|
|
|
May 2002 Workshop Results
|
|
|
|
|
|
Workshop Goals
Global biogeochemical research must increasingly address
the problems of "detection" or quantification of changing fluxes to
the atmosphere, and "attribution" or explanation of those fluxes in
terms of specific mechanisms. Today, neither our measurement nor
analysis capabilities are sufficient to meet the twin challenges of
biogeochemical detection and attribution with sufficient accuracy and
resolution. We have held an Advanced Study Institute to study analysis
techniques (inverse and assimilation modeling), observing system
design and the synergism of new measurements with new analysis
frameworks. This Advanced Study Institute involved lectures from a
broad and distinguished group of scientists on biogeochemical cycles,
current and planned measurement capability, process and data
analytical modeling, and new approaches in applied math.
The Institute had as its centerpiece a hands-on simulation
exercise. Estimates of global terrestrial and oceanic fluxes were
produced from existing data and models, combined to produce flux
fields with reasonable time-space variability (designated the "true"
fluxes). They were distributed in a global simulated atmosphere using
an atmospheric transport model to produce a 4-D data set of CO2 concentrations ("true"
concentrations). The participants formed teams to reconstruct the
"true" surface fluxes, given the "true" concentrations at measurement
locations of their choosing. The teams were free to choose any
strategy they wished in choosing the measurements used to estimate the
"true" fluxes, though the setup and operating expenses of their
network had to fall within a total cost constraint. A mesoscale carbon
assimilation technique developed at NCAR was available for the
participants to use in their network design, and to demonstrate
concepts and application of data assimilation in biogeochemistry. This
4-D Var approach iteratively searches for the surface fluxes that best
match the concentration data, starting from a realistic initial guess
of those fluxes (designated the "prior" fluxes).
|
|
Reference Global Atmospheric CO2: Generating the "Truth"
A full 4-D field of simulated atmospheric CO2
concentrations (the "true" concentrations) was generated using the
PCTM transport model and model estimates of the terrestrial biosphere
and oceanic fluxes, along with the Andres et al. fossil fuel emissions
(designated the "true" fluxes). The ocean flux comes from the NCAR
ocean model, run with anthropogenic forcing (meaning that modern day
atmospheric CO2 concentrations
contribute to the gradient across the air-sea interface). The
terrestrial biosphere fluxes were the NEP calculation from LPJ (
Lund-Potsdam-Jena model, McGuire et al. 2002). For this workshop,
the diurnal cycle in the land biosphere fluxes was not modelled.
The participants were charged with designing a measurement strategy
that would lead to the best estimate of the (unknown) "true" fluxes,
given the "true concentrations sampled for their network. The 4-D Var
data assimilation code was to be used to estimate the "true" fluxes
from the data for each network. The participants were given a
realistic first guess of these fluxes (the "prior") consisting of land
biosphere fluxes from the CASA model, and oceanic fluxes from
Takahashi, et al. (1999). The fossil fuel flux used in the prior was
the same as that used in the truth, not a bad assumption, given that
this flux is quite well known. The PCTM transport model used in the
assimilation is the same as that used in generating the "true"
concentrations, so therefore we have removed the transport model error
issue from the assimilation problem.
C-DAS Grid
- Map with color-coded backgrounds: PDF File (13KB) EPS File (659KB)
- Map with color-coded numbers: PDF File (9KB) EPS File (650KB)
- ascii file (26KB)
|
|
Team Network Data
Existing Network:
Red Team's Network:
Blue Team's Network:
In Situ Aircraft Network:
Manual Flask Network:
|
|
|
|
Carbon Data Assimilation
We have developed a variational data assimilation method (4-D Var) to solve for surface CO2 fluxes (air-sea and air-land) at the resolution of our atmospheric transport model [2° x 2.5°, at a 15 minute time step]. This system has been used, as part of this workshop, to evaluate which of the measurement networks designed by the participants would best pin down the real-world CO2 fluxes, when solved for in this type of inversion framework. A set of modeled oceanic, land biospheric, and fossil fuel CO2 fluxes (unknown to the participants) were designated as the "true" fluxes, and a set of CO2 measurements at the chosen measurement sites was derived from these, with appropriate random measurement errors and biases added on. The measurements for the different networks were then used in this 4-D Var assimilation approach to obtain estimates of the "true" fluxes; those networks that gave estimates closest to the truth (in terms of a pre-defined metric) were then identified.
The 4-D Var method works by iteratively improving an initial guess of the CO2 fluxes (or "Prior" fluxes). The initial guess of fluxes used here (known to the participants) was taken from the CASA land biosphere model, and from Takahashi, et al (1999) for the oceans. The same fossil fuel fluxes used in the "truth" case were used as well in the "prior" case. Both sets of fluxes ("truth" and "prior") were then run through the PCTM atmospheric tracer transport model for 2 years, starting from a realistic first guess of concentrations for January 1st.
The evolution of the CO2 concentrations across the full 2 years is given below:
- Two year movie of surface layer CO2 concentrations given by the "True" fluxes: AVI File (23MB)
- Two year movie of surface layer CO2 concentrations given by the "Prior" fluxes: AVI File (24MB)
These concentration fields were then sampled at the different sites chosen during the CDAS Workshop. The difference between these two sets of concentrations provides information about the differences in the two flux fields.
These concentration differences (prior - truth), sampled at the measurement sites and weighted by the inverse of the data error covariance matrix, are then used to force the adjoint transport model as surface flux fields. Starting from zero initial conditions at 24:00Z, December 31, the adjoint transport model, run backwards in time with this forcing, generates adjoint flux fields.
The adjoint flux fields obtained using measurement differences from the Blue network for Year 1 are shown here after running the adjoint backwards for 3 months (Oct 1), 6 months (June 1), 9 months (March 1), and 12 months (Jan 1).
- One year movie of Adjoint fluxes, for Year 1, using measurements from the Blue network: AVI File (14MB)
These adjoint flux maps are then used to compute corrections to the first guess fluxes. These corrected fluxes are then run through the forward transport model again, compared to the true measurements, and the differences used in an iterative optimization technique to converge to the final flux estimate.
|
|
