Norwegian Climate Prediction Model

From Norcpm
Revision as of 08:33, 19 December 2018 by Fco041 (talk | contribs)

The Norwegian Climate Prediction Model (NorCPM) is aiming at providing prediction from seasonal-to-decadal time scale. It is based on the Norwegian Earth System Model (NorESM, [1]) and the Ensemble Kalman Filter (EnKF, [2]) data assimilation method. NorESM is a state of the art Earth system model that is based on the Community Earth System Model (CESM, [3]) but uses a different aerosol/chemistry scheme and ocean model (evolved from MICOM).

The Norwegian Climate Prediction Model

The EnKF is a sequential data assimilation method that allows for fully multivariate and flow-dependent corrections using a covariance matrix produced by a Monte-Carlo ensemble integration.

NorESM model versions used in NorCPM

Version atmosphere/land resolution ocean/sea ice resolution description reference
NorESM1-L T31 bipolar gx3v7 (~3°) Zhang et al., 2012
NorESM1-LT 1.9°x2.5° tripolar 1° Wang et al. 2017, Kimmritz et al. 2018
BCCRFAST 1.9°x2.5° tripolar 1° Gao et al. 2018
NorESM1-ME 1.9°x2.5° bipolar 1° CMIP5; and a version in preparation for CMIP6 Bentsen et al. 2013, Tjiputra et al., 2013
NorESM1-ACPL 1.9°x2.5° bipolar 1° Anomaly coupling modifies coupling fields by a fixed seasonal climatological correction,
such that biases in SSTs and surface wind stress are obviated.
Toniazzo and Koseki 2018
NorESM2-MH bipolar 1/4° Langehaug et al. 2018

NorCPM Versions

FREE: refers to an ensemble simulation of NorESM carried without assimilation and starting from the same initial ensemble than the assimilation experiment. The initial ensemble is generated by spinning up an ensemble of states (sampled from a long preindustrial forcing run) with real historical forcing from 1850. Such an experiment provides a benchmark for data assimilation and is also used to disentangle the part of prediction skill related to natural internal variability from the part driven by external forcing.
Version 0 (V0): refers to version of NorCPM that assimilates SST only. SST is the only observational dataset available in the ocean for a period of time sufficient (> 100 years) to clearly demonstrate the skill of decadal prediction. Assimilation updates vertically the full ocean while the remaining components of the Earth system model (atmosphere, sea ice, land) are left unchanged but they will adjust dynamically between the monthly assimilation step (an approach referred as weakly coupled data assimilation).

V0 was first tested in idealised twin experiment (Counillon et al. 2014). It is found that assimilation reduces error and can constrain well the variability of the ocean – with the largest improvements in the near surface and sea ice with some benefit over land for temperature and precipitation. The system beats persistence forecast and shows skill for the heat content in the Nordic Seas that is close to the upper predictability limit.
V0 was tested in a real framework and a long stochastic reanalysis was produced for the period 1950—2010 which assimilates the anomaly of the HadiSST2 within the period 1950-2010. HadiSST2 provides a stochastic reconstruction of SST - that is, providing a three-dimensional estimate of the measurement and its accuracy - for the period 1850-2010. A method referred to as upscaling (Wang et al. 2016) is used to ensure that the assimilation does not introduce a drift when updating the non-Gaussian distributed layer thickness variables. The system can reproduce well the North Atlantic variability and shows good agreement with the independent objective analysis of the oceanic heat content and salt content globally. It was noted that using a flow-dependent data assimilation method and formulating the ocean covariance in isopycnal coordinates are an important ingredient for efficiently propagating the surface information below the mixed layer in the Labrador Sea and to constrain the formation of deep water convection.
In Wang et al. (sub) hindcasts are started from a shorter reanalysis (started in 1980) reanalysis. The system shows highly competitive skill compared to North American Multi-Model Ensemble and skillful skill for sea ice extent was variability is driven by ocean variability (e.g. in the boreal winter in the Barents Sea, the Labrador Sea and Greenland Iceland Nordic Seas). The skill of decadal hindcasts was tested for 1955:2010. There is some skill for prediction of AMO and AMOC at 26 but the prediction of the SPG is poor despite a very good match during the reanalysis. Bethke et al. in prep identified that the reason for the poor forecast is a combination of poor initialisation of the deep subtropical water temperature and near-surface salinity in the SPG region.


Version 1 (V1): the system is complemented with the assimilation of hydrographic profiles. Assimilating observations in an isopycnal coordinate model is not straightforward as the observation operator must interpolate either from isopycnal coordinates to z coordinates or vice versa. In Wang et al. 2016 it is shown that the approach to interpolate the model onto z coordinates (still keeping the covariance in isopycnal coordinates) is more linear than interpolating the observations to isopycnal coordinates and as such more efficient. Practical implementations of localisation and the representation error were extensively tested and V1 has been run with an optimal setting. The system is shown to be able to constrain well the error in the interior while being reliable. The performance of the prediction was tested for the period 1980-2017 in the real framework with anomaly assimilation. While complementing the system with hydrographic profile yields little benefit on the seasonal time scale, it greatly enhances the skill for decadal predictions in the SPG region (Bethke et al. in prep). There are currently different version of V1. In V1a, all ocean observations are kept and we do not update the sea ice compartment during assimilation. In V1b, we reject observation if it is located in places where there is ice. In v1c all observations are retained but the error for hydrographic profiles is inflated by a factor of 3 in sea ice covered region because there is large uncertainty for the climatology there (with anomaly assimilation). We also update the sea ice compartment (strongly coupled DA). Most difference between v1a, v1b, and v1c is in sea ice covered region and v1c is performing best there. V1c will be the default version for CMIP6 DCPP.


Version 2 (V2): the system is complemented with assimilation of sea ice concentration. In Kimmritz et al. 2018, we tested different implementations of the data assimilation system in an idealised twin experiment. It is shown that a joint update of the ocean and the sea ice state during the assimilation is beneficial (strongly coupled data assimilation) with a flow-dependent covariance method. It is also strongly beneficial to include the different thickness categories in the state vector. Assimilation is able to constrain well errors in sea ice and in the near-surface ocean. The method is tested in the real framework in Kimmritz et al. in prep. The system show reduced error for sea ice thickness. Prediction of sea ice extent are also greatly enhanced in many regions were sea ice yields predictability.


For each of the NorCPM version the assimilation are either carried using full field assimilation or anomaly assimilation. In full field assimilation the observations are assimilated as they are. An advantage of that approach is that it constrains the model bias, but there is a risk that models are attracted to their bias climatology. In such a case, the redundant corrections will transfer the bias from the observed variables to the non-observed variables via the covariance during assimilation. In anomaly assimilation the anomaly of model and observation are calculated from their respective seasonal climatology before they are compared. As a consequence, bias is left as they are and we aim only at synchronising the variability. This approach also has some drawback as bias in the mean state comes with a bias in its variability (for example with the classical bilocation of the Gulf Stream). Carrassi et al. 2014 details the advantage and inconvenient of both approaches.


Projects funding the NorCPM activities


Current:NFR-SFE (2018-2021), EU-Blue-Action (2016-2019), Norforsk-ARCPATH (2016-2020), EU-INTAROS (2016-2020), BFS-BCPU(2018-2021); EU-TRIATLAS(2019-2024)
Completed:NFR-SNOWGLACE (2015-2018), NFR-EPOCASA (2014-2017), EU-PREFACE (2014-2017), SKD-PARADIGM (2015-2017), SKD- INCREASE (2015-2017), SKD-PRACTICE (2012-2015)


NorCPM activities receive a grant for computer time from the Norwegian Program for supercomputer (NOTUR2, project number NN9039K) and a storage grant (NORSTORE, NS9039K).

Publications, outreach etc.

NorCPM is contributing to the WMO Lead Centre for Annual-to-Decadal Climate Prediction under the name BCCR. We intend to contribute to CMIP6 DCPP.

