OneScience-Group/Functional_Attention

Functional Attention

Model Introduction

Functional Attention is a resolution-independent operator-learning framework developed by researchers affiliated with the Technical University of Munich. It predicts continuous functions such as PDE solutions, aerodynamic fields, and three-dimensional point-cloud segmentation outputs.

Paper: Functional Attention: From Pairwise Affinities to Functional Correspondences

Model Description

Functional Attention uses adaptive basis functions and functional mappings. This package trains and evaluates the model on AirfRANS aerodynamic data for PDE solving, 3D segmentation, and regression across discretizations and resolutions.

Intended Uses

Use case	Description
2D airfoil-flow surrogate	Predict pointwise velocity, pressure, and turbulent-viscosity fields on unstructured AirfRANS meshes.
Reynolds OOD evaluation	Use reynolds_train -> reynolds_test to measure out-of-distribution generalization across Reynolds numbers.
CFD surrogate acceleration	Approximate RANS simulations for fast field prediction and design screening.
Hugging Face/OneCode execution	Download the standalone model package, install its dependencies, and run the scripts directly.

Usage

1. OneCode

Launch the OneCode AI-for-Science environment

2. Manual Setup

Hardware requirements

• A GPU, DCU, or HCU is recommended for full training.
• A CPU can be used for import checks and very small connectivity tests.
• DCU users should install DTK 25.04.2 or later, or the OneScience-recommended version for the cluster.

Download the model repository from Hugging Face

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

Install the runtime environment

DCU environment

# Activate DTK first.
conda create -n onescience311 python=3.11 -y
conda activate onescience311
pip install onescience[cfd] -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/airfrans --repo-type dataset --local-dir ./data

Verify that the data path in config/config.yaml points to the downloaded dataset. The original data is also available from:

https://data.isir.upmc.fr/extrality/NeurIPS_2022/Dataset.zip

Train

The default command runs the Reynolds OOD experiment:

python scripts/train.py \
  --config config/config.yaml \
  --task reynolds

The checkpoint with the best validation Lv + Ls is saved to weight/best_model.pth. It contains the model, optimizer, learning-rate scheduler, training epoch, experiment configuration, and normalization statistics.

Inference

Run inference on three Reynolds OOD test cases:

python scripts/inference.py \
  --config config/config.yaml \
  --checkpoint weight/best_model.pth \
  --task reynolds \
  --max-cases 3

Reported metrics include relative L2 errors for four physical fields, surface-pressure relative L2, and explicitly labeled pressure_only_* lift and drag metrics.

Visualization

result.py reads the training history and inference .npz files and produces training curves plus ground-truth, prediction, and absolute-error fields:

python scripts/result.py \
  --config config/config.yaml \
  --task reynolds

Outputs are saved to:

results/figures/
results/visualization_manifest_reynolds.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: Functional Attention: From Pairwise Affinities to Functional Correspondences.
• Upstream code and weights: xjffff/FUNCATTN and JayXI/FuncAttn.
• This repository uses the Hugging Face-compatible MIT identifier (mit). AirfRANS and other third-party resources retain their own licenses and terms.