The code in this repository reproduces the results of our nudge-to-fine-res paper using the yearlong fine-grid run. The fine-grid C3072 simulation was run on NOAA's Gaea supercomputer. This data was made available to us at an intermediate coarsened C384 resolution and stored in the AI2 Climate Modeling Google Cloud Storage buckets. Below is an outline of the workflow to reproduce the training data, ML model training, offline and prognostic evaluation, and figures.
Before submitting any argo workflows, make sure to make deploy in the workflows directory to use the correct images.
The fine-grid simulation data is stored at C384 resolution (coarsened down from its native C3072 resolution).
Running scripts/coarsen_restarts.sh and scripts/coarsen_pire_dyn.sh will coarsen this data to C48 resolution, which is the resolution of our coarse-grid FV3GFS model.
The nudged run has the sea ice fraction, land sea mask, SSTs, precipitation, and surface radiative fluxes prescribed to the values from the coarsened fine-res reference.
Running the scripts scripts/prescribed_seaice.py, scripts/prescribed_ssts.py, scripts/prescribed-sfc-fluxes-precip.py will create the datasets of prescribed fields for use in the nudged run.
The nudged run is created by submitting the argo workflow via workflows/nudge-to-fine-run/run.sh.
Models used in the final results as well as sensitivity tests are trained by running the */run.sh scripts in the subdirectories of train-evaluate-prognostic-run.
The offline diagnostics are also created automatically by those argo workflows.
Baseline and ML-corrected coarse-grid simulations can be reproduced by running the run.sh scripts in the correspondingly named subdirectories of prognostic-run.
Figures from the paper are created in the notebooks in the folder figure_notebooks. The notebook figure_notebooks/composite_figures.ipynb reduces the size of the eps files and combines subfigures into a single file.