OneScience-Sugon/cfdbench

CFDBench

Dataset Description

CFDBench is a large-scale benchmark dataset for machine learning methods in computational fluid dynamics, designed to evaluate the generalization capabilities of neural operators under unseen boundary conditions, fluid properties, and geometries.

The dataset contains four classic CFD problems: lid-driven cavity flow (cavity), laminar pipe flow (tube), step dam-break flow (dam), and flow around a cylinder (cylinder). For each problem, subsets are organized by variations in boundary conditions (bc), geometry (geo), and fluid properties (prop).

Paper: CFDBench: A Large-Scale Benchmark for Machine Learning Methods in Fluid Dynamics

Supported Tasks

Scenario	Description
Flow field time-series prediction	Predict subsequent flow states from historical velocity fields.
CFD surrogate modeling	Learn the mapping from physical parameters, boundary conditions, or geometry to velocity field evolution.
Generalization evaluation	Evaluate model performance under unseen boundary conditions, fluid properties, and geometries.
Neural operator research	Provide unified CFD evaluation data for models such as FNO and Transformer.

Dataset Format and Structure

The data is organized by problem type, subset, and case:

data/<problem>/<subset>/case*/

Each case contains a physical parameter file and a two-dimensional velocity field time series:

File	Format	Description
case.json	JSON	Fluid properties, boundary conditions, or geometric parameters.
u.npy	NumPy NPY	Horizontal velocity component with shape [T, 64, 64] and dtype float64.
v.npy	NumPy NPY	Vertical velocity component with the same shape as u.npy.

Some cases also provide u.png and v.png as preview images; they are not required training data.

How to Use the Dataset

This dataset is compatible with the OneScience-Sugon/CFDBench model. Download the dataset and model:

hf download --dataset OneScience-Sugon/cfdbench --local-dir ./data

Official OneScience Information

Platform	OneScience 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

• Luo, Y. et al. CFDBench: A Large-Scale Benchmark for Machine Learning Methods in Fluid Dynamics, 2023.
• This dataset is organized from the original CFDBench data and is licensed under the same Apache License 2.0.