Adding NN evaluation for FV3

[tse-chunchen]

Description

To facilitate neural network (NN) based correction to the state variables, I would like to introduce subroutines that define and evaluate a typical feedforward NN and call it from the model (similar to https://github.com/AlexBelochitski-NOAA/fv3atm_full_physics_nn_emulator/).

Solution

• Current plan is to implement it in a similar way as the IAU update since they are both adding an extra term to the RHS of the tendency equations.
• The architecture and the trained weights and biases will be read in from an ascii file. This design is to ensure maximal compatibility with different machine learning packages with minimal effort.
• There will mainly be two subroutines: “init_NN” (initializing the NN) and “eval_NN” (evaluation of the NN) with flexibility of number of layers, neurons, etc. Both will be placed in the “atmos_cubed_sphere/tools/” (the same as the “fv_iau_mod.F90”).
• The two will be called in the “atmos_model.F90”:
  • The “init_NN” will be called in subroutine: atmos_model_init
  • The “eval_NN” together with the preparation of the input data will be called after the physics update in subroutine: “update_atmos_radiation_physics”
  • The output will then update the state variables through the API “atmosphere_state_update” of “atmosphere.mod”

Alternatives

No.

Testing:

No. Just raising the issue for now.

Dependent PRs:

ufs-community/ufs-weather-model/issues/#1184
NOAA-GFDL/GFDL_atmos_cubed_sphere/issues/#186
NOAA-EMC/fv3atm/issues/#522

@pjpegion @frolovsa @AlexBelochitski-NOAA