MITONet for 2D Shallow Water Equations

This repository consolidates the MITONet 2d shallow-water emulator workflows for both the Shinnecock Inlet and Red River studies that appear in the paper "A Neural Operator Emulator for Coastal and Riverine Shallow Water Dynamics". Each case retains its own training scripts, configuration files, and notebooks, while the shared data directory and licensing live at the repo root.

Repository Layout

MITONet/
├── cases/
│   ├── Shinnecock/
│   │   ├── scripts/         # Training + optimization entry points for every model flavor
│   │   ├── settings/        # Configuration for MITONet and the baseline operators
│   │   ├── notebooks/       # Visualization utilities for the Shinnecock study
│   │   └── src/             # Model components and data-processing utilities
│   └── RedRiver/
│       ├── scripts/         # MITONet training + optimization scripts for Red River
│       ├── settings/
│       ├── notebooks/
│       └── src/
├── data/                    # Placeholder for hydrodynamic simulation data
├── LICENSE
└── README.md

Each case directory is self-contained, so you can cd cases/<Case>/scripts and run the same commands you used previously. The shared data/ directory lives alongside cases/ and can store datasets that are reused across both scenarios. Add subfolders inside data/ if you wish to keep case-specific raw files separated (e.g., data/shinnecock/, data/redriver/).

Environment Setup

Create and activate the conda environment (from anywhere inside the repo):

conda create -n mitonet -c conda-forge tensorflow=2.14 matplotlib tqdm scikit-learn optuna optuna-integration netcdf4 pandas cmocean
conda activate mitonet

Any additional packages that are case-specific can be added within each case directory's README or scripts as needed.

Data

Each case has its own dataset hosted on DesignSafe-CI. Download the appropriate archive, then store it under data/<case>/... so files do not mix:

Shinnecock: https://designsafe-ci.org/data/browser/public/designsafe.storage.published/PRJ-5716
Red River:  https://www.designsafe-ci.org/data/browser/public/designsafe.storage.published/PRJ-6207/#detail-5855ec2a-cd4d-423b-80a4-93769c023f29

Extract each archive directly under data/PRJ-5716 and data/PRJ-6207 and keep the publisher-provided folder names (for example, Simulation--2d-adcirc-simulation-of-tidal-flow-in-shinnecock-inlet-ny-parameterized-by-bottom-friction-coefficient). The settings modules already resolve the nested Model--adcirc-model/data and Output--processed-2d-adh-output/data paths, so no manual path editing is required as long as the extracted folders remain intact. Keeping the raw, processed, and intermediate files separated by case prevents accidental cross-contamination of meshes, checkpoints, or scalers.

Running Training / Evaluation

1. cd cases/Shinnecock/scripts (or cases/RedRiver/scripts).
2. Run the desired script:
   • python mito_net.py for MITONet training.
   • python mito_optuna.py to perform hyperparameter search before training.
   • Use the additional Shinnecock-specific entry points (mio_net.py, mdo_net.py, etc.) as needed.
3. Update configuration parameters inside the corresponding settings/ directory.
4. Visualization notebooks live in cases/<Case>/notebooks/Visualization.ipynb once training artifacts exist.

Version Control

This folder is the canonical Git repository for both cases. Commit changes from the repo root so diffs clearly show whether you touched shared data or a particular case under cases/.

Citation

If you use this repository, please cite our paper:

@misc{riveracasillas2026neuraloperatoremulatorcoastal,
      title={A Neural Operator Emulator for Coastal and Riverine Shallow Water Dynamics}, 
      author={Peter Rivera-Casillas and Sourav Dutta and Shukai Cai and Mark Loveland and Kamaljyoti Nath and Khemraj Shukla and Corey Trahan and Jonghyun Lee and Matthew Farthing and Clint Dawson},
      year={2026},
      eprint={2502.14782},
      archivePrefix={arXiv},
      primaryClass={cs.CE},
      url={https://arxiv.org/abs/2502.14782}, 
}