Postprocessing
After finishing, the inversion run automatically creates a postprocessing script "*_integrate.script"
(where the * stands either for the number of iterations reached,
or for APO [= a-posteriori] if the iteration was stopped by threshold criteria,
or for PRI to process the a-priori fluxes).
This script can be used to
- create
NetCDF output files
of the a-posteriori flux field integrated over a set of regions and/or filtered/subsampled in time;
- do forward runs of the atmospheric transport model from the a-posteriori fluxes;
- calculate a-priori uncertainties;
- calculate a-posteriori uncertainties (if RHS has been run before);
- calculate further diagnostics.
To do so, edit the script to modify the postprocessing parameters as described below,
and submit it (on Mistral using sbatch SCRIPT).
This can be repeated (later) with other postprocessing parameters
(but avoid submitting the script several times simultaneously).
Integrated/filtered fluxes
- Choose the a-posteriori case:
- Set μ value(s) (prior strength) in parameter pp_mu
(a blank-separated list allows using multiple μ values sequentially in a single job).
- Set interation in iter (normally set already).
- If you want to write the difference from the prior instead of the full posterior fields,
set subt_pri = T.
- Choose the
set of integration regions
by pp_regionset (comma-separated list for multiple output).
- Choose the
temporal filter
and/or the
temporal subsampling
by pp_timemask (comma-separated list for multiple output).
- Further parameters: See the comments in "*_integrate.script".
Forward runs
- Choose the a-posteriori case as in ``Integrated/filtered fluxes'' above.
- Set pp_do_fwd = T to enable the forward run.
- Output:
- Atmospheric mole fractions will be written out for the locations/instants of the observations
from a given set of stations.
The list of stations and the data version to be used are specified by the parameters
stnlist and dataversion, respectively.
In order to use the same set of observations that constrained the inversion,
comment out the stnlist and dataversion settings.
The output is written into the new directory
"*_output1.dataversion.stnlist/".
For each station, there is an ASCII file with the
same format as the input data, with the modelled values added.
- As a complement to the atmospheric mole fraction output,
the reduced chi squared diagnostic χr2
is written into the file
"*_output1.dataversion.stnlist.chi2r.d".
It is calculated with an effective number of observations (n_eff)
taking into account that the model-data mismatch uncertainties
have been inflated by the data density weighting (see Tech.Rep.).
- If ninst3D = 21600 is set (=uncommented), additionally
full 3-dimensional atmospheric mole fraction fields
are written every 21600 seconds into yearly NetCDF files.
- If ninststn = 21600 is set (=uncommented), additionally
regular time series of atmospheric mole fractions
at a set of fixed locations are written.
The coordinates of the locations need to be provided in a local ASCII file
(one line per location, with name, latitude, longitude, height).
The name of this coordinates file is specified to the script by name_stationlist.
- If ninstsfc = 21600 is set (=uncommented), additionally
the mole fractions of the lowest TM3 model layer are written every 21600 seconds into a NetCDF file.
The file additionally contains Surface squared wind speed
and Surface pressure. It is used to obtain the atmospheric boundary conditions for
sea-air flux calculations.
(In all the ninst3D, ninststn, and ninstsfc parameters,
the output interval of 21600 seconds can be changed to any other interval compatible with the model time step;
alternatively, ninst* = -2 means output every month.)
- Further parameters: See comments in "*_integrate.script".
A-priori (and a-posteriori) uncertainties
- Activate by setting pp_out_cross = T or pp_out_cross_rhs = T.
- Importantly, the a-posteriori (co)variances will be correct only if
sufficient iterations along appropriate directions in the parameter space have been performed.
Convergence for fitting the measurements are generally not sufficient
for this purpose.
See the RHS mode on how to perform the additional iterations.
Further diagnostics
- See the further parameters and their comments in "*_integrate.script".