| --- |
| license: apache-2.0 |
| language: |
| - en |
| tags: |
| - OneScience |
| - fluid-dynamics |
| - shallow-water-equations |
| frameworks: PyTorch |
| --- |
| <p align="center"> |
| <strong> |
| <span style="font-size: 30px;">PINN-SWE-CircularDamBreak</span> |
| </strong> |
| </p> |
| |
| # Model Introduction |
|
|
| PINN-SWE-CircularDamBreak reproduces the dCDB9 circular dam-break experiment for the two-dimensional shallow water equations. It predicts water depth and velocity without labeled training fields. |
|
|
| Paper: [Physics-Informed Neural Networks for Solving the Two-Dimensional Shallow Water Equations with Terrain Topography and Rainfall Source Terms](https://arxiv.org/pdf/2501.11372) |
|
|
| # Model Description |
|
|
| The network maps `(x, y, t)` to `(h, u, v)`. It has five fully connected hidden layers of width 300 with N-LAAF activations. Training uses shallow-water PDE residuals, initial conditions, and periodic boundary conditions. |
|
|
| ## Applications |
|
|
| | Application | Description | |
| | :--- | :--- | |
| | Circular dam-break prediction | Predict water depth and velocity over time | |
| | Shallow-water solver verification | Compare PINN fields with a numerical reference | |
|
|
| # Usage |
|
|
| ## 1. OneCode |
|
|
| [Open the OneCode AI4S environment](https://web-2069360198568017922-iaaj.ksai.scnet.cn:58043/home) |
|
|
| ## 2. Manual Setup |
|
|
| **Hardware** |
|
|
| - Formal training, inference, and evaluation were verified on one DCU compute node; the training script requires a Slurm DCU job environment. |
| - At least 30 GB of host memory is recommended. |
| - CPU is suitable for configuration inspection, not the recorded full training run. |
|
|
| ### Download the model |
|
|
| ```bash |
| hf download OneScience-Group/PINN-SWE-CircularDamBreak --local-dir ./PINN-SWE-CircularDamBreak |
| cd PINN-SWE-CircularDamBreak |
| ``` |
|
|
| ### Install dependencies |
|
|
| **DCU** |
|
|
| ```bash |
| # Activate a compatible DTK and Conda environment first. |
| conda create -n onescience311 python=3.11 -y |
| conda activate onescience311 |
| pip install onescience[cfd-dcu] -i http://mirrors.onescience.ai:3141/pypi/simple/ --trusted-host mirrors.onescience.ai |
| pip install numpy PyYAML matplotlib |
| ``` |
|
|
| **GPU** |
|
|
| ```bash |
| conda create -n onescience311 python=3.11 -y libstdcxx-ng=12 libgcc-ng=12 gcc_linux-64=12 gxx_linux-64=12 |
| conda activate onescience311 |
| pip install onescience[cfd-gpu] -i http://mirrors.onescience.ai:3141/pypi/simple/ --trusted-host mirrors.onescience.ai |
| pip install numpy PyYAML matplotlib |
| ``` |
|
|
| The formal scripts were verified only in a Slurm DCU environment. The GPU command is provided for dependency setup, not as a tested training claim. |
|
|
| ### Training data |
|
|
| No external dataset is needed. `scripts/fake_data.py` creates collocation points and targets from the configuration's PDE, initial condition, and periodic boundary conditions. |
|
|
| ```bash |
| python scripts/fake_data.py --config config/config.yaml |
| ``` |
|
|
| ### Train |
|
|
| The downloaded package includes a trained checkpoint. To train from scratch, first archive the bundled weights, then run on an allocated Slurm DCU node: |
|
|
| ```bash |
| mkdir -p weight/pretrained |
| mv weight/best_model.pth weight/last_model.pth weight/pretrained/ |
| python scripts/train.py --config config/config.yaml |
| ``` |
|
|
| The formal run takes 5,000 optimizer steps and writes `weight/best_model.pth`. |
|
|
| ### Trained weights |
|
|
| `weight/best_model.pth` was selected at step 4,100 after the full 5,000-step run and is ready for inference. File size: 1,466,146 bytes. SHA256: `26953fa8d218a585b05dd2775e1d3574a86ec33a2d68fb2bd58c1c78a8b3734e`. |
|
|
| ### Inference |
|
|
| ```bash |
| python scripts/inference.py --config config/config.yaml |
| ``` |
|
|
| ### Evaluation and visualization |
|
|
| ```bash |
| python scripts/result.py --config config/config.yaml |
| ``` |
|
|
| Outputs are written to `results/`. At `t=0.8 s`, the recorded water-depth RMSE is `0.01066` versus `0.01000` in paper Table 2. The recorded metric uses an independently implemented HLL reference, so it is not strict parity against the paper's exact reference array. |
|
|
| # OneScience Repositories |
|
|
| | Platform | Main repository | Skills repository | |
| | --- | --- | --- | |
| | Gitee | https://gitee.com/onescience-ai/onescience | https://gitee.com/onescience-ai/oneskills | |
| | GitHub | https://github.com/onescience-ai/OneScience | https://github.com/onescience-ai/oneskills | |
|
|
| # Citation and License |
|
|
| - Paper: [Physics-Informed Neural Networks for Solving the Two-Dimensional Shallow Water Equations with Terrain Topography and Rainfall Source Terms](https://arxiv.org/pdf/2501.11372). |
| - Authors' open repository: [PINN_SWE_open](https://github.com/tianyongsen/PINN_SWE_open), licensed under Apache License 2.0. This package is an independent reproduction and retains attribution. |
|
|
