betaFeedback

WorksPhysics and Astronomy

A multi-fidelity dataset of two-dimensional steady laminar channel flow around obstacles

Abstract

Surrogates that correct inexpensive flow simulations need the same flows solved at several fidelities and an estimate of the numerical error of the reference solutions. This dataset contains 4,500 two-dimensional, steady, laminar channel flows past a single obstacle. Each is solved in five numerical configurations: second-order convection on meshes with cell sizes of 2, 1, 0.5 and 0.25 mm, and first-order convection on the 1 mm mesh. A seeded generator draws the obstacles from five shape families in three size classes. All 22,500 OpenFOAM runs were iterated to initial residuals of 10−610^{-6}, and each run's residuals, iterations and provenance are recorded. Every run provides the velocity and pressure on a regular grid with geometry channels, and the pressure and wall shear along the obstacle. Its 20 flow attributes, computed on the simulation mesh, include drag and lift split into pressure and friction parts. The four second-order levels give an estimate of the discretisation uncertainty of 19 attributes of every case. Fixed splits include three out-of-distribution test sets.

Background & Summary

Computational fluid dynamics (CFD) lets engineers test a design before they build a prototype. Every change to the design needs a new simulation. A fine-mesh simulation can take hours, and a coarse-mesh one takes minutes but is less accurate. Machine-learning surrogates for CFD predict a flow without a new simulation. They are trained on solved flows, and many public datasets provide such flows[1, 2, 3, 4, 5, 6, 7, 8, 9]. A surrogate can also correct a cheap solution, for example by predicting the fine-mesh flow or the drag from a coarse-mesh solution, and so shorten the design cycle. Training such a surrogate needs the same flows solved at several fidelities. These flows must cover geometries that vary enough for the surrogate to learn how the flow interacts with the obstacle. The surrogate’s accuracy can only be assessed with an estimate of the numerical error of its reference solutions, because surrogates inherit the discretisation error of their training data[10]. In a study of wind-turbine wakes, networks trained on data with a larger numerical uncertainty were less accurate[11]. If a surrogate is more accurate than its labels, only a comparison with a finer reference can show it[12].

Among the datasets compared here (Table 1), most offer one resolution. In FlowBench[3], the levels are sub-sampled from one simulation. In a double-delta wing set, one level is computed with a cheaper vortex-lattice model[13]. MegaFlow2D, like this dataset, solves each of its two levels on its own mesh, but its flows are transient[14]. The RANS study of Sætnik et al.[15] has two levels that differ in the near-wall treatment. A later set from the same group has eleven meshes for each of 1,627 airfoils, and every run was stopped after a fixed number of iterations[16]. For chaotic turbulence, BLASTNet notes that matching pairs of solved levels cannot be obtained, and it filters one solution instead[17]. Steady laminar flow has no such limit, so each level can be a converged solution on its own mesh rather than a filtered copy of the finest one.

DatasetFlowCasesFidelity levelsStopping criterionLabel uncertaintyPhysical targetsRaw cases
This dataset2D laminar, channel with one obstacle4,5005 configurations: 4 meshes (2–0.25 mm) and first order at 1 mm, all convergedper run, <10−6<10^{-6}, released and auditedestimated for 19 attributes per case20 on the simulation mesh, forces split; surface pp, τw\tau_wall 22,500 runs
DeepCFD[18]2D laminar, channel with one obstacle9811n.r.n.r.noneno
AirfRANS[1]2D RANS, airfoils1,0001force convergencen.r.CDC_D, CLC_L, boundary layer forcesyes
Setinek et al. 2025[15]2D RANS, airfoils611 pairs2 (wall-resolved, wall functions)n.r.n.r.forcesn.r.
Setinek et al. 2026[16]2D RANS, airfoils1,62711 meshes (5k–212k nodes), solved independentlyfixed 2×1042 \times10^{4} or 4×1044 \times10^{4} iterationsn.r. (errors against the finest mesh)CDC_D, CLC_Lno
UniFoil[8]2D RANS, airfoils, turbulent and transitional>500>500k1 mesh; two turbulence models on 4,800 airfoilsn.r.; historyn.r. (mesh study “impractical”)CDC_D, CLC_L, surface fieldsyes
SuperWing[9]3D RANS, transonic wings28,8561residual <10−10<10^{-10}, steady CLC_Lreference wings onlyCDC_D, CLC_L, CMC_M, surface CpC_p, CfC_fno
Double-delta wings[13]3D RANS, wings2,4482 (vortex lattice, RANS)continuity 10−610^{-6}, steady CDC_D; historyone geometryCDC_D, CLC_L, CMC_M, split in raw filessolver output
DrivAerNet++[2]3D RANS, cars8,00017,000 iterations, RMS <10−5<10^{-5}one reference carCDC_D, CLC_L, surface pp, τw\tau_wyes
FlowBench[3] (steady)2D/3D cavity with an object≈9,000\approx9,000 (2D)3, sub-sampled from one solutionn.r.reference cases onlyCDC_D, CLC_L, Nusseltno
MegaFlow2D[14]2D transient, circles, ellipses, a nozzle3,0002, solved independently– (transient)mesh chosen oncenoneno (graph snapshots)

Table 1. This dataset and related steady or multi-resolution datasets. Flows are steady unless marked transient. Fidelity levels: number of levels and how they differ. Label uncertainty: estimate of the numerical error of the labels, or the scope of the mesh study. Raw cases: whether the solver cases are released. RANS, Reynolds-averaged Navier–Stokes; CDC_D, CLC_L, CMC_M, drag, lift and moment coefficients; CpC_p, CfC_f, pressure and skin-friction coefficients; pp, pressure; τw\tau_w, wall shear stress; RMS, root mean square; k, thousand; history, solver history of every run released; n.r., not reported.

The iterative solution converges slowly on fine meshes, so the stopping criterion decides how close a fine-mesh label is to its steady solution. Where the other datasets give a criterion, it applies to the dataset as a whole (Table 1). The criteria given are a residual target between 10−510^{-5} and 10−1010^{-10} (sometimes together with steady forces), a fixed number of iterations, or the convergence of the forces. In the double-delta set, 57 of 2,448 runs reached the target only with a lower-order scheme [13]. UniFoil and the double-delta set release the solver history of every run, but none of the datasets compared shows how a looser criterion would change the labels. In a study reported in Technical Validation, 113 cases were stopped at residuals of 10−310^{-3} (velocity) and 10−210^{-2} (pressure). When the runs start from rest, the 0.25 mm labels lose about a third of their recirculation length. If the coarse level also stops early, the coarse-to-fine difference is overstated by a factor of 36. When the runs start from a coarse solution instead, the same criterion leaves the labels close to that solution, and the difference is understated.

Mesh studies are usually made once, to choose the mesh[13, 14], or on a few reference cases [2, 3, 9]. UniFoil judged a mesh study impractical for its half a million runs [8]. The eleven-mesh airfoil set gives the error of every level against the finest one, which is taken as the reference [16]. None of the datasets compared estimates the numerical uncertainty of each label. In a coarse-to-fine task, the error of a cheap solution and any improvement on it have to be read against the uncertainty of the label used for comparison. Where the steady datasets compared give forces, these are totals. Only the raw solver output of the double-delta set splits them into pressure and friction parts [13].

