---
license: other
language:
- en
tags:
- OneScience
- computational-fluid-dynamics
- darcy-flow
- neural-operator
---
# VINO-DarcyFlow

This package reproduces the 64 x 64 Darcy-flow experiment from [Variational Physics-informed Neural Operator](https://arxiv.org/abs/2411.06587). It predicts pressure from permeability using a variational energy objective; training labels are not used to update the model.

## Download

```bash
hf download OneScience-Group/VINO-DarcyFlow --local-dir ./VINO-DarcyFlow
cd VINO-DarcyFlow
pip install -r requirements.txt
```

The original `darcy2d_s64.pt` comes from the [authors' data directory](https://seafile.cloud.uni-hannover.de/d/4341006631ab427b8270/). A provenance-preserving copy is available on [ModelScope](https://www.modelscope.cn/datasets/OneScience/VINO-Darcy). Place that file at `data/darcy2d_s64.pt`; this Hugging Face repository contains the model only. Only load PyTorch-serialized data from a trusted source.

## Run

```bash
python scripts/inference.py
python scripts/result.py
# To repeat the full 500-epoch run on a GPU/DCU:
python scripts/train.py
```

The included `weight/best_model.pt` is from epoch 435 and has SHA256 `e9b47a201bc03cc9362b4803706b6529426b8fc889ce9f185b8fec52cf9c51f8`. Its mean test relative L2 was 1.109766%, versus 0.93% in the paper. The best epoch was selected using the same test set, so that result is not an independent test estimate.

See `config/config.yaml` for the 1000/100 split and `results/` for the recorded run. Data licensing is separate from this model package; consult the original source.