Add SWE repartitioning for CLM #313
Conversation
If a variable has both a _FillValue attribute and a missing_value attribute, make sure the values are identical or we cannot support it with a single DART missing value. Handle _FillValue and missing_value separately because variables may have one or the other but not both. Adding a couple trivial unrelated changes not worthy of their own pull request, removing unused variables, making a hardcoded filename a parameter, that sort of thing.
The ultimate goal for models that support missing values is to read the variables and replace both _FillValue and missing_value items with the DART missing value. As long as r8 /= r4, there is no loss of precision when writing BACK to the original files. If DART is built with r8 = r4, and it reads from a NF90_DOUBLE and writes to a NF90_DOUBLE, there may be trouble ...
The game plan is to ultimately replace the DART missing value with whatever is originally in the file with the missing values in a post-processing 'dart_to_model' program.
Basically, this replaces the use of the progvar structure with the state_structure. There are some bits in the static_init_model that have been moved to subroutines to improve readability.
clm_to_dart.f90 is needed to reliably replace the CLM variables that have missing_value and/or _FillValue attributes but do not actuall use them. This happens in the restart files for the snow layers, for example.
This is a standalone program and replaces bogus values in layers of snow with the _FillValue value. It replaces the values in-place - i.e. it modifies the input file. To use it, you should COPY the clm restart file to the desired input filename. That filename should be the source of the variables used in pmo or filter or ...
Explores what happens to related snow variables when there is 'no snow' according to SNLSNO. T_SOISNO in the layer closest to the ground still has values even if there is no trace of snow (H2OSNO = 0.0). This seems like a problem with CLM logic. frac_sno seems like a reliable indicator to prevent replacing values in the layer thought to be 'empty'.
Using 'simple table' format for namelist which allows specification of newlines in table cells. Described in https://rest-sphinx-memo.readthedocs.io/en/latest/ReST.html#tables
CLM update 1
levsoi and levdcmp are new dimensions ... and carbon/nitrogen vars use them, so we need them.
The get_grid_vertval() routine still needs work to fully support the different vertical coordinates.
Replacing the 'kind' nomenclature when reporting QTY values.
The vector history files have variables declared with the column varying fastest, i.e. float SOIL1N_vr(time, levdcmp, column) ; while the other vector format file (the restart) always has vertical varying fastest, i.e. double H2OSOI_LIQ(column, levtot) ; That makes a difference when creating the 'model_size' metadata arrays.
Using the netcdf_utilities_mod routines makes it a lot clearer.
get_grid_vertval() still needs work, as does dart_to_clm.f90
The H2OSOI variable (QTY_SOIL_MOISTURE) is currently throwing a model_interpolate() error ... the COSMOS_NEUTRON_INTENSITY obs are exercising that FO. SOIL_MOISTURE is a particularly ugly interpolation - will deal with that later. SMINN_vr(time,levsoi,lat,lon) also uses the 'levsoi' coordinates.
Has the variables necessary for several of the observations in the obs_seq.in Avoiding soil moisture for now.
The exit from the case statement was in a funny place and probably not doing what it should have been doing. The code did the right thing, it just checked extra dimensions.
The dart_to_clm.f90 still needs support for the repartitioning of snow.
Merge remote-tracking branch 'upstream/clm_update_2' into clm_update_2
Using integer in repartition_swe >0 Co-authored-by: hkershaw-brown <[email protected]>
Using integer in repartition_swe >0 Co-authored-by: hkershaw-brown <[email protected]>
|
My latest batch of commits are associated with the observation converters mostly used with clm, that never got migrated to the the DART repository during the prior clm_update PRs. These commits include soil moisture and leaf area observations converters coming from NASA (NASA_Earthdata folder) and NSIDC (NSIDC folder). I updated the links to which acquire the raw netcdf or hdf5 data files, tested functionality, and created documentation for each, which was either sparse or non-existent. Changes to HDF5_utilities_mod.f90, obs_seq_utilities.f90, EASE_utilities_mod.f90 were needed to support the obs converters within the NASA_Earthdata and NSIDC folders. |
converters' page. Minor changes to formating to stop readthedocs complaining
Remove SMAP_L2_to_obs in NASA_Earthdata doc Co-authored-by: hkershaw-brown <[email protected]>
|
@hkershaw-brown I re-tested this current version of the branch for all 3 snow re-partitioning schemes and for the most part it looks pretty good, after I added the |
Let's check this out. I'm not sure if CLM is always bit-for-bit idenical, but DART should be. diff --git a/assimilation_code/modules/io/io_filenames_mod.f90 b/assimilation_code/modules/io/io_filenames_mod.f90