  1. Counillon, F., Bethke, I., Keenlyside, N., Bentsen, M., Bertino, L., & Zheng, F. (2014). Seasonal-to-decadal predictions with the ensemble Kalman filter and the Norwegian Earth System Model: a twin experiment. Tellus A, 66. doi:10.3402/tellusa.v66.21074
  2. Wang Y, Counillon F, Bertino L. Alleviating the bias induced by the linear analysis update with an isopycnal ocean model. Quarterly Journal of the Royal Meteorological Society. 2016. https://doi.org/10.1002/qj.2709
  3. Counillon F, Keenlyside N, Bethke I, Wang Y, Billeau S, Shen M-L, et al. Flow-dependent assimilation of sea surface temperature in isopycnal coordinates with the Norwegian climate prediction model. Tellus. Series A, Dynamic meteorology and oceanography. 2016;68:32437.
  4. Gleixner S.; Keenlyside N.; Dimissie T., Counillon F., Wang Y.; Viste E. Seasonal predictability of Kiremt rainfall in CGCMs, Environmental Research Letters 2017.
  5. Wang Y, Counillon F, Bethke I, Keenlyside N, Bocquet M, Shen M-L. Optimising assimilation of hydrographic profiles into isopycnal ocean models with ensemble data assimilation. Ocean Modelling. 2017;114.
  6. Kimmritz M., Counillon F., Bitz C.M., Massonnet F., Bethke I., Gao Y. Optimising assimilation of sea ice concentration in an Earth system model with a multicategory sea ice model, Tellus A., 2018
  7. Wang et al. Seasonal predictions initialised by assimilating sea surface temperature observations with the EnKF, submitted
  8. Kimmritz et al. Added value of sea ice assimilation for seasonal prediction in the Arctic, in prep
  9. Jackson et al. North Atlantic circulation: a perspective from ocean reanalyses.
  10. Bethke et al. Impact of subtropical North Atlantic initialisation errors on subpolar gyre prediction in prep.
  11. Ogawa et al. Arctic sea ice has no influence on AO/NAO, in prep.
  12. Counillon et al. Relating model bias and prediction skill in the tropical Atlantic, in prep.
  13. Fransner et al. What yields predictability of biochemistry for seasonal to decadal time scale, in prep.

User Resources

All the code for running the system is available on GitHub. The repository is private so please request to an account to access it. [https://github.com/BjerknesCPU/NorCPM]


Contact informations


Leader Prof. Noel Keenlyside;
NorESM related questions Dr. Ingo Bethke, Alok Gupta;
Data assimilation related questions Dr. François Counillon, Dr. Yiguo Wang;
Sea ice prediction Dr. Madlen Kimmritz, Dr. François Counillon
Atmospheric nudging Dr. Mao-Lin Shen, Dr. Fei Li;
Biochemistry predictions Dr. Filippa Fransner

Existing runs

Following is a table that summarise the different experiment runs available so far:

All data are available on Norwegian storage facilities NIRD norwegian . All path below are given relative to the path /projects/NS9039K/shared/norcpm/cases/NorCPM/. If you have access to NIRD but not to NS9039K contact Noel Keenlyside or Ingo Bethke otherwise contact the person with relevant expertise as detailed in the Contact information section.


