GENIE:GENIE Tuning rel-2-7

=Tunings of release 2.7.n=

This page provides comprehensive details of the parameter estimation exercises employed for releases 2.7.n.

The tuning exercise strategy was discussed at a GENIE-LAMP meeting at UEA on 23/03/2010 and defines three phases:


 * 1) Four standard physical tuning of the 8 and 16-level model with and without the ENTS land scheme  . The "improved equation of state" is to be switched on by default for these tunings. To control meridional heat transport over the Southern Ocean an extra 13th free parameter will be introduced/exposed.
 * 2) * 8-level  without ENTS, 13 free parameters, 4 target fields
 * 3) * 8-level  with ENTS, 13 free parameters, 4 target fields
 * 4) * 16-level  without ENTS, 13 free parameters, 4 target fields
 * 5) * 16-level  with ENTS, 13 free parameters, 4 target fields
 * 6) Two tunings of the 16-level model without ENTS land
 * 7) * Mixed layer switched on (to invetigate effects of this improved scheme), (13+3 free parameters)
 * 8) * Equation of state switched off (to compare with legacy model)
 * 9) Two tunings of the biogeochemistry parameters for the   model holding the tuned physical parameters constant with the mixed layer scheme switch on and off.
 * 10) * 16-level  without ENTS, fixed physics, Mixed Layer ON, 8 free parameters, 2 target fields
 * 11) * 16-level  without ENTS, fixed physics, Mixed Layer OFF, 8 free parameters, 2 target fields

Summary
The source code repository has been tagged at release  to indicate a stable release of the code base following the completion of major developments in a number of the modules (see release page for details). The branch  has been created at the same point in the repository in order to manage the tuning process for this version of the model. It is intended that the branch will be merged back into the trunk when tuning is complete in order to propagate the tuned configurations, and potentially adjustments that were required to perform the tuning, to the default version.

Following revision of the tuning exercise a new branch will be created.

Base configuration
The initial physical tuning of the model is applied to the 36x36x16 equal-area-grid resolution of the model and is based upon the configuration file.

The source code is exported from http://source.ggy.bris.ac.uk/subversion/genie/branches/rel-2-7-tuning.

The model is built in the usual way using genie_example.job:

./genie_example.job -f configs/eb_go_gs_itfclsd_16l_tuning.xml

An archive of the model binary and supporting data files and libraries is generated:

zip rel-2-7-tuning.zip $(cat ArchiveContents.txt)

Tuning specification
The config file  is a copy of   and provides the specific changes required for the tuning exercise. In particular the duration of each simulation is set to 5000 years to ensure approximate equilibration of some degree in the model for each parameter set that is evaluated. The frequency of output of the eb_go_gs modules is adjusted in order to facilitate restarts of the model every 500 simulated years:

Specific configuration details for the ac_bg muldules. The duration of the BIOGEM simulation is set to the length of the individual restarted simulation time slices (500 years):

Tuning configuration
The parameter estimation problem is:

Find:

Boundary conditions
For the record the boundary conditions are (ie. these files are read by the model during initialisation):

Target data
Detail the target data files and the code that is used to compare the model output with the data to produce the RMS error scores.

The root mean square error function code is found in. This code compares model output with observational data across the four fields ocean temperature, ocean salinity, atmospheric temperature and atmospheric humidity. Four root mean square error values are returned, one for each field, plus a composite error function value. The multi-objective optimisation algorithm works with the four individual field errors and works to generate a pareto-optimal set of solutions to the problem.

The model configuration includes entries that define which data files are to be used to make the model-data comparison:

Validation
An independent expert will provide details of the error score for the vanilla build of  and ideally will provide NetCDF output from their run. The tuning software must be able to reproduce the vanilla results for the default parameter set using the infrastructure supporting the tuning exercise (ie. the DAGMan execution of the Win32 model build on the UoS Condor Pool must reproduce known results before the tuning exercise can be performed).