GENIE:GENIE Tuning rel-2-7

=Strategy for tunings of release 2.7.n=

This page provides an overview of the parameter extimation exercises employed for GENIE releases 2.7.n. It also provides comprehensive details of the parameter estimation exercises for phase 1 of the exercises. Details for the subsequent phases 2 and 3 can be found on GENIE:GENIE_Tuning_rel-2-7_phase_2 and GENIE:GENIE_Tuning_rel-2-7_phase_3, respectively.

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), University of Bristol (26/05/2010), University of Bristol (23/09/2010-24/09/2010) 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 climatology from reanalysis data instead of observation-based data and replaced wind speed forcing climatology of advective transport in the atmosphere, thermobaricity term in equation of state), of which the second modification adds a tunable parameter to the set of tuned parameters. The effect of the latter two modifications will be tested in individual experiments (Tunings 1c and 1e):


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

An alternative plan for phase 3, which was brought up during a meeting in Southampton (15/12/2010), is to make use of phase 2 to establish a new default configuration which includes ENTS, on which the setup of phase 3 will be based:
 * Phase 3 (Ocean/Atmosphere/Sea ice/Land and ocean biogeochemistry)
 * Tuning 3a: 16-level  with ENTS, fixed parameters for , Mixed Layer off, number of free parameters to be decided, target fields to be decided, base configuration of the   part to be decided
 * Tuning 3b, 3c, ...: Testing sensitivities of further configurations (e.g., MLD on, different production schemes in BIOGEM, etc.)

An additional experiment (Tuning 1f) is added to the list of experiments for phase 1 (13/01/2010), partly motivated by a discussion during a meeting in Southampton (15/12/2010) about large model-data differences of salinity in the vicinity of the Indonesian Archipelago, which are observed in results from other tuning experiments of phase 1. The topography used for tunings 1a, 1c, 1d, and 1e has been modified in order to enable Indonesian Throughflow and barotropic flow through the Bering Strait. Additionally, this modification also adjusts the extent of the Atlantic basin, which includes part of the Mediterranean Sea when using the default topography. The modified topography is planned to be used as the default topography for phase 2 of the tuning exercises.

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 moved further below and modified/integrated in the description of the tunings of version 2.7.1.

Phase 1, tuning of base version of (model version 2.7.1)
Abandoned and is superseded by tuning of version 2.7.4 (October 2010). The branch  have been abandoned and is superseded by. The original description of the planned tuning exercises for version 2.7.1 (which have been set up for the  release, but not yet started) has been updated and modified to reflect the use of release.

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 (unless they have been made to the trunk version in the meantime) that were required to perform the tuning (see below for a list of such modifications), 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.

Model setup and compilation
and subsequently configured for the local environment (see GENIE_running_an_expt).
 * The source code is exported from  into the directory


 * Initially, a set of external datasets have to be downloaded:


 * The model can be built in the usual way using  (the option   triggers the generation of an executable which can be used for offline calculation of the RMS error after the model has been run):


 * After building the model, an archive of the model binary and supporting data files and libraries can be generated:

.

Modifications, corrections, and updates made in branch 'rel-2-7-4-tuning'

 * 2010-10-28: Corrected incorrect attributes ("logical" instead of "boolean") of  tags in   in order to enable the tuning software to correctly handle the corresponding parameters.


 * 2010-01-03: Added functionality for the off-line computation of a range of diagnostic values from previously produced model output to the trunk version (revision 6044) and merged this revision into  branch. With this revision in place, the option   of   triggers the generation of an executable called , which can be used to compute diagnostic values from model output generated by the corresponding model executable.

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 the publication by Cao et al. (2009) (http://www.biogeosciences.net/6/375/2009/bg6-375-2009-supplement.pdf) and has 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 replaced wind stress and wind speed forcing fields) were made. The resulting configuration files for a 5000 year spin-up simulation are provided in the directory   (nm stands for the respective tuning setup).

Base parameters
The config files  provide 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 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 enable restarts of the model every 500 simulated years. The parameters below are non-varied parameters for the base configuration  (this includes some default settings):

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

Wind-related boundary conditions
The wind-speed and wind-stress fields used by the  module are long-term annual averages for the years 1979-2009 of zonal and meridional components of wind fields at 1000mb, and zonal and meridional components of surface-to-atmosphere momentum fluxes, respectively, from the NCEP/DOE 2 Reanalysis [Kanamitsu et al., 2002], using monthly fields derived from the NCEP/DOE 2 Renalysis data provided by the NOAA/OAR/ESRL PSD, Boulder, Colorado, USA, from their Web site at http://www.esrl.noaa.gov/psd/.

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

For the evaluation of the individual simulations, annually-averaged model output from the final year of the simulations is compared 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 which specify the filename of the datasets, variable names and scaling and offset values which are used to read in observationally-based or reanalysis-based fields which are subsequently interpolated onto the model grid and used for the model-data comparison:

Parameters for the climatology of in-situ ocean temperature
The in-situ ocean temperature field used in the model-data comparison is the objectively-analysed annual-mean climatology of ocean temperature [Locarnini et al., 2006] of the World Ocean Atlas 2005 (WOA05; http://www.nodc.noaa.gov/OC5/WOA05/pr_woa05.html) from the National Oceanographic Data Center (NODC; http://www.nodc.noaa.gov/) and the parameters listed below configure its use in the computation of the corresponding RMS error score.

