BEAR-CFD-Dataset

This dataset is released with paper Data-driven operator learning for energy-efficient and indoor air quality-aware building ventilation control.

The data was generated using ANSYS FLUENT 2023R2 and includes both steady-state and transient flow simulations. It is designed to support research in scientific machine learning, particularly in learning neural operators and data-driven PDE solvers. This dataset is released for research purposes only. Please cite the authors if used in published work.

📚 BibTeX

@article{bian2025data,
  title={Data-driven operator learning for energy-efficient building control},
  author={Bian, Yuexin and Shi, Yuanyuan},
  journal={arXiv preprint arXiv:2504.21243},
  year={2025}
}

💻 Code Repository

All code used for data generation, preprocessing, and baseline models is available on GitHub:
🔗 GitHub Repository

🧪 Dataset Description

Steady-State Simulations

• 10 distinct cases were simulated, each with unique boundary conditions.
• Each simulation converges to a time-independent solution.
• These cases represent equilibrium airflow behavior under varying physical setups.

Transient Simulations

• 300 cases, each initialized from a corresponding steady-state solution.
• Each simulation spans 30 minutes of physical time.
• Simulations use a time step size of 10 seconds and output results every 30 seconds.
• This results in 60 time steps per case, capturing unsteady, time-varying airflow dynamics under changing boundary conditions.

Transient simulation data is used to train neural operators for modeling the evolution of fluid dynamics in time.

📂 Dataset Structure

The dataset repository contains the following subfolders:

├── raw_data/ # Transient simulation results in pickle format
├── processed_data/ # Processed datasets split into train/test sets
├── models/ # Trained neural operator models
├── steady_case_data/ # Original ANSYS steady-state result files

• Pickle Files (.pkl): Each transient simulation case is stored as a pickle file for convenient loading in Python.
• ANSYS Data: The steady-state simulation results are in native ANSYS formats (e.g., .cas.h5, .dat.h5).

📚 Use Cases

• Training neural operators to learn the temporal dynamics of fluid flow.
• Studying how steady-state solutions influence transient behavior.
• Benchmarking data-driven PDE solvers and physics-informed machine learning models.

📥 Loading Example

import pickle

with open("raw_data/unsteady_10.pkl", "rb") as f:
    data = pickle.load(f)

import ansys.fluent.core as pyfluent

solver = pyfluent.launch_fluent(mode="solver", precision="double",processor_count=6)
solver.file.read_case_data(file_name='steady_case_data/steady_seed0.cas.h5')
solver.settings.file.read_data(file_name = "steady_case_data/steady_seed0.dat.h5")