This dataset follows the setting of DeepCFD[18]: two-dimensional steady laminar flow past one obstacle in a 120 mm high channel. The inlet velocity and viscosity are the same, and the five obstacle families are similar. DeepCFD offers fields only, at one resolution of about 1 mm, and reports neither a stopping criterion nor a mesh study.

MS-COFD-2DSL (Multi-Scale Central Object Flow Dataset, 2D Steady Laminar) provides 4,500 obstacle geometries in five families and three size classes. The obstacles are rotated and shifted, and their blockage is limited. Each geometry is solved to steady state in five numerical configurations (Table 2): with second-order convection on four meshes with cell sizes of 2, 1, 0.5 and 0.25 mm, and with first-order convection on the 1 mm mesh. Each of the 22,500 runs was iterated to initial residuals of 10−610^{-6}, and its residuals, iterations, cost and provenance are released with it. The four second-order levels form a refinement study of every case, from which the discretisation uncertainty of 19 scalar attributes is estimated. Each run provides the fields on a regular grid (180 × 60 to 1440 × 480 points) with geometry channels, and the pressure and wall shear along the obstacle. It also provides 20 flow attributes computed on the simulation mesh. Fixed splits and three out-of-distribution test sets (blockage, rotation, position) are part of the release, so that results can be compared. The solver setup was checked against the DFG 2D-1 benchmark, and every run was audited. Runs that were processed or solved again reproduced the released data. A separate repository defines tasks, metrics and baselines on these data and holds the uncertainty estimates of every case [19].

RunCell sizeConvectionStarted fromCellsArray gridIterations to 10−610^{-6}Solver time
0.25x2 mmsecond orderrest110,889180×60180 \times60209 (1,094)8 s
base_fo1 mmfirst orderrest43,121360×120360 \times120491 (1,426)77 s
base_lu1 mmsecond orderrest43,121360×120360 \times120587 (2,377)110 s
4x0.5 mmsecond orderbase_lu171,555720×240720 \times2401,075 (3,501)19 min
16x0.25 mmsecond orderbase_lu683,6811440×4801440 \times4803,497 (11,109)5.6 h

Table 2. The five runs of every case. Run: the folder name in the release. Cell counts, iterations and solver times are medians over the 4,500 cases, with the 95th percentile of the iterations in brackets. Solver times are for one core. a^{\mathrm{a}}The 27 relaxed 2 mm runs started from the converged base_lu solution (Flow solution).

The data can be used for coarse-to-fine correction and super-resolution with paired, converged levels, and for surrogate models of fields and of physical quantities. They also allow multi-fidelity learning across mesh size and discretisation order, and studies of how the stopping criterion and the mesh affect learned models.

Methods

Overview

Each case is produced by one automated pipeline [20] (Figure 1b). The obstacle is drawn from a seeded random generator and meshed with gmsh. The flow is then solved to steady state with OpenFOAM in the five configurations of Table 2, and each run is iterated until the initial residuals of velocity and pressure are below 10−610^{-6}. The converged fields are resampled to regular grids, and the flow attributes are computed from OpenFOAM output on the simulation mesh. The complete case is then archived.

Figure 1. Flow configuration and the five runs of every case. a, The channel, to scale, with its boundary conditions. The obstacle keeps clear of the shaded zones, and its centre is shifted within the dashed range around the nominal centre (+). The obstacle shown is the normal right triangle of Methods, Arrays. b, The meshes of a case (median cell counts), its five runs and the fields from which each run starts.

The 0.25 mm run provides the labels. The 2 mm run and the two 1 mm runs are low-fidelity inputs. The 2 mm run is coarse but second order, whereas the first-order 1 mm run is as fine as the second-order 1 mm run but more diffusive. The 0.5 mm run is an intermediate refinement, so that the meshes can be compared in cost and accuracy. The four second-order levels form a refinement study of every case, from which the discretisation uncertainty of 19 scalar attributes is estimated. The two 1 mm runs differ only in the convection scheme, so they separate its effect from that of the mesh. Fidelity here means mesh resolution and convection order. The configurations are not ordered by accuracy for every quantity. For example, the first-order 1 mm run is closer to the label than the 2 mm run in pressure drag but further from it in total drag (Technical Validation). In what follows, a configuration is one of the five numerical set-ups, a run is the solution of one configuration for one case, and the levels are the four second-order meshes.

The runs have the same names as the folders of the release (Table 2). The 1 mm mesh is the base mesh. On it, base_lu is solved with second-order linearUpwind convection and base_fo with first-order upwind convection. The runs 0.25x, 4x and 16x use the other meshes. They are named after their number of cells relative to the base mesh, which is about a quarter, four times and sixteen times as large.

Design choices

The flow is steady and two-dimensional, so every configuration can be iterated to convergence on its own mesh, and the four-level refinement could be afforded for 4,500 geometries. Even in two dimensions, a 0.25 mm run takes 5.6 core-hours (median, Table 2). The resampled fields also lie on regular grids of the kind coarse-to-fine models use. The inlet conditions are fixed, so the Reynolds number varies only through the obstacle height (Flow configuration). With the inlet velocity and viscosity fixed, the effect of the geometry and of the numerical configuration is isolated. At this inlet velocity, the steady solver converged up to the blockage limit (Geometry). The data do not show how models transfer to other Reynolds numbers, to unsteady flow or to three dimensions.

The five shape families are modelled on those of DeepCFD. Rectangles, diamonds and triangles have sharp corners at which the flow separates, and the spline “circles” have rounded outlines. The two triangle families, one pointing upstream and one downstream, add an asymmetry. The rotation is limited to 45∘45^\circ to keep each family recognisable. The mesh is refined uniformly because the uncertainty estimates of a case are based on four second-order levels of the same type of mesh. As a result, the near-wall region is resolved no more finely than the rest of the domain. Locally refined meshes are left to a later version. In a preliminary study of 30 geometries, the second-order 1 mm solution was already within 1–3% of the mesh-independent value for almost every integral attribute. The 2 mm level was added because a coarse-to-fine task needs a coarser input than this. The 0.25 mm level was the finest that could be afforded for all cases (Computation and provenance).

The first-order 1 mm run was cheap to add (77 s per case, median). Attribute errors complement field errors, since a field error can be small while the flow near the obstacle or in the wake is wrong. The pressure and friction parts of the drag also show errors that cancel in the total. The out-of-distribution sets (Splits) test a mild extrapolation in blockage, rotation or position. They contain no new shape families or flow conditions.

Flow configuration

The domain is a channel of length L=360 mmL = 360\,\mathrm{mm} and height H=120 mmH = 120\,\mathrm{mm}, with the inlet at x=0x = 0 (Figure 1a). The fluid is incompressible and Newtonian, with kinematic viscosity ν=1×10−4 m2 s−1\nu= 1 \times10^{-4}\,\mathrm{m^2\,s^{-1}}. The steady equations ∇⋅u=0\nabla\cdot\mathbf{u} = 0 and (u⋅∇)u=−∇p+ν∇2u(\mathbf{u} \cdot\nabla)\mathbf{u} = -\nabla p + \nu\nabla^2\mathbf{u} are solved for the velocity u\mathbf{u} and the kinematic pressure pp. At the inlet, the velocity is uniform, Uin=0.1 m s−1U_{\mathrm{in}} = 0.1\,\mathrm{m\,s^{-1}}, and the pressure has zero gradient. At the outlet, the velocity has zero gradient and p=0p = 0. On the channel walls and the obstacle, the velocity is zero (no-slip) and the pressure has zero gradient. The Reynolds number is 120 based on the channel height, and Reh=Uinh/ν=10Re_h = U_{\mathrm{in}}h/\nu= 10–48 based on the obstacle height hh, the height of its bounding box (h/Hh/H is the blockage ratio). The inlet velocity is the same in all cases, while the shape, size and placement of the obstacle vary (Geometry). OpenFOAM meshes are three-dimensional, so the mesh is one cell layer deep in zz, normal to the plane of the flow. The layer is 10 mm thick, and its front and back faces are empty patches, which makes the solution two-dimensional. Forces are reported for this depth. All fields are kinematic.

