This repository trains a PyTorch convolutional neural network (CNN) to predict subgrid ocean momentum forcing from ocean surface velocity, intended for coupling with larger GCMs to provide a performant, high-fidelity parameterization in coarse-resolution climate models.
Command-line scripts for preparing training data, training up a model, testing model performance, and using the model to make predictions (inference mode) are provided.
For further detail and discussion, please see Arthur P. Guillaumin, Laure Zanna (2021). Stochastic-deep learning parameterization of ocean momentum forcing which originally introduced this work. Documentation in this repository will refer back to sections from the paper e.g. Guillaumin (2021) 2.1 to provide context and further reading. (A snapshot of the code used in the paper can be found on Zenodo.)
This repository also aims to enable reproducing the 2021 paper. The Jupyter
notebooks at resources/jupyter-notebooks
generate some figures shown in the paper.
Model training and usage is separated into a handful of steps. Steps are executed via a command-line interface (CLI) Python script, and will save some data to disk to then be loaded in the next step.
In the "data" step, we generate training data using simulation data from the CM2.6 climate model (which we refer to as the CM2.6 dataset, or just CM2.6). We calculate the subgrid forcing needed for coarse-resolution models using the high-resolution ocean velocity data in the CM2.6 dataset, then coarsen. This coarsened, with-forcings dataset is saved to disk. You may generate training data using either the "control" CM2.6 simulation, or the "1-percent annual CO2 increase" one. (See Guillaumin (2021) 2.1.)
In the "training" step, we train a NN to predict the true forcing from the coarse velocity data generated above. This forcing term tends to have a large amount of uncertainty. Rather than a single value, we predict both the mean and standard deviation of a Gaussian probability distribution for the forcing. This allows for stochastic implementations in online models. (See Guillaumin (2021) 2.3 for a more in-depth explanation and how to interpret the NN output.)
In the "testing" step, we test a trained model on an unseen region of data (the subset not used in the previous training step).
We also provide a basic script for predicting forcings on a prepared dataset.
src: source code (library and CLI scripts)tests: pytest testsdocs: detailed project documentation, implementation notesresources: CLI configs, Jupyter notebooksflake.nix, flake.lock: helper files for building on Nix (ignore)Python 3.9 or newer is required. We primarily test on Python 3.11.
To avoid any conflicts with local packages, we recommend using a virtual environment. In the root directory:
python -m venv venvor using virtualenv:
virtualenv venvThen load with source venv/bin/activate.
With pip installed, run the following in the root directory:
pip install -e .(An alternate pyproject.toml file is provided for building with
Poetry. To use, rename pyproject-poetry.toml to
pyproject.toml (overwriting the existing file) and use Poetry as normal. Note
that the Poetry build is not actively supported-- if it fails, check that the
dependencies are up-to-date with the setuptools pyproject.toml.)
Note that if you are running Python 3.9 or older, you may also need to install
the GEOS library, due to cartopy requiring it. (Newer
versions moved away from the C dependency.)
Execute these commands from the repository root.
See docs directory for more details.
For command-line option explanation, run the appropriate step with --help e.g.
python src/gz21_ocean_momentum/cli/data.py --help.
Most CLI scripts support reading in options from a YAML file using a
--config-file flag. In general, a flag --name value will be converted to a
top-level name: value line. Examples are provided in
resources/cli-configs. CLI options override file
options, so you may provide partial configuration in a file and fill out the
rest (e.g. file paths) on the command line.
There are a handful of unit tests using pytest, in the tests
directory. These assert some operations and methods used in the steps. They may
be run in the regular method:
pytestcli/data.py calculates coarse
surface velocities and diagnosed forcings from the CM2.6 dataset and saves them
to disk. This is used as training data for the neural net.
You must configure GCP credentials in order to download the CM2.6 dataset.
See docs/data.md for more details.
Example invocation:
python src/gz21_ocean_momentum/cli/data.py \
--lat-min -80 --lat-max 80 --long-min -280 --long-max 80 \
--factor 4 --ntimes 100 --co2-increase --out-dir forcingsAlternatively, you may write (all or part of) these options into a YAML file:
lat-min: -80
lat-max: 80
long-min: -280
long-max: 80
ntimes: 100
factor: 4
co2-increase: trueand use this file in an invocation with the --config-file option:
python src/gz21_ocean_momentum/cli/data.py \
--config-file resources/cli-configs/data-paper.yaml --out-dir forcingsSome preprocessed data is hosted on HuggingFace at datasets/M2LInES/gz21-forcing-cm26.
The cli/train.py script trains the model using data generated
previously. You may configure various training parameters through command-line
arguments, such as number of training epochs, loss function, etc.
Example invocation:
python src/gz21_ocean_momentum/cli/train.py \
--lat-min -80 --lat-max 80 --long-min -280 --long-max 80 \
--factor 4 --ntimes 100 --co2-increase --out-dir forcings \
--train-split-end 0.8 --test-split-start 0.85 \
--subdomains-file resources/cli-configs/train-subdomains-paper.yaml \
--forcing-data-path <forcing zarr dir>You may place options into a YAML file and load with the --config-file option.
Notable parameters:
--subdomains-file: path to YAML file storing a list of subdomains to select
from the forcing data, which are then used for training. (Note that at
runtime, domains are be truncated to the size of the smallest domain in terms
of number of points.)--train-split-end: use 0->N percent of the dataset for training--test-split-start: use N->100 percent of the dataset for testingThe --subdomains-file format is a YAML list of bounding boxes, each defined
using four labelled floats:
- lat-min: 35
lat-max: 50
long-min: -50
long-max: -20
- lat-min: -40
lat-max: -25
long-min: -180
long-max: -162
# - ...lat-min must be smaller than lat-max, likewise for long-min.
Note: Ensure that the subdomains you use are contained in the domain of the forcing data you use. If they aren't, you may get a confusing Python error along the lines of:
RuntimeError: Calculated padded input size per channel: <smaller than 5 x 5>.
Kernel size: (5 x 5). Kernel size can't be greater than actual input sizeThe cli/infer.py script allows loading a trained model and a set
of (low resolution) velocity data, and predicts forcings.
Example invocation:
python src/gz21_ocean_momentum/cli/infer.py \
--model-state-dict-file model.pth \
--input-data-dir <forcing zarr dir> \
--device cuda:0We have tested with data from the forcing generation step -- the forcings are not used, it is to obtain low-resolution velocities.
See Guillaumin (2021) for detail on how to use model output.
The resources/jupyter-notebooks folder stores notebooks developed during early project development, some of which were used to generate figures used in the 2021 paper. See the readme in the above folder for details.
There is GZ21 Ocean Momentum data available on HuggingFace
As of 2023-12-08, these are currently low-resolution: forcings generated for few time points (100 vs. 4000 available), and a model trained on that data.
We are not currently accepting contributions outside of the M2LInES and ICCS projects until we have reached a code release milestone.
This repository is provided under the MIT license. See LICENSE for
license text and copyright information.
See CITATION.cff for citation metadata for this software. (For further
details on usage, see https://citation-file-format.github.io/ .)