AreaWeights filling

[JPXKQX]

What happened?

The current behaviour to compute the area weights of a set of nodes is explained below step by step:

• The globe is split into polygons, where each polygon should correspond to one and only one node.
• Each node is linked to a polygon, also referred to as a region. These regions are defined with a list[int] where each index corresponds to a node/vertex.
• The issue [bug] Buffer dtype mismatch, expected 'npy_intp' but got 'double' #45 was happening when a node was associated with no region, an empty list ([]), whose dtype was np.int64, and it was raising this "dtype mismatch" error.
• The current implementation of AreaWeights solves this issue by filling these values with 0 (or any variable defined in the fill_value attribute). This default behaviour can be dangerous as it seems some cases there may be many of these nodes, and their contribution to the loss function would be 0.

Some options to address/alleviate this problem are:

• Log a warning when the filling is done.
• Support filling with the min or mean.
• Set "np.nan" as the default value and handle it later (in anemoi-training?, in a post-processor?, ...)

Version

develop

Platform (OS and architecture)

x86_64 GNU/Linux