JDecke/ubend-cfd

U-bend Conjugate Heat Transfer Dataset

Parameterized U-bend CFD dataset for surrogate modeling and scientific machine learning.

Five randomly selected samples showing the geometry and the solution fields (velocity magnitude, pressure, temperature, turbulent kinetic energy, turbulent viscosity).

Dataset Description

This dataset contains ~9,000 conjugate heat transfer simulations of a parameterized U-bend cooling channel, generated with OpenFOAM (chtMultiRegionSimpleFoam). Each sample represents a unique channel geometry defined by 28 design parameters, with corresponding flow and temperature solution fields.

The dataset is designed for scientific machine learning tasks including:

• Parameter-to-objective regression: predict pressure loss and cooling effectiveness from geometry parameters
• Field prediction / Neural operators: predict full solution fields (velocity, pressure, temperature, turbulence quantities) from mesh coordinates
• Physics-informed neural networks (PINNs): mesh coordinates and boundary conditions available for physics-constrained training
• Surrogate modeling: fast approximation of expensive CFD simulations for design optimization

Configurations

params — Tabular parameter data

Lightweight CSV/Parquet with 28 input parameters and 2 scalar targets per sample.

from datasets import load_dataset
ds = load_dataset("JDecke/ubend-cfd", "params")

Column	Description
input_0 ... input_27	28 Sobol-sampled geometry parameters
target_0	Pressure loss (area-averaged inlet pressure)
target_1	Cooling effectiveness (integrated squared temperature deviation)
target_0_mean, target_0_std	Mean and std over last 100 solver iterations
target_1_mean, target_1_std	Mean and std over last 100 solver iterations

fields — Solution fields on the mesh

Per-sample safetensors files containing mesh coordinates and solution fields, cropped to the bend region (240 x 60 cells).

from safetensors.numpy import load_file
data = load_file("fields/sample_0.safetensors")

Key	Shape	Description
sample_id	(1,)	Sample index
params	(28,)	Geometry parameters
targets	(6,)	Target values (last, mean, std for both objectives)
coords	(3, 240, 60)	Cell center coordinates (x, y, z)
U	(3, 240, 60)	Velocity field (Ux, Uy, Uz)
p	(240, 60)	Pressure field
T	(240, 60)	Temperature field (fluid)
k	(240, 60)	Turbulent kinetic energy
nut	(240, 60)	Turbulent viscosity
p_rgh	(240, 60)	Pressure minus hydrostatic (equal to p for constant density)
solid_coords	(3, 240, 15)	Solid region cell centers
solid_T	(240, 15)	Solid temperature field

Simulation Setup

• Solver: chtMultiRegionSimpleFoam (run in the jdecke/foamy container)
• Turbulence: k-omega SST with wall functions
• Reynolds number: 40,000
• Fluid: Air, constant density (rho = 1.2 kg/m³)1
• Boundary conditions:
  • Inlet (fluid): fully-developed turbulent velocity profile, T = 300 K
  • Outlet (fluid): inletOutlet for U, p = 0, zeroGradient for T
  • Fluid/solid interface: thermally coupled (turbulentTemperatureCoupledBaffleMixed), noSlip
  • Heated wall: constant heat flux q = 100,000 W/m² on the outer solid wall in the U-bend (curved) section only
  • All other solid outer walls: adiabatic (zeroGradient)
  • Front/back: empty (2D simulation)
• Mesh: Structured 2D, 57,000 cells (45,600 fluid + 11,400 solid), wall grading ratio 10. The fluid region contains the flow channel (60 cells across), the solid region models the heated channel wall (15 cells across). Both regions are thermally coupled at the interface.
• Convergence: Residual-based termination (p_rgh < 1e-6), max 15,000 iterations

Parameterization

The geometry is controlled by 28 parameters sampled via a scrambled Sobol sequence (seed=42) to achieve space-filling parameter space. The 28 parameters comprise:

• 6 anchor point displacements (outer wall shape)
• 6 midpoint displacements (inner wall shape)
• 16 cubic Bezier control weights (wall curvature)

Dataset Statistics

• Total samples: 10,000
• Successful simulations: ~8,950 (89.5%)
• Mesh generation failures: ~990 (9.9%) — extreme parameter combinations producing invalid geometries
• Solver divergence: ~65 (0.7%) — physically unsteady flow in strongly constricted geometries
• Pressure loss range: 24 – 800 (successful samples)
• Temperature range: 3.7 – 10.0 (successful samples)

