A curated dataset for data-driven turbulence modelling.

Area of Science:

• Computational Fluid Dynamics
• Machine Learning
• Turbulence Modeling

Background:

• Reynolds-averaged Navier-Stokes (RANS) models have limitations in accurately predicting turbulent flows.
• Machine learning (ML) offers a promising avenue to augment and improve RANS models.
• Existing datasets are not optimally structured for ML-augmented corrective turbulence closure modeling.

Purpose of the Study:

• To develop and release the first open-source dataset specifically for machine learning-augmented corrective turbulence closure modeling.
• To provide a structured resource for training, testing, and benchmarking new RANS models.
• To facilitate advancements in turbulence modeling through accessible data.

Main Methods:

• Curated a dataset comprising RANS simulations alongside matching Direct Numerical Simulation (DNS) and Large-Eddy Simulation (LES) data.
• Included data from four turbulence models: k-ε, k-ε-ϕt-f, k-ω, and k-ω SST.
• Compiled data across 29 cases for various DNS/LES reference cases, including periodic hills, square duct, and more, with 895,640 data points.

Main Results:

• The dataset contains extensive RANS features with DNS/LES labels at each data point.
• Feature set includes standard quantities and additional fields for novel feature generation.
• The dataset encompasses diverse flow scenarios to ensure broad applicability.

Conclusions:

• The released dataset significantly reduces the effort required for developing and validating ML-augmented RANS models.
• This resource is expected to accelerate innovation in turbulence modeling.
• The open-source nature promotes collaboration and reproducibility in the field.