NOAA-GFDL's Sea Ice Simulator version 2
The United States Department of Commerce (DOC) GitHub project code is provided on an "as is" basis and the user assumes responsibility for its use. DOC has relinquished control of the information and no longer has responsibility to protect the integrity, confidentiality, or availability of the information. Any claims against the Department of Commerce stemming from the use of its GitHub project will be governed by all applicable Federal law. Any reference to specific commercial products, processes, or services by service mark, trademark, manufacturer, or otherwise, does not constitute or imply their endorsement, recommendation or favoring by the Department of Commerce. The Department of Commerce seal and logo, or the seal and logo of a DOC bureau, shall not be used in any manner to imply endorsement of any commercial product or activity by DOC or the United States Government.
This project code is made available through GitHub but is managed by NOAA-GFDL at https://gitlab.gfdl.noaa.gov.
To use SIS2 with ML-based bias correction, you can compile MOM6-SIS2 in coupled mode, as outlined in the MOM6-examples wiki. You will just need to make sure that when you fork MOM6-examples, you change the path to the SIS2 repo in .gitmodules to https://github.com/m2lines/SIS2.git. Then you can do:
git clone --recursive https://github.com/yourGithub/MOM6-examples.git MOM6-examples
cd MOM6-examples/src/SIS2
git checkout dev/m2lines
One you compile the source code, all you then need to do is run a simulation with the correct overrides. An example job script can be found here. To run in a forced ice-ocean configuration, see the following overrides in the job script:
#override DO_ML = True
#override ML_CPL = False
#override CNN_HALO_SIZE = 4
#override CNN_WEIGHTS = "/path/to/SIS2/ML_weights/NetworkA_weights_IO_1982-2017.nc"
#override ANN_WEIGHTS = "/path/to/SIS2/ML_weights/NetworkB_weights_IO_1982-2017.nc"
The flag ML_CPL will determine what normalization statistics are applied to the network inputs. If running in ice-ocean, set this to False, while if running in SPEAR, set this to True (Unfortunately the atmosphere and land model source code for SPEAR are not made publicly available by GFDL). You do not need to change CNN_HALO_SIZE. The ML_FREQ flag then determines how often to do inference (in seconds)---this same correction will then be applied at every timestep between inference steps. The CNN_WEIGHTS and ANN_WEIGHTS overrides are then the paths to the network weight netcdf files. These are ravelled vectors of the weights saved from PyTorch.
Tip - Besides the SIS_ML.F90 code, look for the !WG flags in the other SIS2 files to understand how the ML inputs are gathered and where inference is called.