Data-Driven Modelling of Sea Ice Dynamics

[Authors: Francisco Amor Roldán (NTNU, NERSC), Anton Korosv (NERSC)]

This repository contains the main code for an ML-based model of neXTSIM dynamics. It has a main source code directory in 'src', a directory with 'notebooks' containing different jupyter notebooks to interpret or communicate ideas in a more visual way, and some other directories for data,figures... The repository structure is the following:

├── environment.yml
├── example_data
├── LICENSE
├── notebooks
│   ├── examples
│   ├── Data_correlations.ipynb
│   ├── results.ipynb
│   └── roll_out.ipynb
├── README.md
└── src
    ├── datasets
    │   ├── create_MGN_dataset.py
    │   ├── Ice_graph_dataset.py
    ├── ice_graph
    │   ├── ice_graph.py
    ├── models
    │   ├── ConvGNNs_old.py
    │   ├── GUnet.py
    │   ├── MGN.py
    │   └── training_utils.py
    ├── sweep_config.yml
    ├── trainer_GNN.py
    ├── fine_tune_roll_out.py
    └── utils
        ├── graph_utils.py
        ├── metrics.py
        └── Tri_neighbors.py

Notebooks

Contains different notebooks to explore and visualize data or ideas in a more graphical and interactive way. The example folder contains didactical examples of how to open and process data from neXtSIM outputs. Data_correlations.ipynb contains scripts for preliminary data analysis, results.ipynb contains scripts to interpret model results, and roll_out.ipynb contains scripts to apply models iteratively.

Source (src)

Contains the main code developed in the project.

• datasets: Scripts to generate the datasets and data structures used by PyTorch Geometric to train the GNN models from neXtSIM outputs.
  • create_MGN_dataset.py
  • Ice_graph_dataset.py
• ice_graph: Contains the main code to handle and process neXtSIM output data in an object-oriented fashion. Methods to load, interpolate, find neighbors, and other utilities are implemented as methods.
  • ice_graph.py
• models: Implementation of the different GNN models used in the project alongside some training utilities.
  • ConvGNNs_old.py: Deprecated code for sea ice floe-based predictions used in a previous approach of this project.
  • GUnet.py: Code for Graph-Unets.
  • MGN.py: Code for the MeshGraphNet model.
  • training_utils.py: Scripts used during training.
• utils: General utilities used in the project.
  • graph_utils.py: Functions for graph data handling and normalization.
  • metrics.py: Functions for the metrics used.
  • Tri_neighbors.py: Class to work with node-level neighborhoods in the neXtSIM mesh.
• trainer_GNN.py: Code used for training the GNN models.
• fine_tune_roll_out.py: Code used for fine-tuning GNN models during roll-out.
• sweep_config.yml: Parameters used by Weights and Biases for hyperparameter optimization.

Setup Environment

To run the conda environment exported in environment.yml, it may take a while to install.

conda env create -f environment.yml
conda activate nextsim_ml