Align FieldArray with Python xarray and netCDF data model

[sjdaines]

The PALEOmodel.FieldArray is intended to work in the same way as the Python xarray, which itself uses the data model from netCDF CF conventions. Currently there are some key differences:

• Dimensions vs Coordinates. This is the major difference, these do not currently follow the xarray/CF data model. We do want to clearly separate 'dimension' and 'coordinate' in order to allow different coordinate systems (pressure vs height for example). However it's not clear that a single 'dimension coordinate (xarray) / coordinate variable (CF)' with the same name as a dimension is what we want, as that makes it more complicated to connect eg cell centre and face coordinates. See xarray discussion Terminology for the various coordinates pydata/xarray#1295. It might be clearer to attach a list of 'preferred' coordinates to named dimensions, instead of using the xarray/CF naming convention.

[sjdaines]

This will need changes in both PALEOboxes and PALEOmodel in order to update the abstractions used for grids, dimensions and coordinates, and more clearly define the data model that is used to store model output internally (PALEOmodel.OutputWriters.OutputMemory) and on disk (currently just a serialisation of the internal format using jld2)

Currently the internal (and hence serialized external) format for model output uses the NamedDimension and AbstractGrid implementation. So the first priority has to be to implement a netCDF data format for storing model output on disk, so that the internal implementation can be updated without affecting existing work-in-progress that relies on saved output. Then the data model for grids/dimensions/coordinates and the representation of model output can be updated.

One design choice that should then be reconsidered is how to represent grid information internally: via an abstract interface (as at present), vs an explicit lower-level data model (as netCDF does). Currently PALEOboxes grids implement an interface, and don't expose dimensions etc in a low level generic way. Instead there are three api calls:

• internal_size used to define dimensions for arrays in CellSpace
• create_default_cellrange used to create CellRange with appropriate indices
• get_region used to take slices and return appropriate data and coordinates in an xarray-like container for output analysis