Reanalysis
Name NorESM version full_field/anom NorCPM version Obs data set ens size assim freq period forcing var updated localisation Path Publications using it
Reana_twin_PI NorESM1-L full_field V0 Micom SST 30 monthly 610-710 PI ocn(all) poin hor. no vert tape Counillon et al. 2014
Free_PI NorESM1-L FREE - - 30 - 610-710 PI - - tape Counillon et al. 2014
Free_NorESM1 NorESM1-ME FREE - - 30 - 1850-2010 CMIP5 hist - - ../NorESM1-ME_historicalExt_noAssim/ Counillon et al. 2016
Reana-V0-long NorESM1-ME anom V0 HadiSST2 30 monthly 1950-2010 CMIP5 hist ocn(all) poin hor. no vert True_Obs-1950-2010/ME/ Counillon et al. 2016, Bethke et al. in prep
Reana-V0-short NorESM1-ME anom V0 HadiSST2 30 monthly 1980-2010 CMIP5 hist ocn(all) 30 obs True_Obs-1980-2000/ana_19800115_me/ Wang et al. sub
Reana-bccrfast_v0 BCCRFAST anom V0 HadiSST2 30 monthly 1950-2010 CMIP5 hist ocn(all) poin hor. no vert ana_19500115_bccrfast none
Reana-bccrfast_v1 BCCRFAST anom V1 HadiSST2+EN4.1 profiles 30 monthly 1980-2010 CMIP5 hist ocn(all) Wang et al 2017, no vert ana_19791115_bccrfast_SSTTS none
Free_bccrfast BCCRFAST FREE - - 30 - 1850-2010 CMIP5 hist - - ../bccr-fast_historical_18500101/ none
Reana-v1a NorESM1-ME anom V1 Hadisst2+EN4.1 profiles 30 monthly 1980-2010 CMIP5 hist ocn(all) Wang et al 2017, no vert NorCPM_V1/ana_19800115_me/ Jakson et al. in prep, Bethke et al. in prep
Reana-v1-F NorESM1-ME full-field V1 Hadisst2+EN4.1 profiles 30 monthly 2000-2010 CMIP5 hist ocn(all) Wang et al 2017, no vert tape Wang et al. 2017
SFE NorESM1-ME anom V1 NOAA-SST+EN4.2 profiles 30 monthly 2006-present CMIP5 hist ocn(all) Wang et al 2017, no vert True_Obs-2006-2017/reanalysis/ none
Reana-acpl NorESM1-acpl anom V1 Hadisst2+EN4.1 profiles 30 monthly 1980-2010 CMIP5 hist ocn(all) Wang et al 2017, no vert Anomaly_coupled/ Counillon et al. in prep
Free_twin_V2 NorESM1-LT FREE - - 30 - 1000-1020 PI - - Twin_experiment/TWIN_Free Kimmritz et al. 2018
Reana_twin_V2 NorESM1-LT full field V2 micom_icec 30 monthly 1000-1020 PI ocn(all) + ice(all) 800 km Twin_experiment/TwinA56/ Kimmritz et al. 2018
Reana-v2-F NorESM1-ME full field V2 Hadisst2+EN4.1 profiles 30 monthly 1980-2010 CMIP5 hist ocn(all) + ice(all) Wang et al 2017, no vert, 800 km NorCPM_V2/ana_me_ICEC-SST-S-T-1980-2010/
Reana-v2-a NorESM1-ME anom V2 Hadisst2+EN4.1 profiles 30 monthly 1980-2010 CMIP5 hist ocn(all) + ice(all) Wang et al 2017, no vert, 800 km NorCPM_V2a/ana_me_ICEC-SST-S-T-1980-2010/ Kimmritz et al. in prep, Ogawa et al. in prep
Reana-v1b NorESM1-ME anom V1 Hadisst2+EN4.2 profiles; masked in sea ice 30 monthly 1980-2010 CMIP5 hist ocn(all) Wang et al 2017, no vert NorCPM_V1b/ana_19800115_me/ Ogawa et al. in prep
Reana-v1c NorESM1-ME anom V1 Hadisst2+EN4.2 profiles 30 monthly 1980-2010 CMIP5 hist ocn(all)+ice(all) Wang et al 2017, no vert NorCPM_V1c/ana_19800115_me/


Hindcats
Name Reana used length Model period frequency enssize path papers
Dec-hind-V0 Reana-V0 10 years NorESM1-ME 1955-2010 Every 2 years 20 True_Obs-1950-2010/ME_hindcasts/ Bethke et al. in prep
seas-hind-V2F Reana-v2-F 13 months NorESM1-ME 1985-2010 4 times per year 9 NorCPM_V2/SeasHind_ana_me_ICEC-SST-S-T-1985-2010/ none
seas-hind-V2a Reana-v2-a 13 months NorESM1-ME 1985-2010 4 times per year 9 NorCPM_V2a/hindcast/ Kimmritz et al. in prep
Dec-hind-V1a Reana-V1a 10 years NorESM1-ME 1985-2010 Every 2 years 5 NorCPM_V1/ana_19800115_me_dec_19851015/ Bethke et al. in prep
seas-hind-V1a Reana-V1a 13 months NorESM1-ME 1985-2010 4 times per year 9 NorCPM_V1/ana_19800115_me_hindcasts/ Wang et al. in prep
seas-hind-V1b Reana-V1b 13 months NorESM1-ME 1985-2010 4 times per year 9 NorCPM_V1b/ana_19800115_me_hindcasts/
seas-hind-V1c Reana-V1b 13 months NorESM1-ME 1985-2017 4 times per year 9 NorCPM_V1c/ana_19800115_me_hindcasts/
seas-hind-acpl Reana-acpl 13 months NorESM1-ME 1985-2017 4 times per year 9 Anomaly_coupled/acpl_19800115_me_hindcasts_* Counillon et al. in prep