Failed samples and error codes

The params configuration provides three splits: train (≈8,050 samples) and test (≈895 samples) — a 90/10 split with seed=42 over successful simulations — plus failed (≈1,055 samples) containing all unsuccessful runs. Failed samples are indicated by sentinel values in the target columns:

target_0 value	Meaning
10000	Mesh generation failed (checkMesh reported errors) — geometry too extreme for valid mesh
11000	Simulation diverged (final residual >= 1e-4) — flow is physically unsteady or geometry too constricted
other	Actual pressure loss value (successful simulation)

The fields configuration contains only successful samples (~8,950) since no solution fields are available for failed cases. It uses the same train/test split as params.

Note that the failed samples are not randomly distributed in parameter space — they cluster in regions with extreme parameter combinations (sharp bends, strong constrictions). This means the parameter space has gaps where no valid simulation results exist. Users training surrogate models should be aware of this and filter out samples with target_0 >= 10000.

Mesh Quality

The computational mesh was selected based on a mesh convergence study across multiple wall grading configurations. The final mesh provides a good balance between numerical accuracy and simulation stability. Convergence quality is confirmed by recording the mean and standard deviation of each target over the last 100 solver iterations (median relative std < 0.05%).

Quick Start

Parameter regression

from datasets import load_dataset

ds = load_dataset("JDecke/ubend-cfd", "params")
print(ds["train"][0])     # ~8,050 training samples
print(ds["test"][0])      # ~895 test samples (90/10 split, seed=42)
# ds["failed"]            # samples with error codes (mesh/divergence)

Field prediction with PyTorch

from torch.utils.data import Dataset, DataLoader
from safetensors.numpy import load_file
from pathlib import Path
import torch

class UBendFieldDataset(Dataset):
    def __init__(self, data_dir):
        self.files = sorted(Path(data_dir).glob("*.safetensors"))

    def __len__(self):
        return len(self.files)

    def __getitem__(self, idx):
        data = load_file(str(self.files[idx]))
        return {
            "params": torch.from_numpy(data["params"]),              # (28,)
            # Fluid region
            "coords": torch.from_numpy(data["coords"]),             # (3, 240, 60)
            "U": torch.from_numpy(data["U"]),                        # (3, 240, 60)
            "p": torch.from_numpy(data["p"]),                        # (240, 60)
            "T": torch.from_numpy(data["T"]),                        # (240, 60)
            "k": torch.from_numpy(data["k"]),                        # (240, 60)
            "nut": torch.from_numpy(data["nut"]),                    # (240, 60)
            # Solid region
            "solid_coords": torch.from_numpy(data["solid_coords"]),  # (3, 240, 15)
            "solid_T": torch.from_numpy(data["solid_T"]),            # (240, 15)
        }

train_ds = UBendFieldDataset("fields/train")
test_ds = UBendFieldDataset("fields/test")
train_loader = DataLoader(train_ds, batch_size=16, shuffle=True, num_workers=4)
test_loader = DataLoader(test_ds, batch_size=16, shuffle=False, num_workers=4)

Visualization

A ready-to-use visualization script (visualize_sample.py) is included. It downloads N random samples from the hub and renders the geometry and all solution fields as PNGs:

python visualize_sample.py --n 5

This produces individual plots per sample plus an overview.png (as shown above).

Version History

This is an updated version of the original U-bend dataset with improved mesh quality (mesh convergence study), corrected Sobol sampling, and additional output fields (mean/std of targets, solution fields in safetensors format). The underlying geometry parameterization and simulation setup remain the same.

Citation

If you use this dataset, please cite the original publication:

@article{decke2023dataset,
  title={Dataset of a parameterized U-bend flow for Deep Learning Applications},
  author={Decke, Jens and W{\"u}nsch, Olaf and Sick, Bernhard},
  booktitle={Data in Brief},
  volume={50},
  year={2023},
  publisher={Elsevier}
}

Authors

Jens Decke, Olaf Wunsch, Bernhard Sick

Acknowledgments

This research has been funded by the Federal Ministry for Economic Affairs and Climate Action (BMWK) within the project "KI-basierte Topologieoptimierung elektrischer Maschinen (KITE)" (19I21034C).