Geometry

Five obstacle families are used (Figure 2): rectangles, diamonds, left triangles, right triangles and “circles”. The apex of a left triangle points upstream, and that of a right triangle points downstream. The “circles” are closed curves of four splines through perturbed points, which give a rounded but irregular shape. The obstacles come in three size classes (smaller, normal and larger), with 300 cases for every family and class. Each case is drawn with its own random generator, seeded with the string “seed-size-family-index” (global seed 2026), so that any case can be regenerated on its own. Every mesh level of a case shares the same geometry.

Figure 2. Obstacle families and size classes. Columns show the five families and rows the three size classes. Each panel shows the first 25 cases of its family and class as grey outlines, with the first one filled, on the part of the channel between x=45x = 45 and 205 mm. The “circle” family consists of closed curves of four splines through perturbed points. These outlines are rounded rather than circular, but the family keeps its name in the files.

A case is drawn in four steps. First, the width WW and the height are the nominal size of the family (rectangle and diamond 50×25 mm50 \times25\,\mathrm{mm}, triangles 42.5×42.5 mm42.5 \times42.5\,\mathrm{mm}, circle 25×25 mm25 \times25\,\mathrm{mm}) times the scale of the size class (0.65, 1 or 1.3). The width is further multiplied by (1+uw)(1 + u_w) and the height by (1+uh)(1 + u_h), with uw∼U(−0.10,0.10)u_w \sim\mathcal{U}(-0.10, 0.10) and uh∼U(−0.20,0.20)u_h \sim\mathcal{U}(-0.20, 0.20). Second, the corner points are placed around the nominal centre (130, 60) mm, and each is moved by a random offset of up to 0.17W0.17W (0.14W0.14W for circles, 0.085W0.085W for diamonds). A triangle is taken from such a perturbed rectangle, with the apex at the midpoint of its two upstream or downstream corners. A circle is closed by four interpolating splines through the corners and one perturbed point near the middle of each side. The exact construction is in the pipeline code [20]. Third, the obstacle is rotated about its centroid by θ∼U(−45∘,45∘)\theta\sim\mathcal{U}(-45^\circ, 45^\circ) and shifted by Δx∼U(−30,10) mm\Delta x \sim\mathcal{U}(-30, 10)\,\mathrm{mm} and Δy∼U(−15,15) mm\Delta y \sim\mathcal{U}(-15, 15)\,\mathrm{mm}.

The shift is biased upstream to leave room for the wake. Fourth, the obstacle is drawn again if it comes closer than 5 mm to a channel wall, 60 mm to the inlet or 150 mm to the outlet. It is also drawn again if its outline crosses itself or if its blockage ratio (bounding-box height over HH) exceeds 0.40. The number of attempts is recorded.

The blockage limit was set in pilot runs of the production pipeline. Above a blockage of about 0.45, a growing share of obstacles did not converge on the 1 mm mesh, and their residuals stalled at 10−310^{-3}–10−210^{-2}. At 0.40, almost every obstacle converged. The stall may indicate the onset of unsteady flow in a confined wake, but this was not tested with transient simulations. In the dataset, the blockage ratio ranges from 0.08 to 0.40 (median 0.28). A case whose 1 mm second-order run did not reach 10−610^{-6} within its iteration cap was dropped with all its runs. It was replaced by the next reserve geometry of its size and family (indices 301–330, drawn from the same seed). No other run of a case started before its 1 mm second-order run had converged, so every case in the dataset has all five configurations. Eighteen cases were dropped and replaced: 14 normal left triangles, 3 normal right triangles and 1 larger left triangle. All had a blockage of 0.36–0.40 and largest residuals of 10−410^{-4} to 2×10−22 \times10^{-2} at the 10,000-iteration cap. Dropping them removes part of the highest-blockage normal left triangles from the sampled distribution. A dropped case may still have a steady solution, since other stalled runs converged once their starting fields and solver relaxation were changed (see Flow solution). The dataset is therefore conditioned on convergence under the production protocol.

Meshing

Each geometry is written as a gmsh geometry file and meshed with gmsh 4.15.2 [21], using the “Delaunay for quads” algorithm with recombination into a quad-dominant mesh. Every level has one uniform characteristic cell size (2, 1, 0.5 or 0.25 mm) and no local refinement (Design choices). The mesh is extruded by one 10 mm layer and converted with gmshToFoam. It is accepted when checkMesh passes all its default checks (“Mesh OK”). A mesh that failed was generated again with additional smoothing (Mesh.Smoothing = 5). This was needed for 3 meshes at 2 mm, 3 in each 1 mm run, 5 at 0.5 mm and 14 at 0.25 mm. The mesh variant is recorded for every run. Meshing ran on one thread, so the meshes are deterministic.

Flow solution

The steady incompressible Navier–Stokes equations are solved with simpleFoam of OpenFOAM v2512 (ESI)[22] using the SIMPLEC algorithm[23]. Convection is discretised with linearUpwind (second order) in all runs except base_fo, which uses first-order upwind. Each run is set up from the OpenFOAM case template of its configuration. Gradients use the Gauss linear scheme and Laplacians the Gauss linear corrected scheme, with one non-orthogonal corrector. Pressure is solved with GAMG and velocity with symmetric Gauss–Seidel, and the velocity is under-relaxed with a factor of 0.9.

Every run is iterated until the initial residuals of uxu_x, uyu_y and pp are all below 10−610^{-6}. The initial residual of a discretised equation Aψ=bA\psi=b is computed at the start of an iteration, before its linear solve, with the default normalisation of OpenFOAM v2512[24]:

r=∑∣b−Aψ∣∑(∣Aψ−Aψˉ∣+∣b−Aψˉ∣)r = \frac{\sum\lvert b-A\psi\rvert}{\sum\left(\lvert A\psi-A\bar{\psi}\rvert+\lvert b-A\bar{\psi}\rvert\right)}

where the sums run over the cells and ψˉ\bar{\psi} is the field’s mean. This residual measures how far the fields are from satisfying the discretised equations, rather than how much they change between iterations. In every template, the linear solves within each iteration stop at r<10−8r<10^{-8} or once rr has fallen by a factor of 100 (pressure) or 10 (velocity). These solver tolerances do not decide when a run ends. The largest final residual in the dataset is 9.99999×10−79.99999\times10^{-7}. A run that reached its iteration cap with its residuals still falling was continued from its last time step until it converged. The longest runs needed 40,716 iterations at 2 mm, 3,893 and 16,579 at 1 mm (first and second order), 15,647 at 0.5 mm and 52,748 at 0.25 mm. The iterations, cap and number of solver runs of every run are in the metadata. At 2 mm, 27 runs whose residuals had stalled between 10−310^{-3} and 3×10−23\times10^{-2} were instead started again from their converged 1 mm solution. They used SIMPLE and under-relaxation (pressure 0.3, velocity 0.7) in a template of their own, the relaxed template, but otherwise identical settings and discretisation. These relaxed 2 mm runs converged after 575–7,652 iterations.

