U-bend Parametric Thermal-Fluid Dataset
My rating
The U-bend dataset is a large parametric simulation set of a heat-exchanger core component (the bent pipe): 10,000 independent conjugate-heat-transfer runs, each described by 28 geometric design parameters. Every sample is published in three distinct representations - a parameter vector, five-resolution 2D images, and the cell values of the numerical mesh - so the same design can be studied from the parameter, image and graph/mesh viewpoints. What's included
- Geometry: 2D U-bend with a fluid region and a solid steel region, parameterised by 6 boundary points (2 degrees of freedom each) and 8 Bezier curves (2 parameters each) - 28 design parameters in total
- Targets: pressure loss between inlet and outlet in Pa, plus a cooling-performance figure in K2m2
- Images: shape.png plus U_fluid.png, p_fluid.png and T_fluid.png, each provided at five different resolutions
- Mesh cells: solid.pt and fluid.pt as PyTorch tensors carrying cell-centre coordinates and the solution variables
Specs
- Solver: OpenFOAM chtMultiRegionSimpleFoam (steady-state, buoyancy-driven, turbulent)
- Turbulence: two-equation k-omega SST
- Flow regime: Re = 40,000, Mach = 0.05 (incompressible assumption)
- Media: air in the fluid region, construction steel in the solid region
- License: CC BY-NC 4.0
Use cases: Heat-exchanger shape optimization via surrogate models; multimodal representation learning on parameters, images and meshes; transfer learning across coupled PDEs; teaching parametric CFD workflows.
Tags: