STRONGLY COUPLED ENKF DATA ASSIMILATION WITH … · Oct 19, 2016 Sluka - CDAW Toulouse 30 • Naïve implementation of vertical localization fixed to 50m • Detrimental impacts below

Travis Sluka Acknowledgements: Eugenia Kalnay, Steve Penny, Takemasa Miyoshi

STRONGLY COUPLED ENKF DATA ASSIMILATION

WITH THE CFSV2

CDAW – Toulouse Oct 19, 2016

Outline

Oct 19, 2016 Sluka - CDAW Toulouse 2

1. Overview of strongly coupled DA with LETKF

2. OSSE with intermediate complexity model SPEEDYNEMO-LETKF

3. Initial real data experiments with CFSv2-LETKF

OCEAN ATMOSPHERE

Weakly Coupled DA

A single coupled model generates background for separate DA systems. Already operational or planned at many centers.

• EnKF

GFDL CDA (Zhang et al. 2007)

• 3D/4DVAR CDAS, Met Office, ECMWF (Saha, 2010 Laloyaux et al., 2015; Lea et al. 2015)

Oct 19, 2016 Sluka - CDAW Toulouse

Coupled OCN/ATM

model

Data Assimilation

Data Assimilation

analysis

analysis forecast

forecast

ATM obs

OCN obs

3

Strongly Coupled DA

Studied w/ simple models:

• EnKF (Lu et al., 2015; Liu et al., 2013; Han et al., 2013; Luo & Hoteit, 2014; Tardif et al., 2013)

• 4DVAR (Smith et al., 2015)

Studied w/ realistic models:

• EnKF (Lu et al., 2015)

• 4DVar (Sugiura et al., 2008)

• 3D En-var (Frolov, 2016)


Coupled OCN/ATM

model

Data Assimilation

analysis forecast

ATM obs

OCN obs

4

Choice of Data Assimilation

4DVar vs EnKF?

• 4DVar works best with longer time windows (since flow-dependent error covariance is not saved, though propagated implicitly within window)

• EnKF works best at shorter windows Explicitly saves background error covariance

EnKF has several benefits • Don’t need to develop coupled TL/adjoint

• Cross domain covariance generated automstically by the ensemble

• Assimilation can be performed at the shorter window of the atmosphere

• Most importantly… should be easier, I want my PhD at some point!


Local Ensemble Transform Kalman Filter

• LETKF – one flavor of the EnKF, operates by independently and in parallel creating analysis one grid point at a time, using a subset of observations around it.

• Parallelization scales very well

LETKF (Hunt et al., 2006)


observation

model grid point



analysis mean analysis ensemble perturbations

Observation operator

LETKF calculates analysis for each grid point in parallel using subset of obs

around it. For weakly coupled DA, y contains only obs from its own

domain. For strongly coupled DA, y contains obs from all domains.

7


• Cross-domain observation impacts analysis mean


• W varies smoothly across domain interface, allows individual ensemble members to stay “matched” together, allowing for better balance of ensemble members

analysis mean analysis ensemble perturbations

Observation operator

8

Coupled LETKF

• Separate LETKF for each domain helps keep implementation simpler. Ocean LETKF and Atmosphere LETKF can be developed independently.

• Sharing of observational departures allows system to act as single strongly coupled system.

Data assimilation systems, normally separate “weakly coupled DA”

Addition of cross-domain observational departures “strongly coupled DA”


II. SPEEDYNEMO-LETKF Strongly coupled DA with an intermediate complexity CGCM


SPEEDYNEMO nature run – Jet Stream Winds

Sluka, T. C., Penny, S. G., Kalnay, E. & Miyoshi, T. Assimilating atmospheric observations into the ocean using strongly coupled ensemble data assimilation. Geophys. Res. Lett. 43, 752–759 (2016).

SPEEDY-NEMO OSSE

Using the fast SPEEDY-NEMO (one year run takes only 12 hours on 1 core)

• Perfect model OSSE conducted first using only atmospheric observations

SPEEDY-NEMO

• T30 atmosphere

• 2 degree ocean

• Coupling every 6 hours

Experiment parameters

• 40 ensemble members

• Localization: 1000km Hz

• Relaxation to prior spread: 90% for OCN, 60% for ATM


6 hr ATM observations

11

SPEEDY-NEMO Strongly Coupled DA STRONG-WEAK analysis RMSE


Salinity [PSU] Temperature [C]

Surface

Deeper Ocean

(500m-2km)

Northern Hemisphere

Southern Hemisphere

Tropics

Global

Imp

rovem

ent

Imp

rovem

ent

Imp

rovem

ent

Imp

rovem

ent

Analysis RMSE improvement of ocean, from strongly coupled DA of simulated atmospheric observations

12

-52%

-37%



OCN Salinity OCN Temperature

Upper 500m

Pacific

Atlantic

13


• The opposite experiment (assimilating OCN obs into the atmosphere) shows improvement as well

• A possible latitudinal dependence on cross domain impact?


Oce

an

Ob

serv

atio

ns

Atm

osp

her

e

Ob

serv

atio

ns

atm U

STRONG-WEAK, blue is good

SPEEDY-NEMO LETKF

• Assimilating ATM obs into the ocean reduces analysis RMSE for both domains

• The opposite experiment (assimilating OCN obs into the

atmosphere) shows improvement as well

• Problems with SPEEDYNEMO prevent further work with 6 hour DA cycles. System might still be useful for research with climate length runs.

