sign in sign up
<<
Home , Model (person)

Recent development of a - an interactive multi- ensemble

Shuo Wang, François Counillon,, Shunya Koseki, Noel Keenlyside,, Alok Kumar Gupta, Maolin Shen

1. Geophysical Institute, University of Bergen and Bjerknes Centre for Climate Research, Bergen, Norway 2. Nansen Environmental and Remote Sensing Centre and Bjerknes Centre for Climate Research, Bergen, Norway 3. NORCE Norwegian Research Centre AS and Bjerknes Centre for Climate Research, Bergen, Norway Persistent model biases along model development Multi-model precipitation biases across model generations

Tian and Dong 2020 Bias is often larger than the signal we analyze or predict

Palmer and Stevens, PNAS, 2019

Can we learn from model diversity to train a better climate model? Standard modelling VS Supermodelling

A supermodel is an optimal dynamical combination of models that is superior to its individual constituent models Supermodel improved the predictions skill

• Extensively tested with toy models, e.g. Lorenz 63 system (van den Berge et al., 11, Mirchev et al. 12) and AGCMs (Wiegerinck et al., 17; Schevenhoven et al. 17) • Can remove structural error such as double ITCZ with two AGCMs coupled to an OGCM (Shen et al. 16,17) • Improves short term weather forecasts with the intermediate complexity climate model SPEEDO (Schevenhoven et al., 19)

Our work focuses on fully coupled Earth system models An ocean connected super-ESM

Ocean Atmosphere ECHAM 6 Atmosphere MPIESM MPIOM (z-coordinate, 1.5°) ECHAM 6 (2°) CAM5 CAM4

CESM POP (z-coordinate, 1°) CAM5 (1°)

MPIOM POP NorESM MICOM (isopycnal, 1°) CAM4 (2°) MICOM

Ocean

Recursive approach: 1. The three models are propagated for 1 month 2. Generate pseudo-observations (here SST only) from the three individual CMIP5 ESMs (weighted mean) 3. Assimilate the pseudo-observations back into each models (correct the full ocean state) with the Ensemble Optimal Interpolation (Counillon et al. 2009) 11 nodes 1408 CPU, ~10 model-year per day Experimental set-up

We compare the performance of different approaches for 1980 to 2006: 1. A posteriori averaging of the individual models (Standard) 2. Supermodel with equal weight (EW) 3. Supermodel with spatially and monthly varying trained weight (TW). • Weights estimated offline from the performance of the individual model reanalyses with assimilation of real SST • Weights estimated as in Particle Filter (Van Leeuwen 2009); weights are positive and normalised SST bias against OISST for 1980 - 2006 Standard EW TW MPIESM CESM NorESM Multi-model mean SST bias

Standard

EW TW Annual-Mean Climatology of tropical precipitation

AnnualAnnual-Mean-Mean Climatology Climatology of oftropical tropical precipitation precipitation GPCP Unconnected Equal weight (offset) 30N 30N 30N 15N 15N Annual-mean precipitation climatology15N in the tropical Pacific EQ EQ EQ GPCPGPCP Unconnected UnconnectedStandard Equal Equalweight weight (offset) (offset) 15S 15S 30N30N 30N30N15S 30N 30N

Annual-Mean Climatology30S of tropical precipitation30S 15N15N 15N15N30S 15N 15N 0 180 360 0 180 360 0 180 360 EQEQ EQ EQ EQ EQ

15S15S 15S15S 15S 15S 30S 30S 30S 30S 30S 30S 0 180 360 0 180 360 0 180 360 0 180 360 0 180 360 0 180 360 GPCP UnconnectedBBW BBW_EnOI Equal weightOWEW New (offset) PFTW 30N 30N30N 30N 30N30N 30N

15N 15N15N 15N 15N15N 15N

EQ EQEQ EQ EQEQ BBWEQ BBW_EnOI OW New PF 30N BBW30N BBW_EnOI 30N OW New30N PF 15S 15S 15S30N 30N 30N 30N 15S 15S 15N15S 15N 15S 15N 15N 15N 15N 15N 15N 30S 30S30S 30S 30SEQ30S EQ 30S EQ EQ 0 180 3600 0 180 180 360360 0 180 360EQ0 0 180 180 360 360EQ 0 180 360 EQ EQ 15S 15S 15S 15S 15S 15S 15S 15S 30S 30S 30S 30S 30S0 180 360 30S0 180 360 030S 180 360 030S 180 360 2 4 60 8 10 (mm/day) 180 360 0 180 360 0 180 360 0 180 360 BBW BBW_EnOI OW New PF 30N 30N 30N 30N 2 4 6 8 10 (mm/day) 15N 15N 15N 15N 2 4 6 8 10 (mm/day) EQ EQ EQ EQ

15S 15S 15S 15S

30S 30S 30S 30S 0 180 360 0 180 360 0 180 360 0 180 360

2 4 6 8 10 (mm/day) Synchronisation of SST

Standard Synchronisation metric:

time-std(multimodel_mean)/time_avg(std(multimodel)) =1 ; i.e. inter models std = climatological variability <1; i.e. inter models std < climatological variability >1; i.e. inter models std > climatological variability

EW TW

With monthly SST synchronization, the tropical Pacific is somewhat connected Time SST variability: NINO 3.4

Standard EW Nino3.4 Nino3.4 Nino3.4 Weighted SUMOTW SST anomalies 4 4 4

3 3 3

2 2 2

1 1 1

0 0 C 0 °

-1 -1 -1

-2 -2 -2 NorESM NorESM NorESM CESM -3 MPIESM -3 MPIESM -3 MPIESM CESM CESM Super Meam -4 -4 1982 1985 1987 1990 1992 1995 1997 2000 2002 2005 -4 1982 1985 1987 1990 1992 1995 1997 2000 2002 2005 1980 1985 1990 1995 2000 2005 Year

• Internal variability in tropical Pacific is partially synchronised • The variability is little damped, but less than the standard ensemble mean. Standard deviation of SST anomalies Observed Standard models average

EW supermodel mean TW supermodel mean

Deg C

Supermodel mean simulates reduced SST variability (less than standard models mean), with some indications of improved pattern Summary

• A supermodel based on 3 ESMs are connected by ocean assimilation • It reduces mean biases of SST and precipitation • Improvements greater than the standard ensemble mean, because of non-linear properties of the climate system • Synchronisation is achieved in some regions with connection of monthly SST • Internal variability is dampens because averaging of unsynchronised variability in the pseudo observation causes a spurious deflation during assimilation • We working towards improving the system: • improving connectivity (connection of ocean & atmosphere; more frequent) • Isolate the part of variability that can be synchronised • Training the weights online We aim to perform climate prediction and climate change experiments