Apart from the 27 relaxed 2 mm runs, the 2 mm and both 1 mm runs start from rest. The 0.5 and 0.25 mm runs start from the converged 1 mm second-order solution, mapped onto their mesh with mapFields. The effect of this start is tested in Technical Validation.

Arrays

The converged fields of every run are resampled to a regular grid with VTK 9.7.0[25]. The reader vtkOpenFOAMReader interpolates the cell values of u\mathbf{u} and pp to the mesh points. The filter vtkResampleToImage then samples them on nx×nyn_x\times n_y points that cover the whole domain, both ends included, with about one point per cell size (Table 2). ParaView’s OpenFOAM reader and its Resample To Image filter are based on them. Points inside the obstacle lie outside the mesh and are set to zero. Each sample has six channels (Table 3). Three are the flow fields uxu_x, uyu_y and pp. The other three are geometry channels made from the same resampling: the distance to the nearer channel wall, a cell-type mask and the signed distance to the obstacle outline. The wall distance is min⁡(j,ny−1−j)\min(j,n_y-1-j) for grid row jj, so both wall rows have 0. The cell type is 1 on the two channel walls and at the obstacle points, 2 on the inlet column, 3 on the outlet column and 4 elsewhere. The end points of the inlet and outlet columns take the inlet and outlet values, respectively. The signed distance is the Euclidean distance, in grid indices, to the nearest obstacle point with at least one of its four neighbours outside the obstacle. It is zero on these points and negative inside the obstacle. Because the grid spacings in xx and yy differ slightly (2.011 and 2.034 mm on the 2 mm grid), it is not an exact physical distance. Figure 3 shows one case in all five configurations. At its scale, the 0.5 and 0.25 mm fields are hard to tell apart by eye. Their attributes still differ, for example by 0.64% in the recirculation length (median, Technical Validation).

Figure 3. One case in all five configurations. Top: streamwise velocity uxu_x of a normal right triangle, a case chosen at random, on the 0.25 mm grid over the whole channel. Blue is reverse flow. Middle: the dashed window on every grid, as stored (no smoothing), with the exact outline and the recirculation length computed on each mesh. Bottom: ∣ux−ux0.25∣\lvert u_x-u_x^{0.25}\rvert on the grid of every cheaper configuration, in % of UinU_{\mathrm{in}} on a logarithmic scale. The 0.25 mm array is interpolated bilinearly to the points of each of these grids, and points in the obstacle, or whose interpolation touches it, are masked.

ChannelNameContentUnit
0u_xstreamwise velocitym s−1^{-1}
1u_ycross-stream velocitym s−1^{-1}
2pkinematic pressure, zero at the outletm2^2 s−2^{-2}
3wall_distancedistance to the nearer channel wallgrid points
4cell_type1 wall or obstacle, 2 inlet, 3 outlet, 4 fluid–
5obstacle_sdfsigned distance to the obstacle outline, negative insidegrid points

Table 3. Channels of every sample. Points inside the obstacle are zero in channels 0–2.

Attributes

The flow attributes of every run (Table 4) are computed on a copy of the converged case on the simulation mesh, where the obstacle is a patch of its own. OpenFOAM function objects and utilities write the forces, the pressure and wall shear on the obstacle faces, the cell fields and lines sampled across the channel. A Python script computes the attributes from them[20]. The attributes, each defined in Table 4, fall into four groups, and the outlet reverse flow checks the outlet condition. The forces on the obstacle are drag and lift, each with its pressure and viscous (friction) parts. The second group is taken from the pressure and wall shear on the obstacle surface. The pressure drop along the channel and the vorticity and dissipation describe the flow in the channel. The recirculation zone, the flow split at the obstacle and a velocity profile behind it describe the shape of the wake. The recirculation zone is the reverse flow (ux<0u_x < 0) connected to the obstacle, so that a separate thin reverse-flow layer along a channel wall does not count. The vorticity and the dissipation leave out the first 5 mm of the channel, which contain the singular inlet corners. Pressure and wall shear are also stored per obstacle face for every run.

AttributeColumnUnitNote
drag: total, pressure, viscousFx, Fx_pressure, Fx_viscousm4 s−2\mathrm{m}^{4}\,\mathrm{s}^{-2} per 10 mm
lift: total, pressure, viscousFy, Fy_pressure, Fy_viscousm4 s−2\mathrm{m}^{4}\,\mathrm{s}^{-2} per 10 mm
front pressurefront_pressurem2 s−2\mathrm{m}^{2}\,\mathrm{s}^{-2}largest pp of the obstacle faces, not mesh-independent (stagnation point often on a corner)
base pressurebase_pressurem2 s−2\mathrm{m}^{2}\,\mathrm{s}^{-2}unweighted mean pp of the faces in the rearmost 10 % of the obstacle’s xx extent
mean wall shearmean_wall_shearm2 s−2\mathrm{m}^{2}\,\mathrm{s}^{-2}area-weighted ∣τw∣/ρ\lvert\tau_w\rvert/\rho on the obstacle
peak wall shearmax_wall_shearm2 s−2\mathrm{m}^{2}\,\mathrm{s}^{-2}largest face value, not mesh-independent (singular at corners)
pressure drop, inlet to outletdp_inlet_outletm2 s−2\mathrm{m}^{2}\,\mathrm{s}^{-2}difference of the area averages on the inlet and outlet patches
pressure drop, cross-sectionsdp_sectionsm2 s−2\mathrm{m}^{2}\,\mathrm{s}^{-2}the same on lines 10 mm behind the inlet and 10 mm before the outlet
mean vorticitymean_vorticitys−1\mathrm{s}^{-1}area-weighted mean ∣ω∣\lvert\omega\rvert, first 5 mm left out
peak vorticitypeak_vorticitys−1\mathrm{s}^{-1}largest cell ∣ω∣\lvert\omega\rvert, first 5 mm left out, not mesh-independent (singular at corners)
pseudo-dissipationdissipationm4 s−3\mathrm{m}^{4}\,\mathrm{s}^{-3} per metreν∫∣∇u∣2\nu\int\lvert\nabla\mathbf{u}\rvert^{2}, first 5 mm left out
recirculation lengthrecirculation_length_mmmmfurthest reverse-flow cell of the zone, from the downstream end of the obstacle; zero if the zone does not reach behind it
recirculation arearecirculation_area_mm2mm2\mathrm{mm}^{2}area of the reverse flow connected to the obstacle
flow splitflow_split_upper–flow rate above the obstacle over the total, with uxu_x integrated from each wall to the obstacle at the xx of its centroid
wake velocitywake_min_u_xm s−1\mathrm{m}\,\mathrm{s}^{-1}smallest uxu_x on a line across the channel 20 mm behind the rearmost outline point of the obstacle, within the obstacle’s yy extent
outlet reverse flowoutlet_reverse_flow–reverse flow rate in the last 5 mm over the net flow rate (area-weighted over cells), zero in every 0.25 mm run

Table 4. Attributes of every run, computed by a Python script from the output of OpenFOAM function objects and utilities on the simulation mesh (Methods, Attributes). Pressures and stresses are kinematic. Forces are for the 10 mm depth of the mesh (multiply by 100 for values per metre) and equal newtons for ρ=1 kg m−3\rho= 1\,\mathrm{kg}\,\mathrm{m}^{-3}. Dissipation is per metre of depth.

