Global model which blends antarctica and non-antarctica models.

[frodre]

For a while now, online runs have larger drifts appearing in the wintertime Antarctic when using prognostic ML. Time to give it a proper investigation.

Train ML models for 90 - 60 S

• [x] Dense - all scaled
• [x] RNN - all scaled
• [ ] Train models with limiters?

Prognostic Run

• [ ] 10-day
• [x] build model type that combines S Polar + global model
• [x] 6 hr

Analysis:

• [ ] set up a report w/ same structure as Noah's last few reports
  • [ ] compare zonal offline results w/ global model to see if demonstratable difference b4 combined
    • [ ] avg 60-90 S

[frodre]

Added fixes for antarctic only training #1619

[frodre]

Added fixes to fv3fit adapters for use in blending #1633

[frodre]

Composable model blending functionality #1643

[nbren12]

Control for this experiment: rnn-gscond-cloudtdep-cbfc4a.