license: apache-2.0
language:
- en
tags:
- OneScience
- fluid-dynamics
- darcy-flow
- pytorch
library_name: pytorch
FAST-DeepONet-DarcyFlow
Model Introduction
FAST-DeepONet-DarcyFlow predicts a steady Darcy pressure field from a permeability field in a small-data setting.
Model Description
This package reproduces the localized-inclusion Darcy flow experiment in Table 2. A factor-augmented branch network and coordinate trunk predict a pressure field on a 64x64 grid.
Intended Use
| Task | Description |
|---|---|
| Darcy pressure prediction | Estimate steady pressure from permeability fields |
| PDE surrogate research | Study low-data operator learning |
Usage
1. OneCode
2. Manual Setup
Hardware requirements
A GPU or DCU is recommended. CPU can be used for a small inference check.
Download the Model Package
hf download OneScience-Group/FAST-DeepONet-DarcyFlow --local-dir ./FAST-DeepONet-DarcyFlow
cd FAST-DeepONet-DarcyFlow
Set Up the Runtime Environment
DCU environment
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 scipy pyyaml matplotlib
GPU environment
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 scipy pyyaml matplotlib
Training Data
Run the included script from the model package root to generate the Darcy-inclusion data. It follows the paper's experiment settings with seed 33 and needs only CPU, NumPy, and SciPy:
python scripts/generate_data.py
The script creates 100 examples under data/train/ (60), data/tune/ (10), data/calib/ (20), and data/test/ (10), plus data/DATA_MANIFEST.json. Each .npz contains init_value, last_value, and sparse_mask, each with 4096 float32 values. All three arrays in all 100 regenerated examples were checked against the data used for the bundled checkpoint and matched exactly. No generated dataset files are bundled with this model package.
Training
From the model package root, run:
python scripts/train.py
The configured run is 300 epochs. The training script no longer requires a particular Slurm installation; use the package in a suitable GPU or DCU environment.
Checkpoint
The trained checkpoint is weight/best_model.pth (1,240,394 bytes; SHA256 a6dd45b968a3d8f48f180f09bf4117d566c0aa95821f88ad7a608b925ecabdb4). The reproduced test mean relative L2 is 0.0958688; Table 2 reports 0.0925 +/- 0.0019 over five runs and a larger test set, so the figures are only approximately comparable.
Inference
python scripts/inference.py
Evaluation and Visualization
python scripts/result.py
Official OneScience Information
| 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: FAST-DeepONet. Official implementation: fast-deeponet. Verify the generated dataset's release terms before redistribution.