The column dissipation is the pseudo-dissipation ν∫∣∇u∣2 dA\nu\int\lvert\nabla\mathbf{u}\rvert^2 \,\mathrm{d}A. It differs by boundary terms from the viscous dissipation 2ν∫S:S dA2\nu\int\mathbf{S}:\mathbf{S}\,\mathrm{d}A, where S\mathbf{S} is the strain-rate tensor. In five runs checked, one of each configuration, the viscous dissipation over the region used here was 0.16–0.51 % lower, and 0.16 % lower in the 0.25 mm run. The outlet reverse flow is zero in every 0.25 mm run, so no reverse flow reaches the outlet, where the zero-gradient condition would otherwise prescribe the velocity of inflowing fluid. The sensitivity of the solutions to the outlet position was not tested.

Three attributes are included but are not mesh-independent. The peak wall shear and the peak vorticity grow with refinement at the sharp corners of the obstacles, where the exact solution is singular. The front pressure did not converge with the mesh in two thirds of the cases, because the stagnation point often sits on a corner.

Label uncertainty

Except for outlet_reverse_flow, which is zero in every 0.25 mm run, each scalar attribute of each case is computed at four second-order levels. The discretisation uncertainty U of its 0.25 mm value ϕ1\phi_1 is estimated with the least-squares procedure of Eça and Hoekstra [26]. The procedure is designed for refinement data that scatter and are not in the asymptotic range, as is the case for independent unstructured meshes. The step size of level ii is computed from the cell counts as hi=(N1/Ni)1/2h_i=(N_1/N_i)^{1/2}, where N1N_1 is the number of cells of the 0.25 mm mesh. A power law ϕ=ϕ0+ahp\phi=\phi_0+ah^p with an observed order pp, and laws with the fixed orders 1, 2 and 1+2, are fitted by least squares. Each is fitted both unweighted and with weights proportional to 1/h1/h, and the weighting whose power-law fit has the smaller standard deviation σ\sigma is used for all of them. The error estimate is e=∣ϕ1−ϕ0∣e=\lvert\phi_1-\phi_0\rvert, where ϕ0\phi_0 is the extrapolated value of the fit that the observed order selects. For 0.5≤p<2.10.5 \le p < 2.1, this is the power law, and the factor of safety is Fs=1.25F_s=1.25. For 0<p<0.50 < p < 0.5 or p≥2.1p \ge2.1, the fixed-order fit with the smallest σ\sigma is used instead, chosen from the orders 1, 2 and 1+2, or from 2 and 1+2 when p≥2.1p \ge2.1. Then ϕ0\phi_0, σ\sigma and the fitted curve ϕfit\phi_{\mathrm{fit}} are those of this fit, and Fs=3F_s=3. The uncertainty is U=Fse+σ+∣ϕ1−ϕfit(h1)∣U=F_se+\sigma+\lvert\phi_1-\phi_{\mathrm{fit}}(h_1)\rvert. When σ\sigma is at least Δϕ\Delta\phi, a third of the range of the four values, the uncertainty is instead U=(3σ/Δϕ)(e+σ+∣ϕ1−ϕfit(h1)∣)U=(3\sigma/\Delta\phi)(e+\sigma+\lvert\phi_1-\phi_{\mathrm{fit}}(h_1)\rvert). Where the four values do not approach a limit in a way the fits can describe (fitted order p≤0p \le0), U is three times the largest difference between them. This typically happens because the scatter between the independent meshes exceeds the trend. Since all runs converged iteratively, this behaviour concerns the convergence of an attribute with the mesh. The rule applies to 46 % of the attribute–case pairs, including 42 % of the drag values and 93–96 % of the two peak quantities, but none of the dissipation values. Where all four values are equal (217 pairs, all of them zero recirculation lengths and areas), U = 0 by convention, although their numerical error need not be zero.

U estimates the discretisation error of the 0.25 mm solution but does not bound it. For the three attributes that are not mesh-independent (Attributes), the values are not shown to approach a finite limit, so their U is not an uncertainty about a mesh-independent value. U does not include the loss in resampling to the arrays or the iteration error, which was small against U in the cases tested in Technical Validation. The uncertainty file (Usage Notes) gives U both in the units of the attribute and as a percentage of a scale. The scale is the value itself, except for lift, its parts and the parts of the drag. These are taken relative to the drag, since the lift is small and of either sign and the two drag parts shift in opposite directions with the mesh. The wake velocity is taken relative to UinU_{\mathrm{in}}, and the front and base pressure relative to the dynamic pressure 12Uin2\frac{1}{2}U_{\mathrm{in}}^2. The percentage is undefined in the 227 pairs whose 0.25 mm value is zero: 163 recirculation lengths, which are zero when no reverse flow reaches behind the obstacle, and 64 recirculation areas. The level errors in Technical Validation use the same scales, except for the recirculation length and area. Their scale is the larger of the value and the obstacle height (its square for the area), so that short wakes do not give large relative values. The scale differs from the value only in the cases whose zone is shorter than the obstacle height (or smaller in area than its square). The same procedure was applied to the eight attributes that are also computed from the arrays, at every level on its own grid, to measure how well the arrays represent them (Technical Validation). The three-level grid convergence index [27] is released as well, where the convergence is monotonic. The uncertainty is computed by tools/uncertainty.py and the grid convergence index by tools/mesh_study.py[20].

Splits

Three out-of-distribution (OOD) test sets, ood_blockage, ood_rotation and ood_position, are held out so that models can be tested on a mild extrapolation. Each holds cases beyond the parameter range of the training cases in one measure. The sets are tails of the distribution produced by the same generator, each formed from the cases above the 95th percentile of its measure (Figure 4). The measures are the blockage ratio (above 0.389), the absolute rotation (above 41.8°) and the distance of the position shift from the middle of its range, with Δx\Delta x and Δy\Delta y each divided by their half range (above 1.196). The position set therefore holds out the corners of the joint shift domain, and a shift that is extreme in one direction but central in the other can still be in the training split. The three tails hold 224, 225 and 225 cases. Thirty cases lie in two tails (9 in blockage and rotation, 8 in blockage and position, 13 in rotation and position) and none in all three. A case in more than one tail goes to the first of these sets, so the sets hold 224, 216 and 204 cases. No case of the other splits lies in any of the tails. The sets are not balanced over families and sizes. ood_blockage holds mostly larger obstacles (169 of 224) and triangles (159), and a single diamond. ood_rotation holds mostly smaller and normal obstacles (187 of 216). This is presumably because rotation raises the bounding-box height, so that strongly rotated large obstacles more often exceed the blockage limit or fall in the blockage set first. The position set is close to balanced. The other 3,856 cases are split 80/10/10 into training (3,085), validation (385) and test (386) cases within every size class and family (Table 5). The cases of each size class and family are shuffled with a generator seeded with the split seed (2026), the size and the family. The split depends only on the cases and is independent of their order.