index a3a777a15..632cbb3d6 100644
--- a/assimilation_code/modules/io/io_filenames_mod.f90
+++ b/assimilation_code/modules/io/io_filenames_mod.f90
@@ -661,15 +661,15 @@ if ( get_has_FillValue( domid, varid) ) then
select case (get_xtype(domid, varid))
case (NF90_INT)
call get_FillValue(domid, varid, spvalINT)
- call check_attribute_value_int(ncFile, filename, ncVarID, 'FillValue', spvalINT)
+ call check_attribute_value_int(ncFile, filename, ncVarID, '_FillValue', spvalINT)
case (NF90_FLOAT)
call get_FillValue(domid, varid, spvalR4)
- call check_attribute_value_r4(ncFile, filename, ncVarID, 'FillValue', spvalR4)
+ call check_attribute_value_r4(ncFile, filename, ncVarID, '_FillValue', spvalR4)
case (NF90_DOUBLE)
call get_FillValue(domid, varid, spvalR8)
- call check_attribute_value_r8(ncFile, filename, ncVarID, 'FillValue', spvalR8)
+ call check_attribute_value_r8(ncFile, filename, ncVarID, '_FillValue', spvalR8)I'm wondering if this is the difference between your old and new runs. But it could be something else. Cheers, |
Sure -- I am currently running with the files within I have diffused the netcdf restart files for each companion case (the new and older versions) for variable |
|
Odds and ends update: The In reference to the reproducibility of
I am going to re-run the test |
|
The CLM5-DART Western US benchmark simulation with Out of curiosity, I performed some tests running CLM-DART assimilations, with CLM in 'branch' mode (bit by bit reproducible for the start of CLM runs) instead of 'hybrid' mode which is not bit by bit. This didn't explain some of the small differences I was seeing in reproducibility, but doesn't effect the readiness of this PR as far as I am concerned. |
Description:
This adds the capability to repartition snow layer mass and dimensional variables. The repartitioning assures that the adjustments applied to the snow layers are consistent (mass and dimensions are conserved) with the adjustment of the total snow water equivalent. This also resolves the weakness where if a snow layer does not exist for all ensemble members, it is not adjusted for all ensemble members. We allow the user to choose the SWE repartitioning approach within the
dart_to_clmnamelist withinSWE_repartition. A value of '0' provides the previous, default behavior of CLM-DART, a value of '1' repartitions snow to all active snow layers, and a value of '2' repartitions snow to the bottom snow layer only.This
SWE_repartition_fixbranch began fromclm_mainatcommit 337686f(8/27/21). Concurrent changes of theclm_mainbranch have occurred during this time by @hkershaw-brown. I made an attempt to merge theSWE_repartition_fixwith the currentclm_mainatcommit a741218(10/19). This led to expected behavior for all variables except for the clm variableT_SOISNO(snow/soil temperature) which was updated to unrealistic large values for certain snow layers. This led to mass balance errors in CLM. At the suggestion of Helen I reverted the changes toSWE_repartition_fixto the last working version before that attempted merge withclm_main. File changes directly related to adding new SWE repartitioning capability as part of this PR include:~/models/clm/dart_to_clm.f90~/models/clm/dart_to_clm.rst~/models/clm/readme.rst~/models/clm/dart_to_clm.nml~/models/clm/work/input.nml~/models/clm/shell_scripts/cesm2.2/assimilate.cshFixes issue
This directly addresses github issue #253 , updating snow in CLM.
Types of changes
Documentation changes needed?
Tests
Performed perfect data experiment of CLM5-DART across global domain with daily assimilation across 7 day period. See case
/glade/u/home/bmraczka/cases/cesm2.2.0/clm5_SWE1_fixas an example of setup including setup scripts ofCLM5_setup_assimilation.originalandinput.nml. Performed this test for all 3 SWE repartitioning options (clm5_SWE0_fix,clm5_SWE1_fix,clm5_SWE2_fix) that demonstrated: 1) the default implementation (no repartitioning) provides identical bitwise behavior to previous workingclm_mainversion, 2) implementation of repartitioning schemes were mathematically sound (mass/dimensions of snow layers conserved), and 3) Update of CLM restart file is compatible with all internal CLM checks.I intend to perform further scientific testing of repartitioning script across Western United States domain for Utah
NASA CMS collaboration -- mainly intended to explore impact of SWE repartition options upon soil moisture and surface albedo impacts.
Checklist for merging
Additional changes have to occur before this PR is ready to merge.
Testing Datasets
For an example set of CLM files for testing dart_to_clm functionality see:
/glade/scratch/bmraczka/cesm2.2.0/clm5_SWE_PR/rundart_to_clm_input_file = 'clm2_0001.r.2011-01-02-00000.nc.original'dart_to_clm_output_file = 'clm5_SWE_PR.clm2_0001.r.2011-01-02-00000.nc.original'repartition_vhist_file = 'clm5_SWE_PR.clm2_0001.h2.2011-01-02-00000.nc'repartition_analysis_file = 'clm_analysis_member_0001_d03.2011-01-02-00000.nc'