Feature proposal: Regrid nodal data

[vieirasaulo]

Hi there,

Regridding nodal data would be a an important feature to load and wrangle finite element models. @Huite and I had a discussion about a strategy for regridding nodal data here. Perhaps a good interface for this new feature, would be:

regridded = xugrid.regrid(
     source=old,
     target=template,
     dim_methods={"mesh2d_node": BarycentricInterpolator, "grid2d_face": NearestRegridder},
)

Cheers,

Saulo

[Huite]

Hi @vieirasaulo

Based on our discussion, I've made a WIP PR here:
#369

It isn't quite finished yet, but the following Regridders work for nodal data (to nodal data, but also to face):

• BarycentricInterpolator
• InverseDistanceInterpolator
• NearestRegridder

The last two mostly rely on the scipy KD Tree.

There's on odd duck, which is that the overlap regridder can also convert nodal data to face data including the different aggregation methods (mean, min, max, sum, etc.). I think this is fairly intuitive: a number of nodes may fall into a face. The other way around obviously doesn't work out.

It also works for structured data (provided as an ordinary DataArray), they are automatically interpreted as face associated data (with the centroid being the representative vertex for interpolation and the like).

The PR basically works, but I changed a bunch of internal as well which requires updating quite a lot of tests (in particular the regridder .to_dataset() needs some additional logic ... not that anyone uses it as far as I'm aware though).
Secondly, a sigificant number of new tests is required.

Usage looks as follows:

# Using UgridDataArrays
regridder = xugrid.BarycentricInterpolator(source=my_nodal_uda, target=my_nodal_uda)
result = regridder.regrid(my_nodal_uda)

# Node to face, same idea
regridder = xugrid.BarycentricInterpolator(source=my_nodal_uda, target=my_face_uda)
result = regridder.regrid(my_nodal_uda)

# With a Xugrid Ugrid2d object as target
regridder = xugrid.BarycentricInterpolator(source=my_nodal_uda, target=target_grid, target_dim=target_grid.node_dimension)

# OR
regridder = xugrid.BarycentricInterpolator(source=my_nodal_uda, target=target_grid, target_dim=target_grid.face_dimension)

This API is consistent across the different classes, although the number of keyword arguments may differ. Especially inverse distance requires more arguments; I've taken a look at the GDAL inverse distance functionality, I'm curious to hear if you're missing options.

I've also taken a look at supporting regridding of entire datasets. This seems to work, thanks to the generality of xarray's apply_ufunc, but my logic isn't complete yet in terms of erroring on ambiguous arguments (e.g. multiple topologies, both node and face associated data, etc.).

You might be interested in doing a checkout of the PR and trying it on your data.

My feeling is that class based approach is more flexible, since it allows re-use (and even storage) of the regridding weights. The API is also clear in terms of the arguments, which differ per method. With a regrid function or method, we need to forward the kwargs (which is obviously quite manageable) and the weights are single-use only.
Having said that, there are many cases where a function is convenient, so I'm probably easily convinced to add it!