Is your feature request related to a problem? Please describe.
For diffusion spatial downscaling we need a way to interpolate low resolution input data to the target grid.
Describe the solution you'd like
Follow the suggestion in point 1 of #1274
That is,
1. Spatial preprocessing
Map (batch, time, ensemble, source points, features)
to (batch, time, ensemble, target points, features)
while preserving every non-spatial axis and supporting both source-grid and target-grid sharding.
Approach
Introduce a spatial preprocessor (using a cross-grid processor) registered on the model interface.
Spatial preprocessing is applied before all other preprocessors so that every dataset reaches the model on the same grid and existing normalisation strategies (particularly for tendency models) continue to work unchanged.
Describe alternatives you've considered
An interpolation connection as implemented here: #1242
Additional context
No response
Organisation
No response
Is your feature request related to a problem? Please describe.
For diffusion spatial downscaling we need a way to interpolate low resolution input data to the target grid.
Describe the solution you'd like
Follow the suggestion in point 1 of #1274
That is,
1. Spatial preprocessing
Map (batch, time, ensemble, source points, features)
to (batch, time, ensemble, target points, features)
while preserving every non-spatial axis and supporting both source-grid and target-grid sharding.
Approach
Introduce a spatial preprocessor (using a cross-grid processor) registered on the model interface.
Spatial preprocessing is applied before all other preprocessors so that every dataset reaches the model on the same grid and existing normalisation strategies (particularly for tendency models) continue to work unchanged.
Describe alternatives you've considered
An interpolation connection as implemented here: #1242
Additional context
No response
Organisation
No response