GENIE:GENIE Tuning rel-2-7

=Strategy for 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 (  and   model setups) tunings 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) * Thermobaricity term in 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

This strategy was restructured, updated, and refined subsequently and further discussed during meetings at the University of Southampton (19/05/2010) and the University of Bristol (26/05/2010) and elsewhere.

Three different combinations of active modules define three phases, each consisting of a default tuning and exercises enabling individual deviations from the standard setup. In contrast to previous tunings, all default setups include the following model updates (96 instead of 100 oceanic time steps/yr, meridional heat transport reduction over the Antarctica/Southern Ocean, wind stress forcing climatology from reanalysis data instead of observation-based data, thermobaricity term in equation of state), of which the second modification will included via a tunable parameter and the effect of the latter two modifications will be tested in individual experiments (Tunings 1c and 1e):


 * 1) Phase 1 (Ocean/Atmosphere/Sea ice without ENTS):
 * 2) * Tuning 1a: 16-level  without ENTS, 13 free parameters, 4 target fields
 * 3) * Tuning 1b: 8-level
 * 4) * Tuning 1c: Legacy wind stress forcing fields
 * 5) * Tuning 1d: Mixed layer dynamics switched on (to investigate effects of the mixed layer scheme), (13+3 free parameters)
 * 6) * Tuning 1e: Thermobaricity term in equation of state switched off
 * 7) Phase 2 (Ocean/Atmosphere/Sea ice and ENTS):
 * 8) * Tuning 2a: 16-level  with ENTS, 13 free parameters, 4 target fields
 * 9) * Tuning 2b: 8-level   with ENTS, 13 free parameters, 4 target fields
 * 10) Phase 3:(Ocean/Atmosphere/Sea ice and ocean biogeochemistry)
 * 11) * Tuning 3a: 16-level  without ENTS, fixed parameters for , Mixed Layer OFF, 8 free parameters, 2 target fields, base configuration of the   part to be decided
 * 12) * Tuning 3b, 3c, ...: Testing sensitivities of further configurations (e.g., MLD on, different production schemes in BIOGEM, etc.)

Initial physical tuning of version 2.7.0
Abandoned and superseded by tuning of version 2.7.1 (May 2010) Modifications in branch  have been incorporated into the trunk version before creating the new branch for the tuning of version 2.7.1. The description of the preparation of the (not started) tunings for version 2.7.0 has been move further below and modified/integrated in the description of the tunings of version 2.7.1.

Summary
The source code repository has been tagged at release  to indicate a stable release of the code base following the completion of developments in a number of the modules (see release page for details). The branch  has been created immediately following the tag   in order to manage the tuning process for this version of the model. It is intended that the branch will contain a tag  after (partial) completion of the tuning exercises and that it will then be merged back into the trunk in order to propagate the tuned configurations, and potentially adjustments that were required to perform the tuning, to the default version. In case of a new release during the tuning exercise and corresponding revision of the tuning exercise a new branch is planned to be created   and the tunings to be resumed based on the new branch.

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  (with parameters resulting from a previous tuning) updated with a few modifications. This setup corresponds to a model version which has been described in the supplementary material of a publication (http://www.biogeosciences.net/6/375/2009/bg6-375-2009-supplement.pdf) and been used previously in various studies (see e.g., http://www.biogeosciences.net/6/375/2009/bg6-375-2009-supplement.pdf for references). In addition a few further modifications (this includes the reduction of the number of ocean timesteps per year from 100 to 96 and the new wind stress forcing fields) were made. The resulting configuration files for the initial 500 year spin-up simulation and 500 year continuation  are provided in the directory   (nm stands for the respective tuning setup).

The source code is exported from http://source.ggy.bris.ac.uk/subversion/genie/branches/rel-2-7-1-tuning.

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

./genie_example.job -f configs/rel-2-7-1-tuning/rel-2-7-1-nm-S00.xml

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

zip rel-2-7-1-tuning-nm.zip $(cat ArchiveContents-rel-2-7-1-nm.txt)

Tuning specification
The config files  provide the specific changes with respect to the default configuration (specified in  ) required for the individual tuning setups. In particular the duration of each simulation is set to 10x500 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 enable restarts of the model every 500 simulated years. The following parameters are the non-varied parameters for the base configuration  (this includes a few default settings):

Tuning configuration
The parameter estimation problem is finding values for the parameters:

Build macro definitions
{| border="1" ! Parameter (XPath) ! Value ! Description
 * Optimiser controlled variable 01 - scales wind stress
 * Optimiser controlled variable 01 - scales wind stress
 * Optimiser controlled variable 01 - scales wind stress
 * Optimiser controlled variable 01 - scales wind stress

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).