overall plan

[TomMelt]

Non-Local GW project summary

As noted in the resource application the ultimate goal of this project is:

The focus is on coupling a series of offline-trained neural networks written in PyTorch to the Fortran-based climate model CAM.

There are 3 proposed NNs:

• single column (1 x 1)
• 3 x 3 CNN
• global grid (UNet)

Note

After initial discussions, we have agreed that the single column and global UNet approaches will be the simplest to implement (due to difficulties communicating non-local regions on an unstructured grid). So the plan is to target single column NN and global NN first. If this is successful we will look at implementing the 3 X 3 CNN.

Milestones

1. Setup and run and minimum worked example of CESM (based on CAM7)
   • running with intended correct resolution e.g., 2.5 deg or coarser (must be coarse enough to not resolve GWs)
2. Run and trace python NNs
   • we need the traced .pt models for FTorch
3. Create two MWEs (Python / Fortran) that run inference for a given set of inputs
   • given same inputs both Fortran & Python model should predict same output
4. Agree which part of CAM is to be replaced (all GW's but what routines?)
   • Do we predict fluxes or forcings?
5. Create a branch of CAM that uses FTorch to replace GW routines with a single call to one of the 3 proposed NN's

What does success look like?

From the resource request

• The models will be evaluated on robustly defined statistics like the period of the tropical QBO and QBO-related variability, the frequency of Sudden warming events in the midlatitude stratosphere, and lastly, the springtime breakdown time of the southern hemispheric polar vortex. [...]

• In addition, the global time-averaged climatological distribution of the predicted fluxes and forcing will also be tested and compared with existing, documented climatology.

(see here for more context)

Added since kick-off meeting

• Fortran and Python MWEs should predict the same output for a given set of inputs
• CAM model should be stable over long time periods