SWAN_DeepONet

Operator Learning for Predicting Bulk Wave Parameters of Spectral Wave Models

This repository contains the code, example notebooks, and data-generation setup for a DeepONet surrogate of the spectral wave model SWAN. The surrogate predicts bulk wave quantities (significant wave height (H_sig) and the wave-induced radiation-stress forcing), and is evaluated on idealized 1D cases and on the field-scale DUCK case.

The large binary assets (training/validation data, trained model weights, and the sampling-sensitivity model set) are not stored in this repository. They are on the DesignSafe-CI Data Depot under project PRJ-6235 — Operator Learning for Predicting Bulk Wave Parameters of Spectral Wave Models and can be downloaded from there.

---

Repository structure

NeuralOperator-CoastalWaves/
├── analysis_tools/          # Shared Python modules (data extraction, scaling, metrics, plotting)
├── examples/                # End-to-end notebooks that reproduce the paper figures
│   ├── 1d/                  #   Idealized 1D cases
│   └── duck/                #   Field-scale DUCK case
├── baseline_mlp/            # MLP baseline for comparison against the neural operator
│   └── MLP/
├── cross_grid_analysis/     # Cross-grid / grid-refinement generalization study (DUCK)
├── runtime_benchmark/       # Execution-time comparison: surrogate inference vs. SWAN
├── sampling_sensitivity/    # How training-sample selection affects surrogate accuracy
├── swan_data_generation/    # SWAN control files (*.swn), bathymetry, data-gen scripts
│   └── CrossGridValidation/
├── Models/
│   └── scripts/             # Training scripts (trained weights hosted on DesignSafe)
├── README.md
└── gitpush.sh

# Downloaded from DesignSafe (PRJ-6235), placed at repo root
#   Data/                        SWAN input/output for training, validation, evaluation
#   Models/ (weights + scalers)  Trained neural-operator & MLP weights
#   sampling_sensitivity_models/ Model set for the sampling-sensitivity study

Path	Description
analysis_tools/	Reusable Python modules: data extraction, scaling/preprocessing, error metrics, and plotting. Imported by the example notebooks.
examples/	End-to-end notebooks that reproduce the paper figures (1d/, duck/). Start here.
baseline_mlp/	Multilayer-perceptron (MLP) baseline used for comparison against the neural operator.
cross_grid_analysis/	Cross-grid / grid-refinement generalization study (DUCK).
runtime_benchmark/	Execution-time comparison: surrogate inference vs. running SWAN.
sampling_sensitivity/	Study of how training-sample selection affects surrogate accuracy.
swan_data_generation/	SWAN control files (*.swn), bathymetry, and scripts used to generate the training data.
Models/scripts/	Training scripts for the neural-operator models (trained weights live on DesignSafe).

Data on DesignSafe

Download these from PRJ-6235 and place them at the repository root. They are on the

DesignSafe Data Depot under project PRJ-6235 — Operator Learning for Predicting Bulk Wave Parameters of Spectral Wave Models.

Path	Contents
Data/	SWAN input/output used for training, validation, and evaluation.
Models/	Trained neural-operator and MLP weights + fitted scalers.
sampling_sensitivity_models/	Model set for the sampling-sensitivity study.

---

Getting started

Environment

The analysis code uses Python 3.11 with:

numpy  pandas  matplotlib  cmocean  scikit-learn
tensorflow  joblib  tqdm  netCDF4

We recommend a dedicated environment, e.g.:

conda create -n coastalwaves python=3.11
conda activate coastalwaves
pip install numpy pandas matplotlib cmocean scikit-learn tensorflow joblib tqdm netCDF4

Reproducing the results

1. Download Data/ and Models/ from DesignSafe (PRJ-6235) into the repository root.
2. Open a notebook under examples/ (e.g. examples/duck/duck_analysis.ipynb) and run it top to bottom. The notebooks import the shared modules in analysis_tools/.

---

Citation

If you use this code or data, please cite the associated publication and the DesignSafe dataset (PRJ-6235).

@article{
  title   = {Operator Learning for Predicting Bulk Wave Parameters of Spectral Wave Models},
  author  = {Shukai Cai, Sourav Dutta, Mark Loveland, Eirik Valseth, Peter Rivera-Casillas, Corey Trahan, Clint Dawson},
  journal = {Ocean Engineering},
  volume = {367},
  pages = {127866},
  year = {2026},
  issn = {0029-8018},
  doi = {https://doi.org/10.1016/j.oceaneng.2026.127866},
  url = {https://www.sciencedirect.com/science/article/pii/S0029801826037005}
}

```bibtex
@misc{
  title     = {Operator Learning for Predicting Bulk Wave Parameters of Spectral Wave Models[Dataset]},
  author    = {Cai, Shukai and Dutta, Sourav and Loveland, Mark and Valseth, Eirik and
               Rivera-Casillas, Peter and Trahan, Corey and Dawson, Clinton N.},
  publisher = {DesignSafe-CI},
  year      = {2026},
  doi       = {10.17603/DS2-DKHW-PT40},
  url       = {https://doi.org/10.17603/DS2-DKHW-PT40}
}