Single column implementation with 93 vertical levels

[ma595]

This is the 93 vertical level implementation of the single column model in #28.

The PR removes all interpolation routines as the number of levels provided to the model matches CAM's dimensionality exactly for this test problem. We have reverted to a CPU only for performance reasons. No observed benefit was found when using GPU nodes.

This is currently being tested over longer time-scales - 10 months has been successfully executed.

Before merging:

• Check the tendency and flux output - then merge Add tendencies and fluxes to variable output #30
• Stable for runs up to a year - current output here /glade/derecho/scratch/matta/single_column_93_levels/run
• Code neatening - remove commented code
• Rebase
• Add some README information on how to do longer runs
• Clean up commit history
• Prepare for Aman testing:
  • Update README to provide clearer execution instructions

NOTES

• I've reverted to using cpu only . It scales reasonably well to 1024 cpus (8 nodes). I haven't tried more.

Additional instructions to execute and output tendencies and fluxes

For Aman's benefit, checkout the single-column-dev branch and follow instructions in the README.md.

• checkout this branch git checkout single-column-dev
• rebuild
• place scripted model in location and amend model path in user_nl_cam
• ./case.submit

Below is an example of a suitable user_nl_cam demonstrating the output of tendencies and fluxes. See ncar's docs for help on outputting cadence.

user_nl_cam:

use_gw_nlgw=.true.
gw_nlgw_model_path='/glade/u/home/matta/scripted-retrained_L93_ann_era5_global_global_uvthetaw_mseloss_train_epoch70.pt'

nhtfrq=0,5 # output every 5 steps
mfilt=1,60 # 60 outputs per file. 
avgflag_pertape = 'A','I'
fincl2 = 'U','V','T','UTGW_NL','VTGW_NL','UFLUX_NL','VFLUX_NL'