Status of CESM data assimilation using the DART systemStatus of CESM data assimilation using the...
Transcript of Status of CESM data assimilation using the DART systemStatus of CESM data assimilation using the...
Status of CESM data assimilation using the
DART system
Alicia R. Karspeck ([email protected])
DARES personnel: Nancy Collins, Jeff Anderson, John Hendricks, Kevin Raeder, Tim HoarCISL/ASAP: Brian Dobbins
CGD software-engineers:Mariana Vertenstein, Steve Goldhaber, Brian Eaton
CGD personnel:Gokhan Danabasoglu, Fred Castruccio, Steve Yeager, Joe Tribbia
Goals of data assimilation development within CGD
We do not have mandated operational assimilation or prediction responsibilities
Flexibility to explore challenging projects that are relevant to the broader community
Initialized prediction with
the CESM component
modelsmodel
improvement
data assimilation
research
Build tools and capabilities for the CESM research community that are relevant for
”
”
Ensemble Data Assimilation
“DART”
Modular data assimilation:• Can be used with ANY model or
combination of models• Can be used with any observational
data source (provided a forward operator can be coded)
• Can be used to test different algorithms (only ensemble-based at this point)
• Leverage methodological advances across a wide community
Modular Earth System Model:• Can be configured with different
components of the climate system• Different versions of the same
component• Broad community of scientific users
from which to draw expertise
Modularity in Model and DA framework
Different configurations ofCESM + DART
• “Weakly-coupled” DA system keeps operates separately for each component• Coupler exchanges fluxes between component model• “Weakly-coupled” DA system keeps operates separately for each component• Coupler exchanges fluxes between component model
• “Strongly-coupled” DA system has a unified state-vector• Coupler exchanges fluxes between component model + • Observations from one component system can impact other components
• “Strongly-coupled” DA system has a unified state-vector• Coupler exchanges fluxes between component model + • Observations from one component system can impact other components
• “Single component” DA system• “Single component” DA system• “Ensemble OI” uses pre-computed ensemble perturbations• Do not run ensembles of model integrations (much less expensive)• “Ensemble OI” uses pre-computed ensemble perturbations• Do not run ensembles of model integrations (much less expensive)
Provides State‐of‐the‐Art Ensemble Initial Conditions for prediction with these CESM models:Model Type Component Some Univ Collaborators
Atmosphere CAM‐FV, CAM‐SE, WACCM, WACCM‐X,CAM‐CHEM,MPAS
UC Berkeley, SUNY Stonybrook
Ocean POP University of Miami
Land CLM University of Texas, University of Arizona,University of Maryland
Sea Ice CICE University of Washington
Coupled System CAM‐FV/POP
Data Assimilation Research Testbed (DART)
Slide courtesy of N. Collins/J. Anderson
A few highlights from our prototype coupled CESM/DART coupled ocean-atmosphere assimilation experiment
• Took us about 2 years wall-clock to complete about 20-years• Only ~12 years of this was useable (Jan 1970 - Feb 1982)
• 30 ensemble member DART EAKF
• Conventional (in-situ) data from the ocean and atmosphere components were assimilated (primarily radiosondes in the atmosphere and XBTs in the ocean)
Karspeck et al (in prep)
CESMDART weakly-coupled ocn/atm configuration
Constrains interannual variability
Upper ocean heat content Sea level pressure
CESMDART
NoAssim
[1970-1981]
I
CESMDART weakly-coupled ocn/atm configuration
Constrains large-scale modes of climate variability
MJO from J. Eliashiv (Scripps)
I
CESMDART weakly-coupled ocn/atm configuration
Constrains synoptic atmosphere variability in N. hemisphere
6 hour forecasts
+ JRA55 6-hrx CFSR/NCEP 9-hro ERA40 NH 9-hr
I
CESMDART weakly-coupled ocn/atm configuration
Constrains synoptic atmosphere variability in N. hemisphere
Major Hurricane “Harvey”September 1981
time in days
SLP SST SST AVHRR
• System is currently costly due to
1) ensemble integrations of the CESM model (“fixed cost”)2) The DA update within DART (“fixed cost”) 3) *Costs associated with how the CESM and programs DART interact*
• NCAR/CGD is addressing one known bottleneck through “Pause-Resume” • Pause-resume will enable CESM component models to interact with
DART without “stopping integration” • Funds are only sufficient to prototype a proof-of-concept
• CISL/ASAP has devoted some work toward profiling inefficiencies in the joint-system (much more to be done).
• Opportunities for further innovation (a few examples)• Ensemble management/optimization/fault tolerance in CESM• Embedding forward observation operators within component models• Passing data between CESM/DART in memory (as opposed to files)
Progress towards a CESM+DART that is faster and
more user friendly
• The CESM + DART system can be used to provide ensemble initial conditions for all of the major CESM component models
• Investment in the joint-CESM+DART infrastructure is needed make the system more accessible for individual PIs.
• The value of native-model data assimilation (over other strategies) for initialized prediction has been shown in only a handful of controlled studies, but the CESM+DART can provide the framework for this type of applied research
• Should delivering data-assimilation/initialization capabilities for CESM models be a CGD priority? How can we fund those development efforts, interact more effectively with the DART group?
Key take away-points:
Discussion Question:
END
Open questions regarding coupled assimilation that groups are trying to answer:
Best practical path for migrating to coupled DA? • Extend exiting systems? Start from new systems?
Does coupled DA improve initial conditions and/or forecast performance?• Under what conditions? (perfect model? With model error?)
Does strongly-coupled DA improve over weakly-coupled?• At what lead times? For what problems?
For long-lead forecasts, can the need to do “state-of-the-art” DA be relaxed? • Reduced observational networks, reduced complexity models, less
sophisticated DA methods?
• CESMDART (CAMDART/POPDART/CICEDART/CLMDART) have generally shown high performance
• System is very currently costly (too costly for us to do more than prototyping)
• Addressing this (in part) through “Pause-Resume” efforts (funded jointly by the NCAR/CGD directorate funds – will only be prototyped with existing funds)
• Implementation of multi-coupler
• More options available, however, for making CESM-specific models work more effectively with DART and bringing DART into the CESM framework (requires software engineering support that we do not currently have)
• Also DA/CESM systems tend to require “tuning”; maintenance of appropriate data sources; setting up of test cases; etc (not unlike the role of component model developers to provide a useable system)
• Should delivering data-assimilation/initialization capabilities for CESM models be a CGD priority? How can we fund those development efforts?
Summary statements/Discussion questions