---
license: apache-2.0
language:
- en
tags:
- OneScience
- computational-fluid-dynamics
- neural-operator
- pipe-flow
---
# HO-FNO-Pipe-Flow

This package reproduces the Pipe experiment of [Higher-Order Fourier Neural Operator](https://arxiv.org/abs/2606.28122). An eight-layer, second-order spectral operator maps two pipe-coordinate channels to horizontal velocity on a 129 x 129 grid.

## Download

```bash
hf download OneScience-Group/HO-FNO-Pipe-Flow --local-dir ./HO-FNO-Pipe-Flow
cd HO-FNO-Pipe-Flow
```

Install PyTorch, NumPy, PyYAML and Matplotlib in a compatible GPU/DCU environment. The original 2310-sample Pipe dataset is available from the [paper's official repository](https://github.com/AlexColagrande/HO-FNO) and as the [OneScience ModelScope dataset](https://www.modelscope.cn/datasets/OneScience/HO-FNO-Pipe). Download `Pipe_X.npy`, `Pipe_Y.npy`, and `Pipe_Q.npy` into `data/`; no dataset files are included in this Hugging Face model repository.

## Run

```bash
python scripts/inference.py
python scripts/result.py
# To repeat the 500-epoch training run on an accelerator:
python scripts/train.py
```

The included `weight/best_model.pth` is from epoch 494. Its 200-case test mean relative L2 is 0.004887735; the paper reports 0.0049 for the Pipe setting. The weight SHA256 is `9d71053c64b62a6fc6f867108c29d2cc0d277d667a3504dc073a6b56072981d4`. See `results/` for the recorded metrics and plots.

## Source and license

Model code in this reproduction is distributed under Apache 2.0. The underlying Pipe data has separate source terms; consult the [original project](https://github.com/AlexColagrande/HO-FNO) before reuse.