Parameters for the climatology of ocean salinity
The ocean salinity field used in the model-data comparison is the objectively-analysed annual-mean climatology of ocean salinity [Antonov et al., 2006] of the World Ocean Atlas 2005 (WOA05; http://www.nodc.noaa.gov/OC5/WOA05/pr_woa05.html) from the National Oceanographic Data Center (NODC; http://www.nodc.noaa.gov/) and the parameters listed below configure its use in the computation of the corresponding RMS error score.

Parameters for the climatology of atmospheric temperature
The atmospheric temperature field used in the model-data comparison is a long-term annual average for the years 1979-2009 of the 1000mb temperature fields from the NCEP/DOE 2 Reanalysis [Kanamitsu et al., 2002], using monthly fields derived from the NCEP/DOE 2 Ranalysis data provided by the NOAA/OAR/ESRL PSD, Boulder, Colorado, USA, from their Web site at http://www.esrl.noaa.gov/psd/. The parameters listed below configure its use in the computation of the corresponding RMS error score.

Parameters for the climatology of atmospheric relative humidity
The atmospheric relative humidity field used in the model-data comparison is a long-term annual average for the years 1979-2009 of the 1000mb relative humidity fields from the NCEP/DOE 2 Reanalysis [Kanamitsu et al., 2002], using monthly fields derived from the NCEP/DOE 2 Ranalysis data provided by the NOAA/OAR/ESRL PSD, Boulder, Colorado, USA, from their Web site at http://www.esrl.noaa.gov/psd/. The parameters listed below configure its use in the computation of the corresponding RMS error score.

Validation
The model output files,  , and   from two 5-kyr model runs using the same source code and model configuration  but which were independently carried out (same setups as for model runs A and B below) have been found to agree (2010-11-04).


 * Model run A was carried out at the UoS on the Iridis 3 compute cluster within the setup used for the model optimisation where the 13 varied parameters have been set to the middle of their respective ranges. The model executable was compiled using gfortran version 4.4. The RMS model-data errors for this model run are:


 * Model run B was carried out at the UEA on a desktop workstation using the configuration . The model executable was compiled using gfortran version 4.3.



A comparison of model output generated at an earlier stage of the set up of the tuning exercise (using the code from branch ) revealed a difference between results generated by model executables compiled with Fortran compilers from differing families. TODO demonstrate such differences with the  model version.

Phase 1, tunings of modified setups of (model version 2.7.1)
Abandoned and is superseded by tuning of version 2.7.4. The branch rel-2-7-1-tuning has been abandoned and is superseded by rel-2-7-4. The original description of the planned tuning exercises for version 2.7.1 (which have been set up for the rel-2-7-1 release, but not yet started) has been updated and modified to reflect the use of release rel-2.7.4.

Phase 1, tunings of modified setups of (model version 2.7.4)
Some modifications of the base configuration are required to perform tunings of the variants, for   3 additional parameters need to be varied. Apart from these modifications, the setup, as well as varied parameters and ranges, remain unchanged with respect to the base variant :

Changes to selected model components
The model component  is not selected for this experiment.

Changes to boundary conditions
The wind-speed and wind-stress forcing fields are from the interpolated input fields commonly used in previous experiments (e.g., by Edwards and Marsh, 2005). See Edwards and Marsh [2005] for a description and the sources of these fields.

Changes to the list of required input files
Note, the required input files listed above have become available in the  branch with revision 6057.

Phase 2, tuning of
Please see GENIE:GENIE_Tuning_rel-2-7_phase_2 for details.

Phase 3, tuning of
Please see GENIE:GENIE_Tuning_rel-2-7_phase_3 for details.

Refereces

 * Antonov J. I., R. A. Locarnini, T. P. Boyer, A. V. Mishonov, and H. E. Garcia (2006), World Ocean Atlas 2005, Volume 2: Salinity. S. Levitus, Ed. NOAA Atlas NESDIS 62, U.S. Government Printing Office, Washington, D.C., 182 pp.
 * Cao, L., Eby, M., A. Ridgwell, K. Caldeira, D. Archer, A. Ishida, F. Joos, K. Matsumoto, U. Mikolajewicz, A. Mouchet, J. Orr, G.-K. Plattner, R. Schlitzer, K. Tokos, I. Totterdell, T. Tschumi, Y. Yamanaka, and A. Yool (2009), The role of ocean transport in the uptake of anthropogenic CO2. Biogeosciences, 6, 375-390.
 * Edwards N. R., and R. Marsh (2005), Uncertainties due to transport-parameter sensitivity in an efficient 3-D ocean-climate model. Clim. Dyn., 24, 415-433.
 * Kanamitsu M., W. Ebisuzaki, J. Woollen, S-K Yang, J.J. Hnilo, M. Fiorino, and G. L. Potter (2002), NCEP-DOE AMIP-II Reanalysis (R-2). Bul. Atmos. Met. Soc., 1631-1643.
 * Locarnini R. A., A. V. Mishonov, J. I. Antonov, T. P. Boyer, and H. E. Garcia (2006), World Ocean Atlas 2005, Volume 1: Temperature. S. Levitus, Ed. NOAA Atlas NESDIS 61, U.S. Government Printing Office, Washington, D.C., 182 pp.