Suitability of xgcm for interpolation of point cloud data on sub-grid scale

[VeckoTheGecko]

Hello! I am a core developer for Parcels, a Lagrangian simulation toolbox. Our core functionality is releasing particles to be advected on GCM output according to user defined behaviour.

We're looking to migrate our internals to xarray (issue).

I was very excited to find xgcm for its ability to interpret Arakawa grids, and automatically parse metadata from NetCDF datasets. Initially the presence of Grid.interp() gave me hope that we would be able to heavily rely on this package, however looking deeper at interp() I since found that its specialised for interpolating dataarrays onto the grid and not for arbitrary interpolation using the context of the grid.

Also looking at the public API, I don't see how I could plug into xgcm to extract the grid context for use in custom interpolation routines. Any suggestions? Or is xgcm very specialised to the world of finite volume analysis to the point we're better off writing our own code?

cc @erikvansebille

[dcherian]

What kind of interpolation do you want? From gridded field to trajectory?

[VeckoTheGecko]

Yes, from the gridded field onto points (which would be used to compute the trajectory)

[dcherian]

Xarray's interp implements this "vectorized interpolation", but not always the most performant. I am planning on fixing that for method="linear" and method="interp" though. What interpolation methods are you interested in?

[VeckoTheGecko]

We're mainly interested in linear and nearest neighbour interpolation... also on arbitrary curvilinear grids (which looks to be an open issue pydata/xarray#2281 (comment)).

But in the case of a rectilinear C grid, I guess it would be a matter of:

• Ensuring the dataset dimension naming conforms to some standard (e.g., Comodo like here)
• doing ds["V"].interp(XC = ..., YG = ...)

and bypassing xgcm entirely. If the grid file is separate, then that is a separate can of worms if going via xarray.

I was just hoping that we could piggyback a bit more off of xgcm, but I don't know if thats possible.