Matching Neural Net input-outputs to SCAM input-outputs

m2lines / convection-parameterization-in-CAM

Deployment of the convection parameterization scheme (previously from within SAM) deployed into the CAM model

MIT License

3 stars 2 forks source link

Matching Neural Net input-outputs to SCAM input-outputs #2

Closed dorchard closed 1 year ago

dorchard commented 1 year ago

We will need to suitably re-interface the Neural Net model, but something we will need to solve is that the co-ordinate schemes between (previously used) SAM and SCAM differ. Therefore some translation with interpolation will be needed. This will likely need a lot of steer from Judith and Paul.

Identify input-output interface for the model within SCAM
Identify the translation and interpolation needed in the co-ordinate schemes
Design translation scheme
Write tests to ensure that the translation in-and-out is an isomorphism (no junk no noise).
Identify CESM/SCAM metrics we can use to test end-to-end after parameterization inserted (continuous integration test)

dorchard commented 1 year ago

From @paogorman's M2LInES talk the following 'impedance matching' activities are needed (with some rephrasing from me):

Match vertical coordinate: Interpolate from pressure in CAM to/from reference pressure profile in anelastic SAM
Variable conversions: convert tendency of liquid/ice static energy to temperature
Energy conservation: conserve enthalpy in CAM versus static energy in anelastic SAM (use anelastic-primitive equation analogy). @paogorman does this come for free from some variable conversion, or are we going to need to do something extra?
Timesteps: subcycling due to mismatch in timesteps

I think it might be best for us to break these into 4 separate issues actually.

jatkinson1000 commented 1 year ago

The first three may well need to be combined as I suspect they will be interlinked.

The subcycling I agree should be tackled elsewhere, as it is likely non-trivial.

paogorman commented 1 year ago

I expect energy conservation will come for free with correct choice of variables

dorchard commented 1 year ago

Made a separate issue for subcycling then https://github.com/m2lines/convection-parameterization-in-CAM/issues/37

mondus commented 1 year ago

Conversation has moved to a private google document to permit open discussion and detailed comments. Once decisions have been made on this they will be moved into issues.