Increasing scale
;nP X
1X 2X 3X nX... ...
2. pt
u u g u
Vary α
Perturbed Parameters
How can uncertainty be represented in ESMs?
• Multi-model ensembles
• Perturbed parameters
• Stochastic parametrisation
A stochastic-dynamic paradigm for the Earth-System model
Computationally-cheap nonlinear stochastic-dynamic models, providing specific possible realisations of sub-grid motions rather than sub-grid bulk effects
Coupled over a range of scales
Increasing scale
ECMWF Tech Memo 598
SAC 2009
Spectral Stochastic Backscatter Scheme
• Origins: Leith (1990), Mason and Thomson (1992)
• Shutts, G.J. (2005). A kinetic energy backscatter algorithm for use in ensemble prediction systems. Q.J.R.Meteorol.Soc. 131, 3079
• Berner, J. et al (2009). A spectral stochastic kinetic energy backscatter scheme and its impact on flow-dependent predictability in the ECMWF ensemble prediction system. J. Atmos.Sci., 66, 603-626.
SAC2009
Slide 21
Backscatter Algorithm
Streamfunction forcing Pattern using spectral AR(1) processes as SPPT
Dtot is a smoothed total dissipation rate, normalized here by
Btot and bR is the backscatter ratio
Realisations of stochastic pattern generator
In ENSEMBLES we have tested the relative ability of these different representations of uncertainty:
Multi-model ensemblesPerturbed parametersStochastic physics
to make skilful probabilistic seasonal climate predictions.
“Giorgi” Regions
dry wet dry wetAustralia
1 2 2 3MM best
Amazon Basin3 0 1 3
PP bestSouthern South America
1 1 1 1SP best
Central America2 3 3 2
Western North America3 3 3 3
Central North America1 1 3 3
Eastern North America1 1 2 3
Alaska3 1 2 3
Greenland1 3 2 3
Mediterranean3 3 3 3
Northern Europe2 2 3 3
Western Africa3 1 2 3
Eastern Africa3 3 2 2
Southern Africa3 3 3 2
Sahel1 3 2 1
South East Asia1 1 1 0
East Asia3 3 3 3
South Asia3 3 3 3
Central Asia2 3 2 2
Tibet1 1 1 2
North Asia3 2 1 1
precipitationJJA DJF
1991-2005lead times: 2-4 monthsDry=lower tercile
Wet=upper tercile Which is best?Brier Skill Score
cold warm cold warmAustralia
3 3 3 1MM best
Amazon Basin3 1 1 1
PP bestSouthern South America
1 1 3 2SP best
Central America3 1 3 1
Western North America1 1 3 1
Central North America1 1 2 2
Eastern North America3 3 2 3
Alaska3 3 2 2
Greenland1 1 2 2
Mediterranean3 2 1 3
Northern Europe2 2 3 3
Western Africa1 1 3 3
Eastern Africa1 1 2 3
Southern Africa1 2 1 1
Sahel1 2 1 3
South East Asia1 1 2 1
East Asia3 2 1 2
South Asia3 1 3 3
Central Asia1 2 3 2
Tibet3 1 3 3
North Asia1 2 1 2
temperatureJJA DJF
1991-2005lead times: 2-4 months
Brier Skill Score
Cold=lower tercile
Warm=upper tercile
precipitation over Northern Europe land (north of 48ºN) in DJFstart dates: Nov 1st. hindcast period: 1991-2005
lower tercile
stochastic physics #7
BSS(∞)=0.087BSS(∞)=-0.018
perturbed physicsmulti-model
BSS(∞)=-0.031
Multi-model Seasonal Forecast Reliability
Conclusions
• Stochastic parametrisation and perturbed parameter methodologies are competitive with the traditional multi-model approach to representing model uncertainty
• Stochastic parametrisation “wins” overall for atmospheric variables, but needs to be extended to the ocean and the land surface.
• The ECMWF THOR integrations will be started next year using the latest stochastic parametrisation schemes.