Figure 4. The out-of-distribution test sets. Each set holds the cases above the 95th percentile (dashed) of its measure. a, Blockage ratio. b, Absolute rotation. c, Distance of the position shift from the middle of its range, with both shifts divided by their half ranges (outside the dashed circle). A case in more than one tail belongs to the first of these sets.

Table 5: Splits. Each split is one file per run, so every configuration has the same cases. ood: out-of-distribution, the cases above the 95th percentile of the measure given (threshold in brackets). A case in more than one tail is in the first of these sets.

SplitCasesContent
train3,08580 % of the in-distribution cases of every size and family
val38510 %
test38610 %
ood_blockage224blockage ratio (0.389)
ood_rotation216absolute rotation (41.8°)
ood_position204distance of the shift from the middle of its range (1.196)
total4,500300 per size and family

Table 5.

Computation and provenance

Every step that computed the data ran in one container image, built by continuous integration. The image contains OpenFOAM v2512, gmsh 4.15.2, VTK 9.7.0 and Python 3.14 with numpy, scipy, h5py and pyarrow. All 22,500 runs used the same image digest. Two steps that only unpack and read the run archives ran outside the image, because they need the Python package zstandard, which the image lacks. The case of each run continued under a raised iteration cap was restored from its archive on the cluster’s login node, and the solver resumed inside the image. The wake velocity was computed after the production, so its sampled lines were extracted from the archives in Google Colab. The production ran on the KUACC cluster of Koç University, on Intel Xeon E5-2640 and E5-2670 v2 processors, with one SLURM job and one CPU core per run. Peak solver memory was at most 1.05 GB. The 22,500 runs took 35,922 solver core-hours, 94 % of them at 0.25 mm. A 0.25 mm run costs about 185 times as much as a 1 mm run (median). It has 16 times the cells, needs 6 times the iterations and takes 1.9 times the time per iteration and cell on the same CPU model. The solver time of a run is the simpleFoam execution time, summed over all solver runs when a run was continued over several jobs. It does not include meshing, mapping or the prerequisite 1 mm solution of a mapped run, and superseded attempts are not counted. The task time covers all steps of the task that finished the run: set-up and meshing, mapping, the solver, resampling, attributes, archiving and upload. Times are as measured on the two CPU models, without normalisation. Every run records its convergence, set-up, cost and provenance (Data Records, Metadata). The complete case of every run (mesh, the last two time steps, logs and attribute outputs) was archived, so any attribute can be recomputed without solving again.

Use of large language models

Claude (Anthropic), a large language model, was used to help write the analysis and pipeline code, search and compare the literature, and draft and edit the manuscript. ChatGPT (OpenAI) was used for a final review of the text for grammar and consistency. The authors checked all code, numbers and citations and take full responsibility for the content.

Data Records

The dataset is available on Hugging Face [28] under the CC BY 4.0 licence. The release described here is version 1.0.0. It consists of 38 data files and the checksum file checksums.md5, 39 files and 104.0 GB in all. Among the 38 files is a data card (DATACARD.md) with the caveats and provenance notes. The MD5 of checksums.md5 (82a46f931bb1d60b57dbc564305019fb) identifies the release, and any change to the data would be a new version. The archived OpenFOAM case of every run is on Hugging Face [29] (22,500 archives, 1.56 TB). The folder structure is:

v1.0.0/
  0.25x/ base_fo/ base_lu/ 4x/ 16x/    each with train.h5, val.h5, test.h5,
                                      ood_blockage.h5, ood_rotation.h5, ood_position.h5
  metadata.parquet  metadata.csv  mesh_study.parquet  surface.parquet
  label_uncertainty.parquet  dropped_cases.csv  dataset_info.json
  DATACARD.md  checksums.md5

Arrays. Each file <run>/<split>.h5 holds three HDF5 datasets. The array fields is float32 of shape (N,6,nx,ny)(N, 6, n_x, n_y) and is stored with one sample per chunk and no compression, so a single sample can be read without loading the file. The names in case_name identify the case of each row, and time gives the iteration of the exported solution. Row nn is the same case in every run of a split, in the order of the metadata, so the configurations of a case are directly paired. The file attributes give the channel names (Table 3), the run, the grid, and the mean and standard deviation of every channel over the training split of that run. The grid points include both ends of the domain, so the spacing is L/(nx−1)L/(n_x - 1) and H/(ny−1)H/(n_y - 1). The axis xx points along the flow, and the inlet is at index 0. The arrays of 0.25x, base_fo, base_lu, 4x and 16x take 1.2, 4.7, 4.7, 18.7 and 74.7 GB, and the tables add about 0.2 GB.

Metadata. metadata.parquet (also as CSV) has one row per case and 259 columns. For each case, it gives the split, the case name with its size, family and index, the global seed, the sampled width, height, rotation and shift, the number of attempts and the outline points. The derived geometry comprises the bounding box, centroid, area, blockage ratio and the gaps to the walls, inlet and outlet. For every run, columns named <run>.<field> hold the 20 attributes of Table 4 and a record of the run. The record covers the convergence (status, iterations, number of solver runs, final residuals), the mesh (mesh check, mesh variant, cell count) and the set-up (template, starting fields, raised cap). It also gives the cost (solver and task time, peak memory) and the computing environment (host, CPU model, allocated cores, SLURM job, container digest, code revision).

Further tables. surface.parquet holds the pressure and wall shear stress on every obstacle face for every case and run, with the face position. mesh_study.parquet has one row per case and attribute (85,500 rows). It covers 19 attributes, all except outlet_reverse_flow, which is zero in every 0.25 mm run. Each row holds a three-level study of the 1, 0.5 and 0.25 mm levels. It gives the values of all five runs, the refinement ratios, the type of convergence (monotonic, oscillatory, none or equal) and the observed order. Where the convergence is monotonic, it also gives the extrapolated value, the error of every level against it and the grid convergence index of the 0.25 mm value. label_uncertainty.parquet holds the Eça–Hoekstra uncertainty estimate of the 19 attributes of every case, also distributed with the benchmark repository[19]. dropped_cases.csv lists the parameters and final residuals of the 18 dropped cases, whose 1 mm second-order run did not converge. dataset_info.json gives the counts per split, the runs and grids, the code revisions and the image digest. DATACARD.md summarises how to read the files and gives the attributes that are not mesh-independent, the runs that needed a raised cap or a different start, and the provenance notes.

Technical Validation

Solver setup against a reference benchmark

The steady flow past a cylinder in a channel at Re=20Re = 20 (DFG 2D-1[30]) was solved with the discretisation, SIMPLEC settings, 10−610^{-6} criterion and gmsh meshing of the production. Only the geometry, the viscosity and the parabolic inlet were taken from the benchmark. Three uniform meshes were used (D/20, D/40 and D/80), and the two finer runs started from the coarse solution, as in the production (Table 6). On the finest mesh, the drag coefficient and the recirculation length lie inside the reference intervals. Between the meshes they change non-monotonically and by less than the width of their intervals, so they were not extrapolated. The pressure difference is 0.8% low on the finest mesh, and its extrapolated value lies inside the interval. The lift coefficient is 9% above the reference value, a difference of 0.02% of cDc_D, and its iteration error is 0.15%. The lift is mostly viscous, and its pressure part changes sign between D/40 and D/80, which points to the lack of wall refinement in the uniform meshes. No wall-refined mesh was run to confirm this.

