Z_INIT_HOMOGENIZE=True uses non-reproducing and potentially incorrect averages #296
Labels
Comments
Hallberg-NOAA
added
bug
Hallberg-NOAA
added a commit
to Hallberg-NOAA/MOM6
that referenced
this issue
Jan 7, 2023
Added the ability to use reproducing sums to create spatially homogenized tracer fields when Z_INIT_HOMOGENIZE = True, for rotational symmetry and consistency across layouts. The previous version had used non-reproducing sums. This new code is used when HOR_REGRID_ANSWER_DATE >= 20230101, and the comments describing HOR_REGRID_ANSWER_DATE have been updated to reflect this. As a part of this change, the new publicly visible routine homogenize_field was added to the MOM_horizontal_regridding module, so that the homogenization occurs in a single part of the code rather than being spread across several files. By default this commit could lead to answer changes in some cases, depending whether and how HOR_REGRID_ANSWER_DATE is set, but it turns out that the existing single_column test cases that use this have only 4 points and happen to give the same answers with either the older or newer version of the code. This commit addresses MOM6 issue mom-ocean#296 (github.com/NOAA-GFDL/issues/296), which can be closed as soon as this commit is merged in to the dev/gfdl branch of MOM6.
marshallward
pushed a commit
that referenced
this issue
Jan 28, 2023
Added the ability to use reproducing sums to create spatially homogenized tracer fields when Z_INIT_HOMOGENIZE = True, for rotational symmetry and consistency across layouts. The previous version had used non-reproducing sums. This new code is used when HOR_REGRID_ANSWER_DATE >= 20230101, and the comments describing HOR_REGRID_ANSWER_DATE have been updated to reflect this. As a part of this change, the new publicly visible routine homogenize_field was added to the MOM_horizontal_regridding module, so that the homogenization occurs in a single part of the code rather than being spread across several files. By default this commit could lead to answer changes in some cases, depending whether and how HOR_REGRID_ANSWER_DATE is set, but it turns out that the existing single_column test cases that use this have only 4 points and happen to give the same answers with either the older or newer version of the code. This commit addresses MOM6 issue #296 (github.com//issues/296), which can be closed as soon as this commit is merged in to the dev/gfdl branch of MOM6.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
The averaging to generate spatially homogeneous initial conditions that is triggered in the
horiz_inter_and_extrap_tracer()routines or inMOM_temp_salt_initialize_from_Z()by settingZ_INIT_HOMOGENIZE=Trueuses simple sums that are non-reproducing across rotation, PE-count or layout, and does not account for the possibility that some of the values come from land points. Ideally this would be modified to use properly masked and area-weighted reproducing sums, as is done inglobal_layer_mean(), but this would lead to answer changes at the level of roundoff.The one known experiment where
Z_INIT_HOMOGENIZE=Trueis used - MOM6-examples/ocean_only/single_column - is only ever run on a single PE and has just 4 columns that are (always?) ocean points, so this might not be too bad of an issue there, but it is easy to envision other cases where the current version of the code could lead to undesirable consequences.The text was updated successfully, but these errors were encountered: