search

nbren12 / uwnet

Neural Networks based unified physics parameterization for atmospheric models
MIT License
3 stars 2 forks source link

Implement digital filter initialization for SAM #17

Closed nbren12 closed 6 years ago

nbren12 commented 6 years ago

For some reason, the NGAqua data does not conserve mass: https://atmos.washington.edu/~nbren12/reports/uwnet/notebooks/3.2-mass-conservation-problems.html. This leads to very large errors in the large-scale forcing using SAM. The problem is especially bad in the boundary layer.

After reviewing the data assimilation literature, it seems that so-called Digital Filter Initialization (DFI) is an effective way to initialize a model with un-balanced initial conditions. This is what WRF uses. To do this, we will need to run SAM backwards and forwards in time from a certain initial condition.

nbren12 commented 6 years ago

Actually, I am not sure this will help much before we solve the issues with the boundary conditions for the pressure solver.

nbren12 commented 6 years ago

After much discussion, we found that the zonal-striping and mass conservation issues were caused by using a single precision poisson solve for the pressure. Changing to double precision greatly improves the simulation.

That said, DFI might still be useful for filtering the inputs to the machine learning procedure in a model-aware way.

nbren12 commented 6 years ago

I am going to close, this...it's a good idea but not necessary for now.