Table 6: DFG 2D-1 with the production settings. Δp\Delta p in m2 s−2\mathrm{m^2\,s^{-2}}, LaL_a in m.

MeshCellsIterationscDc_DcLc_LΔp\Delta pLaL_a
D/2035,9654105.5848−0.00130.11380.08428
D/40143,9077155.58570.01000.11570.08433
D/80575,7451,9775.58270.01160.11660.08446
Extrapolated–0.01180.1175–
Reference interval5.57–5.590.0104–0.01100.1172–0.11760.0842–0.0852

Table 6.

Iterative convergence and initial fields

All 22,500 runs stopped with every initial residual below 10−610^{-6}. The dataset audit (Completeness, integrity and reproducibility) checked this from the provenance of every run. The median number of iterations rises from 209 at 2 mm to 3,497 at 0.25 mm, with a long tail (Figure 5). The figure shows

Figure 5. Iterations to initial residuals of 10−610^{-6}. Share of the 4,500 runs of each configuration that had converged within a given number of iterations. The dot on each curve marks the median, and the first-order 1 mm run is dashed.

the computational effort. To test whether the criterion is strict enough, 12 converged 0.25 mm runs were continued from their released state with a criterion of 10−810^{-8}. They were the four runs that had needed the most iterations (41,173–52,748) and three runs at the blockage limit of 0.40. The other five were the four with the smallest uncertainty of drag or dissipation and one with the median iteration count. Ten of them reached 10−810^{-8} after 1,790 to 28,097 further iterations. The two slowest were stopped after about 30,000 further iterations, with largest residuals of 6×10−86 \times10^{-8} and 8×10−88 \times10^{-8} that were still falling. The fields changed by 0.03–0.05 % (RMS, largest in uyu_y). Each attribute changed by at most 0.21 of the label uncertainty of its case, with a median of 0.0002 over the 228 attribute–case pairs. Only three pairs exceed 0.1, all in one of the cases chosen for their small uncertainty. In the cases tested, the iteration error left at 10−610^{-6} is small against the discretisation uncertainty.

In production, the 0.5 and 0.25 mm runs start from the mapped 1 mm solution, as do the 27 relaxed 2 mm runs (Methods, Flow solution). Cases were solved on the same mesh both from rest and from these starting fields (Table 7). The remaining differences grow as the mesh gets finer and are of the order of the iteration error left at the 10−610^{-6} stop. At 0.25 mm, the largest attribute difference of each case is 0.3–8 % of its label uncertainty. The mapped starting fields reduced the number of iterations.

LevelCasesFieldsAttributesIterations, mapped vs from rest
2 mm (relaxed runs)40.007 % (uyu_y)0.001 % (base pressure)575–6,157 vs 6,909–14,576
0.5 mm60.021 % (uyu_y)0.033 % (base pressure)760–2,385 vs 1,217–4,782
0.25 mm60.10 % (uyu_y)0.15 % (base pressure)2,466–8,142 vs 3,966–15,885

Table 7. Same mesh, different starting fields. Largest difference over the cases, between runs from rest and from the mapped 1 mm solution. Fields: RMS of the difference over the RMS of the field.

A looser stopping criterion

To quantify how the labels of fine meshes depend on the stopping criterion, 113 cases were solved again with initial-residual targets of 10−310^{-3} (velocity) and 10−210^{-2} (pressure) and otherwise identical settings. They are 105 test cases, drawn from seven bins of the iteration count of their converged 0.25 mm run, and 8 cases of the blockage set. Table 8 and Figure 6 report the test cases, weighted by the inverse inclusion probability. Started from rest, the loose 0.25 mm run stops after a tenth of the iterations that the converged run needs, and its recirculation length is about a third too short. Started from the 1 mm solution, it stops almost at once, and the label stays close to the coarse solution. The coarse-to-fine difference is then understated, at 0.36 % instead of 0.47 % (relative RMS of uxu_x, Figure 6a). With both levels loose from rest, the difference is overstated by a factor of 36 (17.1 %). The loose criterion barely affects the drag and recirculation length of the coarse 1 mm run. Because the direction of the error depends on the starting fields, these results do not support a single correction applied afterwards to all runs.

RunLoose stopConvergedRMS change of uxu_xDragRecirculation length
1 mm, from rest2335710.55+0.01 % (0.0)0.00 % (0.1)
0.25 mm, from 1 mm83,3530.32+1.33 % (2.3)−1.53 % (2.9)
0.25 mm, from rest3553,35318.67+0.50 % (7.8)−32.6 % (67.2)

Table 8. Loose against converged runs. Weighted medians over the 105 test cases, with the 90th percentile of ∣⋅∣|\cdot| in brackets. Iterations are given at the loose stop and for the converged run, and the RMS change of uxu_x in % of UinU_{\mathrm{in}}.

Figure 6. Effect of a loose stopping criterion on the labels. a, Coarse-to-fine difference of every test case (relative RMS of uxu_x, 1 mm interpolated bicubically to 0.25 mm after the zeros inside the obstacle are replaced by the nearest fluid value). It is shown for the converged labels, for loose 0.25 mm labels started from the 1 mm solution and for both levels loose from rest. Ticks mark the weighted medians. b, Change of the recirculation length of the loose 0.25 mm labels, started from the 1 mm solution and from rest, against the number of iterations their converged run needed.

Discretisation error and label uncertainty

Table 9 and Figure 7 compare the error of each cheaper configuration with the Eça–Hoekstra uncertainty of the 0.25 mm label, on the scales of the level errors (Methods, Label uncertainty). The second-order levels converge towards the label. Their median error falls with every refinement for every attribute except the mean wall shear, and by half or more from 1 to 0.5 mm. For the drag it is 1.10, 0.76 and 0.33 % at 2, 1 and 0.5 mm, and for the dissipation 4.68, 1.69 and 0.53 %. The mean wall shear error rises from 2 to 1 mm (1.14 and 1.48 %) before it falls to 1.00 % at 0.5 mm. The first-order 1 mm run differs most from the label where the flow separates. Its recirculation length is 11 % off and its drag 3.9 %, whereas its pressure drag is close to the label (0.97 %). Its recirculation length is shorter than the label in 96 % of the cases, which is consistent with the numerical diffusion of first-order upwinding.

Attribute2 mm 0.25x1 mm, first order base_fo1 mm base_lu0.5 mm 4xU of the label 16x
drag1.10 (1.82)3.93 (5.17)0.76 (1.25)0.33 (0.56)1.45 (3.79)
pressure drag2.72 (4.54)0.97 (1.83)2.18 (3.38)1.02 (1.59)5.98 (11.35)
friction drag1.95 (3.41)2.89 (4.22)1.46 (2.45)0.69 (1.19)4.07 (8.70)
lift0.66 (1.93)1.34 (3.40)0.46 (1.11)0.23 (0.50)0.97 (2.97)
recirculation length2.80 (6.30)11.10 (17.38)1.62 (3.45)0.64 (1.40)2.41 (8.05)
recirculation area1.75 (3.80)7.07 (13.22)1.10 (2.33)0.49 (1.04)2.19 (6.90)
wake velocity0.63 (1.45)2.79 (5.06)0.41 (0.78)0.18 (0.33)0.73 (2.05)
base pressure1.20 (3.12)2.99 (8.43)0.58 (1.70)0.29 (0.94)1.50 (5.44)
dissipation4.68 (5.67)4.56 (7.07)1.69 (2.38)0.53 (0.80)0.49 (1.04)
mean vorticity0.65 (0.94)2.31 (3.63)0.33 (0.52)0.12 (0.20)0.34 (0.98)
Δp\Delta p between sections0.58 (1.02)2.33 (3.00)0.47 (0.76)0.20 (0.35)1.00 (2.35)
flow split0.17 (0.54)0.29 (0.85)0.09 (0.25)0.04 (0.12)0.20 (0.68)
mean wall shear1.14 (3.33)4.90 (7.75)1.48 (3.28)1.00 (2.11)4.58 (10.53)

Table 9. Error of every configuration against the 0.25 mm label, and the uncertainty U of the label, in %. Median over the 4,500 cases, with the 90th percentile in brackets. Thirteen of the 20 attributes are shown, and the uncertainty of all 19 is in the uncertainty file. Errors and U are on the scales of the level errors (Methods, Label uncertainty).

