area_statistics unable to load areacello FX files when diagnostic MIP doesn't match fx MIP

[ledm]

The UKESM1's ocean spatial grid does not change (and perhaps other models too). It will always be the same for all MIPs, scenarios, experiments and time ranges. With this in mind, there is only one copy of the grid area data, areacello. It can be found with:

          - {dataset: UKESM1-0-LL, project: CMIP6, mip: Ofx, exp: piControl, ensemble: r1i1p1f2, grid: gn, start_year: 1850, end_year: 1850,}

(As an aside, I don't understand why we need to provide a time range like start_year: 1850, end_year: 1850 for a time-independent variable!)

If we are looking at data that isn't in the exp: piControl experiment, _get_correct_fx_file is not finding it. For instance, the following recipe:

preprocessors:
  prep_timeseries_2d_ocean:
    area_statistics:
      operator: mean
      fx_files: [areacello, ]

diagnostics:
  time_series:
    variables:
      timeseries_fgco2:
        short_name: fgco2
        preprocessor: prep_timeseries_2d_ocean
        fx_files: [areacello, ]
        additional_datasets:
         - {dataset: UKESM1-0-LL, project: CMIP6, mip: Omon, exp: historical, ensemble: r1i1p1f2, start_year: 1995, end_year: 1999, grid: gn}

produces the error:

    fx_cube = iris.load_cube(fx_file)
  File "/home/users/ldemora/anaconda3_20190821/envs/esmvaltoolprivate/lib/python3.7/site-packages/iris/__init__.py", line 381, in load_cube
    raise iris.exceptions.ConstraintMismatchError('no cubes found')
iris.exceptions.ConstraintMismatchError: no cubes found

@valeriupredoi and I have been discussing this. There are a few suggestions on how to adapt to solve this issue in the short term:

• Make _get_correct_fx_file smart enough to look in picontrol if the areacella, areacello is not found.
• Allow a more complex fx_file argument to area_statistics:

        fx_files: [{short_name: areacello, mip:Ofx, exp: picontrol, ensemble: r1i1p1f2, }]

• I personally think that the cell area/volume data is crucial metadata that should be required by iris when loading the cube. However, we're not in a position to make this change.

I've also noticed that some recipes are using area_statistics, operator:mean without the fx_files: areacello/areacella argument. This may result in weird and unphysical results:

• recipe_cox18nature.yml
• recipe_ecs.yml
• recipe_flato13ipcc.yml
• recipe_ocean_bgc.yml
• recipe_ocean_example.yml
• recipe_ocean_Landschuetzer2016.yml
• recipe_tcr.yml
• recipe_wenzel16jclim.yml

All of these recipes are likely producing incorrect results. In the case of the UKESM data, the difference between the global SST average with and without area-weighting the mean is approximately 4 degrees. This is a non-trivial error!

Also note that in the past, area_statistics used the iris tool iris.analysis.cartography.area_weights to create area_weights, when no fx data were provided. I believe that we should not use this function anymore as it is not as accurate as the areacell datasets.

Finally, just to make things worse, these issues may also apply to volcello in volume_statistics and zonal_statistics preprocessors as well.

[valeriupredoi]

solved in #439

[bouweandela]

Fixed by #1609.