• Improvement strongest in NH extratropics during spring/winter months (over 50% RMSE reduction)

4/11/2016 Sluka - Strongly Coupled DA 15

III. CFSv2-LETKF Strongly coupled data assimilation with the Climate Forecasting System v2, using real observations


CFSv2 - SST

CFSv2-LETKF

• Not at all using operational CDAS for CFS.

• Combined existing GFS-LETKF (Lien, 2013) and MOM-LETKF (Penny, 2013)

• T62/L64 atm 0.5deg ocn (reduced

resolution ATM)

• 50 member ensemble (initialized from CFSR, run freely for 6 months to develop sufficient spread)

• observations from operational ATM PREPBUFR and OCN profiles used by GODAS


OCN

ATM

17

marine surface scatterometer rawinsonde

SATWND land surface

aircraft

T&S profiles (ARGO, XBT, moored buoys)

Weakly Coupled DA - JJA


5m OCN T bias

ATM T bias – SFCSHP obs

Weakly Coupled DA – cross covariances


Mixed Layer depth (depth of T10m ± 0.2°C)

June December

• Cross correlations given by the ensemble for a single date • ATM and OCN temperature max correlation of 0.36, highest values in that

hemisphere’s summer, below 850mb and above top of thermocline • June values likely artificially large do to insufficient spin up time for the ocean

Weakly Coupled DA – cross covariances


Mixed Layer depth (depth of T10m ± 0.2°C)

Humidity Windspeed

• ATM_Q x OCN_T correlations weaker, though same pattern as ATM_T x OCN_T • Increase in windspeed correlated with decrease in surface temperature

• These are the correlations LETKF will be using…

Strongly coupled DA


ocn profiles (argo, XBT,…)

ATM SFCSHP T&q

OCN obs (JJA)

ATM SFCHSP obs (JJA)

• 1 way strongly coupled DA • Strongest cross correlations are between

OCN_T and ATM_T/ATM_q, so… • OCN assimilates surface marine T and q as well,

given by the SFCSHP section of the PREPBUFR

SFCSHP T bias


• Known diurnal bias in SFCSHP T, is a problem (sensors placed over warm deck of a ship, no NSST in model)

• also visible in MERRA2 reanalysis (Carton et al., personal comm)

• Still, there are areas of persistent bias of same sign, caused by SST bias in our weakly coupled run, which strongly coupled DA should improve upon

22

JJA average B-O using SFCSHP T

Options for coupled DA windows

• Atm obs every hour, 6 hour window

• Ocn obs once a day

• Synchronous strongly coupled DA, atm updates Ocn at sane time, ev 6 hours


Options for coupled DA windows


• Atm obs every hour, 6 hour window

• Ocn obs once a day

• Synchronous strongly coupled DA, atm updates Ocn at sane time, ev 6 hours

• Asynchronous Atm obs deps saved and only assimilated ev 24 hours

CFSv2-LETKF single obs test

• After 1 month of weakly coupled DA, several locations, especially near coasts, exhibit large SST errors.

• Yellow Sea is chosen for strongly coupled single observation test (synthetic obs)


Feb 1, 2010 00Z

SST background error Surface atmosphere T error

Feb 1, 2010 00Z

obs location

CFSv2-LETKF single obs test • 2 strongly coupled DA experiments

(single ATM T ob [left], single OCN T ob [right])

• Provide similar analysis increment pattern. But, in this case, OCN observation improves atmosphere MORE than ATM observation


Strongly Coupled CFS - results


-13.1%

-2.1%

-3.8%

• Errors in 6 hour background for ATM T are greatly reduced in the NH


• Strong improvement in both ATM and OCN in NH

• Some problems immediately near coast of NA observations too close to land should probably be excluded Low resolution of ATM likely causing problems near gulfstream


5m

15m

25m

35m



• Naïve implementation of vertical localization fixed to 50m

• Detrimental impacts below the mixed layer better -> <- worse

NH TP

% RMSD Improvment

Dep

th (

m)

Mixed layer depth* JJA average

*MLD defined as level where θ = θ10m ± 0.2°C

Caused by naïve fixed vertical localization of ATM observations into ocn (σ=50m), need to limit impact to mixed layer only

Future work


Currently: Progress has been made on one-way strong coupling of ocean/atmosphere model with a subset of real observations with the CFSv2

Next-

• Fun with localization!

• Full ATM observation set (U/V may be more difficult given lower cross domain covariance),

• Full 2-way strongly coupled DA OCN observations (night-time level 2 SST) impacting the atmosphere

31

Ultimate Goal…

• CFSv3 - NCEP transitioning to hybrid-GODAS, based on LETKF for the ocean.

• Increased potential at that point for an operational strongly coupled hybrid-LETKF global DA system


OCEAN

WAVES SEA ICE

ATMOSPHERE

LAND AEROSOL

32

Travis Sluka, and S. Penny, E. Kalnay, T. Miyoshi

STRONGLY COUPLED ENKF DATA ASSIMILATION

WITH THE CFSV2

Travis Sluka University of Maryland

[email protected]

Documents

STRONGLY COUPLED ENKF DATA ASSIMILATION WITH … · Oct 19, 2016 Sluka - CDAW Toulouse 30 • Naïve implementation of vertical localization fixed to 50m • Detrimental impacts below