Is your feature request related to a problem? Please describe.

Currently when we train on data, we assume a constant land-sea mask for example.

This is tentatively true within a data set up then doesn't work for new data with different NaN locations.

Describe the solution you'd like

Somehow make an inference mode, which does not rely on constant masks.

Describe alternatives you've considered

No response

Additional context

No response

Organisation

ECMWF