The Regional Ecosystem Modeling Intercomparison Testbed Project

1
The Regional Ecosystem Modeling Intercomparison
Testbed Project
Marjorie Friedrichs and Raleigh Hood Larry
Anderson Rob Armstrong Fei Chai Jim
Christian Scott Doney John Dunne Jeff
Dusenberry Masahiko Fujii John Klinck Dennis
McGillicuddy Markus Schartau Yvette Spitz Jerry
Wiggert
2
Objective
To quantitatively compare pelagic ecosystem
models in a standardized one-dimensional framework

• Numerous ecosystem model structures and
  parameterizations.
• Few (if any) quantitative comparisons.
• Which ecosystem structures are most robust?
• How much complexity is justified?
• Is it feasible to develop models that are
• applicable over many diverse ecosystems?

Tool Regional Modeling Testbeds
3
Definition

• A regional testbed provides a framework for
  quantitatively comparing different ecosystem
  models and formulations. This framework includes
• One-dimensional physical forcing fields
• (from 3-D GCM or data)
• Biogeochemical data time series
• (for evaluation or assimilation)
• Tools for data assimilation
• Intra-site and Inter-site comparisons

4
US JGOFS Arabian Sea/EqPac Process Studies
Available data Sediment trap
1 year time series Cruise observations
nitrate chlorophyll production
zooplankton
5
Methods Ecosystem model descriptions

• Models 1-4 N, P, Z, D (NH4,DOM, Cchl, T) (CCMA,
  McCreary, Hood, Anderson/McGillicuddy)
• Models 5-6 2P, 2Z, Fe (Christain, Wiggert)
• Model 7 2P, 2Z, Si (Chai)
• Model 8 2P, 3Z, Si, DOM (Fujii)
• Model 9 2P, 4Z, DOM (Laws/Hood)
• Model 10 C, Alk, P, Z, 2DOM (Schartau)
• Model 11 3P, 0Z, 3DOM, Si, Fe (Dunne)
• Model 12 3P, 1Z, 4DOM, Si, Fe (Dusenberry/Doney/M
  oore)
• MM Mean Model
• LST Least Squares Test (N,P,Z,D)
  (Friedrichs/Hood/Wiggert/Laws)

6
Initial model-data comparison (pre-assimilation)
Model data misfit
LST 1 2 3 4 5 6 7
8 9 10 11 12
Model Number (increasing complexity)
-Model performance is quite variable.-Is this
variability due to model structure or tuning?
7
Methods
Problem How do we determine whether
differences in model results
are due to differences in model structure, or
differences relating to parameter
choices? How do we compare these models in
an objective way?
Approach
Variational Adjoint Method of Data Assimilation
-Objectively optimize all of the models
-Discern differences due to model
structure
8
Methods Variational Adjoint Method
Optimize model parameters using a weighted
least-squares cost function measure of
model-data misfit
2 5 Nij
cost function ? ? ? Wijn (Xdat-Xmod)2 Wijn
sij-2 Nij-1 Si kj
i j n
-Minimize cost using the variational adjont
method by adjusting a subset of model parameters,
selected by examination of the Hessian
matrix -Assess model performance based on the
magnitude of the cost function -The cost function
therefore precisely defines model performance,
i.e., which biogeochemical quantities matter and
how much they matter
9
Experiments

• Experiment 1 Simultaneous assimilation
• Assimilate data from both EqPac and Arabian Sea
  sites
• Experiment 2 Individual assimilation
• Assimilate EqPac and Arabian Sea data
  individually
• Experiment 3 Cross assimilation
• Assimilate EqPac data and use the resulting
  parameters to compute Arabian Sea fields
• Assimilate Arabian Sea data and use resulting
  parameters to compute EqPac fields
• Measure of portability

10
Cost function comparison Exp 1 2
Cost function
MM LST 1 2 3 4 5 6 7 8 9
10 11 12
1P, 1Z multi-Z 3P, 0-1Z
-Most models do significantly better for
individual assimilation-Only 4 models do
substantially better than MM Mean Model-More
complex models (5-12) show greater variability
in performance and they do not necessarily
perform better than the simple NPZD models (1-4)
11
Portability
portability
portability
-Models with variable Cchl ratios and
temperature dependence are more portable
12
Portability
portability
portability
- Models with more P and Z size classes are not
necessarily more portable than models
with single P and Z size classes
13
Variability
Arabian Sea Primary Production
EqPac Primary Production
-Models with 3P and simple Z reproduce PP
variability best in the Arabian Sea-Models with
multiple Z groups reproduce PP variability worst
at both locations
14
Conclusions

• The performance of more complex models is more
  variable and they do not always perform better
  than simple NPZD-type models
• More complex models are not necessarily more
  portable than simple NPZD-type models
• Most models perform less well - at least in terms
  of reproducing the data - than the LST and MM
  models (ouch!)
• Models with variable CChla ratios and
  temperature dependent parameters are more
  portable than models without these attributes
• Models with multiple phytoplankton groups and low
  Z diversity estimate primary production
  variability best
• There appear to be some superior model
  structures (e.g., model 11 with multiple P groups
  and implicit Z)

15
Extra slide Mean production and chlorophyll
All models underestimate PP most underestimate
chlImplicit and 1 Z models are better able to
reproduce PP data