This issue will coordinate the different loss scaling work being conducted.

There are a series of issues that we will work on:

• The use of 'variable' scalars for loss functions and how to use them for validation metrics
• The pressure level scalar configuration does not work as expected (see separate issue Pressure Level Scalings only applied in specific circumstances #7)
• Do we need to change the code pattern to better distinguish between loss scalings we have: ie applied per variables, scalings applied for vertical levels or spatial scalings?
• Related to the above point, where should the scaling code live? At the moment much of the scaling functionality is having to take place on forecaster.py which could quickly become bloated

There is a PR on stretched grid RMSE ecmwf/anemoi-training#199
and a draft PR on using tendency scale losses ecmwf/anemoi-training#186.

There will be a series of future PRs:

1. To fix how the variables are scaled in the validation metrics (@HCookie @jswijnands @havardhhaugen) Fix 'all' validation metrics anemoi-training#202
2. To fix the pressure level scalings (@sahahner; @pinnstorm; @mc4117)

PR's

• 'All' Validation Metric Fix 'all' validation metrics anemoi-training#202
• Pressure level scaling seperation