Post on 14-Mar-2020
Analysis of model drift in a climate forecast system used
for decadal predictions
Emilia Sanchez-Gomez, Christophe Cassou, Elodie Fernandez, Laurent Terray
CERFACS – SUC 1875 (Toulouse) France
Special thanks to the CNRM colleagues
1. Initialisation methodology to minimise the model drift
1. Analysis of the drift dynamics in decadal hindcasts
3. Conclusions
Outline
Initialisation methodology
Aims:
1. Minimising the initial shock and the model drift (initial conditions close
to model mean state)
2. Obtaining initial conditions for sea ice and atmosphere components,
compatible with initialisation date.
Solution :
Nudging of the ocean component in the coupled model towards a ocean
reanalysis (NEMOVAR).
ARPEGE-Climat T127 (1.4°), 31 levels
SURFEX Interface
GELATO v5
NEMO 1° 42 levels
OASIS River Routing
TRIP
The NEMOVAR-COMBINE
ECMWF Ocean Reanalysis
Balsameda et al. 2010
CNRM-CM5 coupled model
Nudging only of the ocean component =
3D Newtonian damping +
Sea surface restoring
Aims:
1. Minimising the initial shock and the model drift (initial conditions close
to model mean state)
2. Obtaining initial conditions for sea ice and atmosphere components,
compatible with initialisation date.
Initialisation methodology
NO-EQ IC No 3D nudging within the 1°S–1°N band
• Several tests have been performed on the nudging parameters to avoid
spurious effects on the ocean and to control the nudging strength (spatial
domain, sea surface restoring).
Initial conditions issued from nudging experiments
3D Newtonian damping + sea surface restoring
NO-TROP IC No 3D nudging Within the 15°S–15°N band
Initial conditions issued from nudging experiments
NO-EQ IC No 3D nudging within the 1°S–1°N band
3D Newtonian damping + sea surface restoring
• Several tests have been performed on the nudging parameters to avoid
spurious effects on the ocean and to control the nudging strength (spatial
domain, sea surface restoring).
Mean 1960-2005
1. Initial conditions to minimise model drift Initial conditions issued from nudging experiments
Ocean Heat Content (300m) NOEQ - NOTROP
Uinitialized NEMOVAR NOEQ NOTROP
ΔH
Indian Ocean Pacific Ocean Atlantic Ocean
SST averaged over 60°N - 35°S NOEQ vs NOTROP
SST observations
Forecast annual means for leadtime 1- 5 (years)
- Only 9 initial dates
- Only atmosphere is perturbed to obtain the members (10 members)
1. Initial conditions to minimise model drift Decadal experiments using different ICs
SST averaged over 60°N - 35°S NOEQ vs NOTROP
SST observations
Forecast annual means for leadtime 1- 5 (years)
- Strong warming at forecast range 1
- Progressive cooling towards the equilibrium state
1. Initial conditions to minimise model drift Decadal experiments using different ICs
SST observations 0.08 °C
0.01 °C
NOEQ decadal forecasts exhibits a stronger warming at the first lead
years NOTROP decadal forecasts completed for CMIP5.
SST averaged over 60°N - 35°S
1. Initial conditions to minimise model drift Decadal experiments using different ICs
NOEQ vs NOTROP
Focus on internal variability modes of the CNRM-CM5 model
2. Understanding the model drift Analysis of the model drift : Methodology
1. Natural modes of variability determined by EOF analysis on PiCTL
(PreIndistrial – Control) experiment :
- ENSO
- Tropical Atlantic
- AMO, PDO
2. Decadal forecasts EOF projection PDF analysis
Seasonal means considered: JFM , AMJ , JAS, OND
(only JFM is shown here)
Leadtim
e (
years
) 2. Understanding the model drift Analysis of the model drift : Equatorial Pacific (ENSO)
Uninitialized
Leadtim
e (
years
) 2. Understanding the model drift Analysis of the model drift : Equatorial Pacific (ENSO)
NOEQ
Uninitialized
Leadtim
e (
years
) 2. Understanding the model drift Analysis of the model drift : Equatorial Pacific (ENSO)
Y1: no drift
Y2: An El Niño event is excited for most of members
Y3: An La Niña event occurs for most of members
- Fast drift (equilibrium at 4 years lead time)
- Oscillatory behaviour
NOEQ
Uninitialized
Leadtim
e (
years
) 2. Understanding the model drift Analysis of the model drift : Equatorial Pacific (ENSO)
Y1: no drift
Y2: An El Niño event is excited for most of members
Y3: An La Niña event occurs for most of members
- Fast drift (equilibrium at 4 years lead time)
- Oscillatory behaviour
- More moderate initial shock for NOTROP initialized
forecasts
NOTROP
NOEQ
Uninitialized
Leadtim
e (
years
) 2. Understanding the model drift Analysis of the model drift : Equatorial Pacific (ENSO)
Mean bias CNRM-CM5 Non Initialized – NEMOVAR
Ocean heat content [0-300m]
The strong heat release can be explained by model
biases
NOTROP
NOEQ
Uninitialized
2. Understanding the model drift Analysis of the model drift : Equatorial Pacific (ENSO)
Color=precip / contour=Z500
(significance hatching 95%)
Atmospheric response :
NOEQ – NOTROP differences
2. Understanding the model drift Analysis of the model drift : Equatorial Pacific (ENSO)
Color=precip / contour=Z500
(significance hatching, 95%)
Atmospheric response :
NOEQ – NOTROP differences
NAO WESPAC
NOTROP
NOEQ
Uninitialized
2. Understanding the model drift Analysis of the model drift : Equatorial Pacific (ENSO)
Color=precip / contour=Z500
(significance hatching 95%)
Atmospheric response :
NOEQ – NOTROP differences
- Y1: Initial shock at the first
winter (positive NAO)
- Y2: negative NAO, positive
WESPAC.
