OneScience-Group/Spectral-Refiner

Spectral-Refiner

Model Introduction

Spectral-Refiner is a physics-residual fine-tuning method for spatiotemporal Fourier neural operators. This project reproduces the two-dimensional forced-turbulence experiment from Table 2 of the paper: an SFNO is first trained on 64 x 64 vorticity trajectories, then its spectral output layer is fine-tuned at 256 x 256 resolution using the (H^{-1}) negative Sobolev norm of the Navier–Stokes PDE residual.

Paper: Spectral-Refiner: Accurate Fine-Tuning of Spatiotemporal Fourier Neural Operator for Turbulent Flows

Model Description

The model takes the first 10 time steps of a two-dimensional vorticity field and predicts the next 40. The base SFNO has four spatiotemporal spectral layers, 12 x 12 spatial modes, 5 temporal modes, and width 20. Spectral-Refiner freezes the base network, expands the output spectral layer to 64 x 64 x 6, and fine-tunes it for 50 steps with an (H^{-1}) PDE-residual objective. This project is an independent OneScience reproduction.

Intended Uses

Use case	Description
2D turbulence prediction	Predict future spatiotemporal evolution from historical vorticity fields.
PDE surrogate	Accelerate periodic fluid problems with a Fourier neural operator.
Physics-residual fine-tuning	Constrain the predicted PDE residual with a negative Sobolev norm.

Usage

1. OneCode

Launch the OneCode AI-for-Science environment

2. Manual Setup

Hardware requirements

• A GPU or DCU is recommended for full training and inference.
• A CPU can run imports and the --smoke connectivity test.
• The reported validation used PyTorch 2.5.1 on one DCU.

Download the model repository from Hugging Face

pip install -U huggingface_hub
hf download OneScience-Group/Spectral-Refiner --local-dir ./Spectral-Refiner
cd Spectral-Refiner

Install 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

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

Download the training dataset from Hugging Face

hf download OneScience-Group/fno \
  --repo-type dataset \
  --local-dir ./data

The required files are:

data/
├── fnodata_extra_64x64_N1280_v1e-3_T50_steps100_alpha2.5_tau7.pt
└── fnodata_extra_fp64_256x256_N16_v1e-3_T50_steps100_alpha2.5_tau7.pt

Set data.train_file and data.test_file in config/config.yaml to these files. The default experiment uses 1,152 low-resolution trajectories for training and 128 for validation, mapping 10 input steps to 40 output steps. The high-resolution data is used for evaluation and (H^{-1}) spectral fine-tuning on the 256 x 256 grid.

Train

python scripts/train.py --config config/config.yaml

weight/best_model.pt stores the base SFNO, Spectral-Refiner output layer, configuration, and weight-selection metric.

Inference

python scripts/inference.py

Predictions are saved to results/predictions.pt.

Evaluation

python scripts/result.py

Metrics are printed and saved to results/metrics.json.

OneScience

Platform	OneScience repository	OneSkills 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: Spectral-Refiner, arXiv:2405.17211.
• Public implementation: scaomath/torch-cfd.
• This repository uses the Hugging Face-compatible MIT identifier (mit). Dataset files and other third-party assets retain their original licenses and terms.