---
license: other
language:
- en
tags:
- OneScience
- computational-fluid-dynamics
- cylinder-wake
- physics-informed-neural-network
---
# TSA-PINN-CylinderWake

This package reproduces the Re=100 cylinder-wake experiment of [Physics-informed neural networks with trainable sinusoidal activation functions](https://publications.polymtl.ca/65937/1/2025_Khademi_Physics-informed_neural_networks_trainable.pdf). A four-layer, 50-unit PINN predicts velocity and pressure from space-time coordinates.

## Download

```bash
hf download OneScience-Group/TSA-PINN-CylinderWake --local-dir ./TSA-PINN-CylinderWake
cd TSA-PINN-CylinderWake
pip install -r requirements.txt
```

The original `cylinder_nektar_wake.mat` is available in the [authors' repository](https://github.com/AmirhosseinnnKhademi/TSA-PINN) and on [ModelScope](https://www.modelscope.cn/datasets/OneScience/TSA-PINN-CylinderWake). Place it at `data/cylinder_nektar_wake.mat`. No dataset is mirrored in this Hugging Face model repository.

## Run

```bash
python scripts/inference.py
python scripts/result.py
# To repeat the full 4000-update run on a GPU/DCU:
python scripts/train.py
```

The included `weight/best_model.pt` is from update 3956; its SHA256 is `5fadc91405b4741288750c775c09d043c205a625e630d1ee2a9e8b5022712e0f`. At t=10, measured DNS relative L2 errors are 0.021981 for u and 0.083490 for v. The paper reports 0.01960 and 0.06630, but its aggregate evaluation convention is not fully specified, so the figures are not a strict like-for-like comparison.

See `config/config.yaml` and `results/` for the experiment settings and outputs. The original data's usage terms should be checked separately.