Add a conversion back to CAM variables before calling the net.

[jatkinson1000]

Add a conversion back to CAM variables before calling the net in order to reduce numerical error in calculation of deltas.

As discussed in #16 this will reduce numerical inaccuracies that could result from the variable conversion schemes and propagate into the tendency calculation.

Closes #16

[jatkinson1000]

There remains a question of what tabs we wish to use as an input to the neural net - the value interpolated from the CAM tabs, or the value after a conversion from SAM-variable-space back to CAM-variable-space?

[jatkinson1000]

As per my comment above the code here works as is, but I think could be clarified/simplified.

We currently pass tabs0_sam (tabs from conversion of interpolated variables in SAM space (t, q) back to CAM space) and tabs_sam (tabs interpolated from CAM) to the nn_convection_flux subroutine.

I think we only need one now that there is no subcycling inside the parameterisation.

• tabs_sam would be consistent with the other inputs to the net being directly as interpolated from CAM.
• tabs0_sam is the value we are actually taking the $\Delta$ in comparison to.

Currently we use the tabs0 as the input to the net, are you happy with this @paogorman or would you prefer to use tabs?

The most recent commit 3a7eb8c reduces to a single temperature input to the scheme.

[jatkinson1000]

@paogorman notes that this choice seems correct, but actually we could consider removing the second conversion as it could in-fact hurt performance* compared to some field campaigns, so consider removing?

Paul will confirm this and we will then either merge this, or remove the second conversion.

*because we are forcing the CAM thermodynamics towards the SAM thermodynamic setup.

[paogorman]

My recollection is that it didn't help, but I can't find it in my notes. In any event, it would be best to check with/without this conversion using the newest version of the code for TOGA and probably an aquaplanet too.

…

On Wed, Sep 11, 2024 at 10:43 AM Jack Atkinson ***@***.***> wrote: @paogorman <https://github.com/paogorman> notes that this choice seems correct, but actually we could consider removing the second conversion as it could in-fact hurt performance compared to some field campaigns, so consider removing? Paul will confirm this and we will then either merge this, or remove the second conversion. — Reply to this email directly, view it on GitHub <#21 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ARPLW64JI335KZ72XF7QYOTZWBJJBAVCNFSM6AAAAABKPHEL5KVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGNBTHA4DCMRXGY> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[jatkinson1000]

I have just rebased this back on top of CAL-ML which contains the new NN and moist-dry conversion.

Next step will be to run a comparison to see if it has any effect on the Toga II case.