Enable downscaling for Atmorep

[grassesi]

In this issue-branch, the tail network for precipitation downscaling will be developed. The IMERG dataset, reprojected onto a 0.083°-grid, will serve as the target data for this downscaling task. While downscaling has been realized for 2m temperature (target data from COSMO REA6) in previous AtmoRep-versions (cf. lessig-dev-branch and arXiv paper), the envisioned approach will tackle the following issues of the previous approach:

• deep network for downscaling incl. continued training of full core model
• no injected auxiliary data (e.g. high-resolved topography)
• fixed target domain

The envisioned network architecture for downscaling has the following key properties:

• use PerceiverIO-module (Jaegle et al., 2022) to ...
  • reduce dimensionality of input data
  • fuse different kind of input data (decoder output from AtmoRep + auxliary information at higher resolution and/or same resolution)
• embedding (incl. positional embedding) for injected data re-using the prepare token function
• optional: SwinTransformer-layers in the core downscaling model
• ensemble tail after output PerceiverIO

This development will be done in parallel with setting up a validation framework in the atmorep_analysis-repo and with training a competitor network (see this repo) based on the Harris WGAN (cf. Harris et al., 2022).

Text set up by @mlangguth89

[clessig]

It would be good to have an outline here for the approach and how it integrates with the overall developments.

CC @iluise

[mlangguth89]

@clessig Yes of course! We had our first detailed discussion on Thursday and I will add summary notes from this meeting in the description of this issue asap.

UPDATE: Summary has been added in the issue-description above.