NAO WESPAC
NOTROP
NOEQ
Uninitialized
2. Understanding the model drift Analysis of the model drift : Tropical Atlantic
North Tropical Atlantic Mode (JFM)
Power sprectrum
95%
5%
2.3yr 8-16yr
2. Understanding the model drift Analysis of the model drift : Tropical Atlantic
North Tropical Atlantic Mode (JFM)
Power sprectrum
95%
5%
2.3yr
Leadtim
e (
years
)
- Y1: the drift starts
- Linear and slow drift (cooling)
- Smaller differences between NOTROP and NOEQ
forecasts
8-16yr
NOTROP
NOEQ
Uninitialized
2. Understanding the model drift Analysis of the model drift : Tropical Atlantic
North Tropical Atlantic Mode (JFM)
Power sprectrum
95%
5%
2.3yr
Leadtim
e (
years
)
- Y1: the drift starts
- Linear and slow drift (cooling)
- Smaller differences between NOTROP and NOEQ
forecasts
8-16yr
NOTROP
NOEQ
Uninitialized
2. Understanding the model drift Analysis of the model drift : AMO and PDO
AMO PDO
NOTROP
Uninitialized
- Y1: Positives AMO and PDO
- Linear and slow drift (slow cooling)
- Likely connected to ocean gyre adjustement.
- AMO: equilibrium is not reached at leadyear 10
- PDO: equilibrium is reached at leadyear 5
2. Understanding the model drift Analysis of the model drift : AMO and PDO
AMO PDO
NOTROP
Uninitialized
- Y1: Positives AMO and PDO
- Linear and slow drift (slow cooling)
- Likely connected to ocean gyre adjustement.
- AMO: equilibrium is not reached at leadyear 10
- PDO: equilibrium is reached at leadyear 5
2. Understanding the model drift Analysis of the model drift : AMO and PDO
AMO PDO
NOTROP
Uninitialized
- Y1: Positives AMO and PDO
- Linear and slow drift (slow cooling)
- Likely connected to ocean gyre adjustement.
- AMO: equilibrium is not reached at leadyear 10
- PDO: equilibrium is reached at leadyear 5
BIAS COR.
Generating initial conditions controlling nudging strength reduces the
model initial shock, nevertheless the model drift is still present.
Strong warming at forecast range 1 + progressive cooling
The initial shock reduction seems to have no impact on the forecast
skill on SSTs (maybe other variables?).
To derive to its own climate, the model precisely uses the internal
modes of variability that we seek to predict, putting some shade and
uncertainties on traditional skill score.
According to this, the model drift can behaves differently depending
on the internal variability mode considered :
- ENSO: fast drift and oscillatory behaviour.
- North Tropical Atlantic: Slow and linear drift (slow cooling)
- AMO and PDO : slow drift and linear behaviour (for PDO
stabilisation at Y5, whereas for AMO at year 10 is not stabilised yet).
2. Understanding the model drift Conclusions
1. Initial conditions to minimise model drift THANK YOU !!
1. Initial conditions to minimise model drift Impact on the skill (correlation) of SST
Initialized NOTROP Initialized NOEQ Uninitialized
Detrended forecasts
Ocean Initialisation: Nudging experiments (3D damping)
)(1
... oTTt
T
)(1
... oSSt
S
β = f(depth, space)
Reanalysis
Current
No nudging within
the Equatorial band
(1oN-1oS) and
Near the coast
(300km)
(1/ β) =0
In the thermocline
(1/ β) =0
Deep Ocean
β = 360 days
Below thermocline
β = 10 days
Ocean Initialisation: Nudging experiments (3D damping)
)( 1 obsk
o
nsns SSTTdT
dQQQ
Heat flux:
Heat flux at
the surface
feedback term.
SSTobs= observations
dT
dQ Feedback coefficient
= -40W/m2/K
Fresh water flux:
1
13
1 )(
k
obsktso
S
SSSSeEMPEMP
Fresh water
budget at
the surface
Feedback term.
SSSobs= observations
Feedback parameter
= -167 mm/day
s
Sea Surface restoring