Dataset of a parameterized U-bend flow for Deep Learning Applications

Jens Decke, Olaf Wünsch, Bernhard Sick

Abstract

Researchers from the University of Kassel have released a massive open-access dataset containing 10,000 high-fidelity fluid flow and heat transfer simulations of U-bend geometries. By providing three distinct data representations—design parameters, 2D images, and raw numerical mesh values—this benchmark enables the development of deep learning models for rapid design optimization in complex multiphysics environments.

Introduction

In engineering, designing components like heat exchangers or cooling ducts often requires Computational Fluid Dynamics (CFD). While CFD is highly accurate, it is computationally "expensive," often taking hours or days for a single simulation. Deep Learning (DL) offers a potential shortcut: if a model can learn the relationship between a shape and its performance, it can evaluate new designs in milliseconds.

However, DL is data-hungry. Most CFD studies focus on a handful of designs, and large-scale, standardized datasets for fluid-thermal systems are rare. This paper fills that gap by providing a comprehensive, labeled dataset specifically designed to train and benchmark DL models for shape optimization and multiphysics problems.

Methodology

The authors developed an automated pipeline to generate 10,000 unique U-bend designs. Each design is defined by 28 parameters that control the position of boundary points and the curvature of third-order Bézier curves forming the walls of the U-bend.

Parameterization of the U-bend geometry showing design variables and boundary conditions.

Physics Simulation and Workflow

The team used OpenFOAM, an open-source CFD toolbox, specifically the chtMultiRegionSimpleFoam solver. This handles "conjugate heat transfer," meaning it simulates how heat moves through both the solid walls (steel) and the fluid (air). To ensure reproducibility and scale, the simulations were run in Podman containers on a high-performance computing cluster, managed by the Slurm workload manager.

The pipeline included automated "Mesh Control" and "Residual Control" steps to discard simulations where the geometry was physically impossible or the numerical solution failed to converge.

Dataset Description

The primary contribution is the U-bend flow dataset, consisting of 10,000 samples. Each sample is organized into a structured hierarchy containing parameters, images, and raw mesh data.

The hierarchical structure of the dataset, showing the organization of parameters, images, and mesh-based cell data for each sample.
  1. Parameters: A vector of 28 design variables and two objective values (pressure loss and cooling performance).
  2. Images: 2D images of the geometry and solution variables (velocity, pressure, temperature) at five different resolutions.
  3. Cells: Raw data mapped to the numerical mesh centers, saved as PyTorch tensors. This allows researchers to use Graph Neural Networks (GNNs) or other architectures that respect the underlying physics mesh.

Analysis and Insights

  • Multimodal Versatility: The most innovative aspect is the triple-representation of data. By providing parameters, images, and mesh-based tensors, the dataset supports a wide variety of DL architectures, from simple Multi-Layer Perceptrons (MLPs) and Convolutional Neural Networks (CNNs) to advanced Geometric Deep Learning.
  • The Physics Trade-off: The dataset captures a classic engineering conflict. Minimizing pressure loss ($J_1$) usually requires lower flow speeds, while maximizing cooling ($J_2$) benefits from high speeds and turbulence. These objectives are "mutually exclusive," making this an excellent benchmark for multi-objective optimization.
  • Real-world Complexity: Unlike "toy" datasets, this involves turbulent flow (Reynolds number of 40,000) and conjugate heat transfer, reflecting the complexity of actual industrial cooling systems.

Results and Availability

A total of 10,000 independent and identically distributed (i.i.d.) samples were successfully generated. Individual simulations took between 15 and 240 minutes on AMD EPYC processors; providing this data pre-computed saves the community thousands of CPU hours. The dataset is licensed under CC-BY-NC-4.0 and is publicly available via the University of Kassel’s repository (DaKS) and Zenodo.

Limitations and Future Directions

To keep data generation feasible, the authors simplified a 3D problem into 2D. While 2D captures core physics, real-world ducts have 3D effects (like secondary flows) that aren't present here. Additionally, material properties for the fluid (air) and solid (steel) were kept constant. Future iterations could vary these properties to train even more generalized models. The authors suggest this data is ripe for research into transfer learning, active learning, and anomaly detection.

Summary updated

BibTeX
@inproceedings{decke2023dataset,
  author = {Jens Decke and Olaf Wünsch and Bernhard Sick},
  title = {{Dataset of a parameterized U-bend flow for Deep Learning Applications}},
  booktitle = {Data},
  year = {2023},
  doi = {10.1016/j.dib.2023.109477},
}
Figure