Automatically select a lazy regridding scheme in `regrid` preprocessor function for 2D lat/2D lon grids when possible

[bouweandela]

Through iris-esmf-regrid and the generic regridding schemes, it is possible to regrid grids with 2D latitude and 2D longitude coordinates lazily. However, these regridding schemes are not automatically selected and instead the non-lazy ESMPyLinear/ESMPyNearest/ESMPyAreaWeighted built-in schemes are selected. This results in poor computational performance by default.

@schlunma @sloosvel Do you think it would be possible to improve the default so we prefer lazy schemes and automatically select those when possible?

[sloosvel]

I'm not sure this is possible because the generic regridding schemes don't work for all models, you can get results such as this one using ESMFAreaWeighted:

[bouweandela]

Could you provide some more details on that image, to help me understand the issue?

• What was the source grid or model?
• What was the target grid?
• Which regridding scheme was used?
• Does the image look as expected?

[sloosvel]

It happens to models that have OceanGridFix applied to them. The generated bounds work fine to regrid with the non-lazy regridders, but for some reason when you call the lazy AreaWeighted scheme, this happens. You can check with the recipe below for instance:

documentation:
  title: test regrid
  description: |
    Test regrid
  authors:
     - loosveldt-tomas_saskia

datasets:
   - {dataset: FGOALS-g3, ensemble: r1i1p1f1, exp: historical, project: CMIP6, grid: gn, timerange: 1951/1951}

preprocessors:
  pp:
    regrid:
      target_grid: 1.5x1.5
      scheme: 
        reference: esmf_regrid.schemes:ESMFAreaWeighted

diagnostics:
  test_regrid:
    variables:
      tos:
        mip: Omon
        preprocessor: pp
    scripts: null

[bouweandela]

An overview of regridding a 2D ocean variable (sea surface temperature) for most models in CMIP6 to a 1x1 grid is available here and with xesmf here. Created with the debug-regrid branch which keeps going when a regridding error occurs and adds a bit of info.

A first inspection reveals the following issues:

• iris-esmf-regrid lazy regridders:
  • Linear:
  • 1D grids not supported: AWI-CM-1-1-MR
  • discontiguous grids: 2D variables supported with use_src_mask: true, 3D (i.e. with vertical dimension) variables not supported: CMCC-CM2-SR5, CMCC-ESM2, CNRM-CM6-1-HR, CNRM-CM6-1, CNRM-ESM2-1, CanESM5-1, CanESM5-CanOE, CanESM5, EC-Earth3-Veg-LR, EC-Earth3-Veg, EC-Earth3, GFDL-ESM4, IPSL-CM6A-LR, IPSL-CM6A-LR, MPI-ESM1-2-HR, MPI-ESM1-2-LR, MRI-ESM2-0, NESM3, TaiESM1
  • masked vertical stripe in CESM2-WACCM, CESM2, CIESM, FGOALS-f3-L, FGOALS-g3, FIO-ESM-2-0
  • Nearest Neighbour:
  • 1D grids not supported
  • variables on discontiguous grids not supported, use_src_mask: true gives wrong results, see https://github.com/SciTools-incubator/iris-esmf-regrid/issues/368
  • Area Weighted:
  • 1D grids not supported: AWI-CM-1-1-MR
  • 3D variables discontiguous grids not supported
  • results look wrong for BCC-CSM2-MR , FGOALS-f3-L, FGOALS-g3, FIO-ESM-2-0, KACE-1-0-G (deformed continents), and INM-CM4-8 (a large non-existent sea appears in Asia)
• Built-in (mostly non-lazy ESMPy based) regridders from esmvalcore.preprocessor.regrid_schemes:
  • Linear:
  • (non esmpy based regridder) 1D grids produce results that look wrong: AWI-CM-1-1-MR
  • suspicous values vertical stripe in NESM3
  • masked vertical stripe in FGOALS-f3-L, FGOALS-g3, FIO-ESM-2-0, MRI-ESM2-0, TaiESM1
  • Nearest Neighbour:
  • (non esmpy based regridder) 1D grids produce results that look wrong: AWI-CM-1-1-MR
  • suspicious values vertical stripe in NESM3
  • masked vertical stripe in CMCC-CM2-SR5, CMCC-ESM2, CNRM-CM6-1, CNRM-ESM2-1, IPSL-CM6A-LR
  • Area Weighted:
  • not supported: AWI-CM-1-1-MR
  • results look wrong for BCC-CSM2-MR, FGOALS-f3-L, FGOALS-g3, FIO-ESM-2-0
  • vertical band of high values in CNRM-CM6-1-HR, EC-Earth3-Veg-LR, EC-Earth3-Veg, EC-Earth3, GFDL-ESM4, IITM-ESM, MPI-ESM1-2-HR, MPI-ESM1-2-LR (band is slightly tilted), MRI-ESM2-0, NESM3
  • masked vertical stripe in CanESM5-1, CanESM5-CanOE, CanESM5, GFDL-ESM4, MRI-ESM2-0, TaiESM1

[bouweandela]

It happens to models that have OceanGridFix applied to them.

@sloosvel I tried the recipe, but am getting poor results with all available regridders. Did I miss something? I ran this recipe and got the following images: Default built-in ESMPy based regridder iris-esmf-regrid regridder xesmf regridder from #2433

[sloosvel]

You are right, I don't know why I was under the impression it worked for the non-lazy schemes.