Experimental
Protocol for Transcom 3 Level 3 (T3L3)
Aim
This page provides
detailed information for participants of T3L3. For background
information on this project see the TRANSCOM
home page and
the
introduction to the TRANSCOM 3
experiment.
So far, the Transcom experiments have had a strong focus on differences
between transport models and the causes and consequences of that
(which,
after all, was the goal of the project). The analyses conducted
in T3L1
and T3L2 allow a rigorous indepth assessment of the estimated
source/sink distributions and their sensitivity with respect to the
different transport models. It is clear, however, that, besides
transport, the results of inversions are also sensitive to many other
aspects of the inversion set-up, such as data selection and errors,
definition of priors and their covariances, and the mathematical
procedures to solve the inversion problem. This sensitivity has not yet
been adressed in the level 1 and 2 excercises, and the results to some
extent give a misleading picture of the constraints on carbon cycle
fluxes that inverse modelling can provide. Transcom 3 level 3 aims at a
more realistic assessment of this constraint, by allowing not only the
transport to be different accross the simulations, but also various
aspects of the optimization procedure. It is hoped, that the outcome of
this exercise will provide a more realistic assessment of surface
fluxes
and their uncertainties which clearly will be extremely useful for the
wider carbon cycle research community.
Overview of the Simulations
The
participants are asked to perform the following inversions:
1) Reproduce T3 Level 2 results with own optimization software (
details).
2) as 1) but using personally preferred optimization procedure ( details).
3) as 2) but, in addition, using personally preferred boundary
conditions ( details ).
Inversion 3) is the
final goal of the exercise and should reflect the range of estimates
that individual groups consider the most realistic they can provide.
Inversion 1) is meant is a consistency check of the various codes that
are used for optimization. Inversion 2) is meant to quantify what part
of the differences in Inversion 3) are due to the different
regularization techniques that are used.
Each participant is
asked to submit simulated fluxes and corresponding uncertainties for
each inversion as well as a detailed specification of the applied
inversion set-up. Note that the range of estimates themselves will be
the main outcome. Given the large degrees of freedom in the experiment,
the interpretation of specific details in the results will likely be
difficult (which, however, is not the primary goal). The strength of
this experiment (relative to comparing just a bunch of publications) is
that the relevant differences between the set-ups are well documented,
and that the results refer to the same fluxes that can therefore
straightforwardly be compared (and be compared with other Transcom 3
levels).
Details of
Inversion 1
Repeat the
simulation as specified in the experimental outline of the interannual
inversion of Transcom 3 level 2 by David Baker (download
here, see also the example of questionnaire
1 - under
submissions - for a summary) for your
own
model. As the target period of the inversion we take 1988-2001
(excluding spin-up). The number of measurement stations is 76, which
are
to be used throughout the target period. The results that were computed
in Transcom 3 level 2 can be downloaded (
here
- NB: these posterior fluxes are WITHOUT fossil fuel).
These results were obtained without any SVD/Fourier truncation or other
regularization tricks except for a prior term in the cost function
(just
the plain Bayesian optimum). Try and get a best as possible match with
these results. You may want to try the basefunctions of your model as
they were reconstructed from the level 2 output (to be downloaded
here). In the meanwhile we found out that
a different region assignment of the presubtracted fields may explain
some of the differences. You can check this using the T3L2 posterior
fluxes without postadding any of the presubtraction fields (to be
downloaded
here) Send in the
most accurate match
that you obtained, and a documentation of any findings/hints that you
think may explain the remaining differences (for details of the
submission see
below ).
Note, this experiment assumes that
participants use the Bayesian concept for their inversion. If you use a
different method, it may be that to carry out this experiment would
require substantial reprogramming. In this case you may skip this part.
Details of
Inversion 2
Start with the same set-up as
described in Inversion 1). In particular the target time period and the
observations and their uncertainties should remain the same as in
Inversion 1). Change the optimization method to one that you consider
the best you can do. This includes changes in the number and the
geometry of the regions (in space and time); and you may want to use
different internal base func
tion structures and a priori covariances
(variances, however, should remain the same). Integrate/interpolate the
estimated fluxes and uncertainties over/to monthly Transcom 3 regions,
and document all changes as compared with Inversion 1 (see
submissions).
Details of
Inversion 3
Start with the same set-up as
described in Inversion 2). Change all boundary conditions to what you
consider the best you can do. By boundary conditions we mean every
aspect of the inversion, other than the target period. This includes
the
number of observations, selection method, uncertainties, and the use of
interannually varying meteorology etc. to drive your transport model.
However, the optimization method and region geometry should be the same
is used in Inversion 2). Integrate/interpolate the estimated
fluxes and
uncertainties over/to monthly Transcom 3 regions, and document all
changes as compared with Inversion 2) (see
submissions).
Submissions
For every simulation one results file and one
documentation file are submitted. The file name convention is:
model-id.inversion-id.format-id
where
model-id:
as used in previous Transcom experiments, or new (e.g. TM3.Houweling)
inversion-id: inversion that
the file contents refer to (inv1, inv2 or inv3)
format-id: 'txt' for
documentation, 'nc' for inversion results in netcdf format.
The .txt files contain a completed
questionnaire in ascii format (download examples here) which is to contain
a detailed and quantitative documentation of all changes that have been
introduced as compared with the previous inversion (such that someone
else could repeat the inversion on the basis of this documentation
only). Because of the various approaches employed by the different
participants this documentation is very important, and possibly we will
get back to you for even more information.
The .txt
file for inversion 1 is optional (just in case there are comments in
connection with efforts to make the two inversions match).
The .nc
file is a netcdf formatted file and contains the following results for
each inversion:
- prior fluxes
- prior flux uncertainties
- posterior fluxes
- posterior flux uncertainties
- presubtracted fluxes (experiment 1 only),
fossil fluxes (experiment 2&3)
- prior simulated concentrations
- posterior simulated concentrations
- used measurements
- χ2
Fluxes and 1 sigma
uncertainties should refer to monthly Transcom regions (see
map ). The unit is PgC/month with respect to the
atmosphere (positive means source). Note that fluxes and uncertainties
refer to the sum of ALL processes (INcluding e.g. fossil fuel use)
Please make use of our standardized fortran output subroutine to create
this file (
output_nc.f ).
Ftp all files to
the a
nonymous ftp-server ftp.bgc-jena.mpg.de; cd pub/incoming/; bin; put
*.*; bye.
Send a notification e-mail of your submission to
all three of us:
Sander Houweling (s.houweling@phys.uu.nl),
Martin Heimann (martin.heimann@bgc-jena.mpg.de),
Stefan Körner (stefan.koerner@bgc-jena.mpg.de).
Analysis Workshop
The Jena group has organized a
workshop
in spring of
2003 (12-15 May)
where the first results were presented and discussed. Information on
the workshop program, attendee list, pictures and links to the
powerpoint presentations can be found
here.
For the T3L3 intercomparison the following has been proposed:
- Sept 1: New version of
the results that were submitted by March 31 (distributed by e-mail)
- Oct 1: Extended deadline
for new submissions
- Dec 2003: New version of
results including Oct 1 submissions (meeting to be announced)
- June 2003: Update of T3L3
analyses at the next TRANSCOM meeting (..?)
Based on our experience so far the protocol has undergone some minor
changes (see the
correction list for
details).
Please send an e-mail to
s.houweling@phys.uu.nl
if you encounter any problems.
Last update: June 4, 2003 by Sander Houweling.
Corrections since the first release (9-12-2002) are listed
here