Figure 7. Error of every cheaper configuration against the 0.25 mm label, and the label uncertainty. a, Dots mark the median over the 4,500 cases, and lines extend to the 90th percentile of the cases (not a confidence interval). The grey band extends to the median uncertainty of the label, which is not a limit for single cases. b, Share of cases whose error exceeds their own uncertainty estimate. Scales as in Table 9.

Because medians do not describe single cases, Figure 7b gives the share of cases whose error exceeds their own uncertainty estimate. Where the estimate is zero, any difference counts as exceeding it. At 0.5 mm this share is below 9 % for every attribute except dissipation (61 %). For the drag and the recirculation length it is 20 and 29 % at 1 mm, 42 and 52 % at 2 mm, and 85 and 90 % for the first-order run. The dissipation error exceeds its uncertainty in nearly every case at 1 and 2 mm, and the mean-vorticity error does so in half the cases at 1 mm. For the first-order run, the share is lowest for the pressure and friction drag (6 and 32 %). For the second-order levels, these shares are not an independent test, because U is estimated from the same four levels.

They describe where each level lies relative to the estimated error of the label.

The meshes of a case are independent unstructured meshes, not nested refinements, so part of the discretisation error is irregular. The three-level grid convergence index is undefined where the convergence is not monotonic. This is the case for 34 % of the drag labels, 31–38 % of its parts, 45 % of the recirculation lengths and 43 % of the wake velocities. The Eça–Hoekstra procedure handles such data with least-squares fits and larger factors of safety. Where the data converge cleanly, as for dissipation, it gives a tighter uncertainty than the simple estimate 3∣ϕ0.25−ϕ0.5∣3|\phi_{0.25}-\phi_{0.5}| (0.49 against 1.59 %). The friction and pressure drag converge slowly on their own (median observed orders 0.29 and 0.36), but relative to the drag their errors at 0.5 mm are about 1 % or less (Table 9). The three attributes that are not mesh-independent often do not converge with the mesh. For example, the front pressure did not converge in 2,947 of the 4,500 cases.

Resampled arrays against the simulation mesh

The dataset audit checked every array of every run for its shape and finite values, zero values inside the obstacle, geometry channels consistent with the outline, and mass conservation. The obstacle area agreed with the sampled geometry to within 0.67 % at 2 mm and 0.02 % at 0.25 mm (median). For mass conservation, the volume flux through cross-sections agreed with the inlet flux to within 0.69 % at 2 mm and 0.022 % at 0.25 mm. To measure what the resampling loses, the eight attributes that can be computed from the fields were also computed from the arrays of every configuration, each on its own grid. They were compared with the values on the simulation mesh (Table 10). Gradient quantities lose the most. For dissipation and mean vorticity, the loss is first order in the grid spacing and halves with every refinement, from 13 to 1.7% for dissipation. The arrays systematically underestimate both, which is consistent with the cell-to-point interpolation smoothing the steep gradients at the walls. The inlet-to-outlet pressure drop of the arrays is about 1.4% low at every level. The array value averages the end columns of the grid, whereas the mesh value averages the patch faces, which include the singular inlet corners. Peak vorticity, taken as the largest grid value, is 22–36% low. It is not mesh-independent on the simulation mesh either.

Attribute2 mm 0.25x1 mm, first order base_fo1 mm base_lu0.5 mm 4x0.25 mm 16x
recirculation length2.00 (+0.87)1.09 (+0.67)1.05 (+0.69)0.52 (+0.39)0.25 (+0.18)
recirculation area2.51 (−2.51)0.76 (−0.76)0.80 (−0.80)0.25 (−0.24)0.07 (−0.07)
flow split0.09 (+0.01)0.02 (0.00)0.02 (0.00)0.01 (0.00)0.00 (0.00)
Δp\Delta p inlet to outlet1.52 (−1.52)1.41 (−1.41)1.46 (−1.46)1.40 (−1.40)1.36 (−1.36)
dissipation13.01 (−13.01)7.17 (−7.17)6.80 (−6.80)3.44 (−3.44)1.74 (−1.74)
mean vorticity4.35 (−4.35)2.25 (−2.25)2.14 (−2.14)1.05 (−1.05)0.52 (−0.52)
peak vorticity35.7 (−35.7)31.3 (−31.3)29.2 (−29.2)24.6 (−24.6)21.5 (−21.5)
wake velocity0.63 (−0.22)0.43 (−0.37)0.34 (−0.22)0.19 (−0.14)0.09 (−0.08)

Table 10. Attributes from the arrays against the simulation mesh. Median ∣|array − mesh∣| over the 4,500 cases in % of a scale, with the median signed difference in brackets. The scale is the mesh value of the same run, except for the recirculation length and area, relative to at least the obstacle height and its square, and the wake velocity, relative to UinU_{\mathrm{in}}.

The uncertainty procedure was also applied to the array values of the four levels, each on its own grid. The median uncertainty of the 0.25 mm array values is then 2.94% for dissipation (0.49% for the mesh values) and 0.90% for mean vorticity (0.34%). For the recirculation length and area, the flow split and the pressure drop, it is about the same as on the mesh. For gradient quantities, the stored arrays represent the flow less accurately than the attributes computed on the mesh.

Completeness, integrity and reproducibility

Before the release, an audit ran over every run of every case, and all its checks passed. It checked the membership (300 cases per size and family, five runs each, the dropped cases and the reserves used) and the convergence and provenance of every run. It also checked the archives (one per run, with the size and MD5 of all 22,500 archives on the remote store) and the arrays and attributes. To detect mixed-up cases, every configuration was compared with 0.25 mm, and no case had outliers in three or more attributes. The audit covered the mesh study and the splits, and found every geometry file regenerated from the seed identical. Its only warnings concern eight 0.25 mm runs whose starting fields were not recorded in their provenance. Their archives show that they started from the 1 mm solution like all the others, and the metadata records the starting fields of every run. In the release files, the splits equal those of the audit, row nn is the same case in every run, and all 38 checksums are correct. Five archives, one per configuration, were restored and processed again with the code at tag data-v1.0.0. They include a run continued over four solver runs, a 0.5 mm run continued under a raised cap and one of the relaxed 2 mm runs. The arrays and attributes obtained from them were identical bit for bit to the released ones. With the same code, three cases were solved again from rest at 2 mm and at 1 mm (second order). They needed the same iterations and gave identical arrays and attributes.

Usage Notes

With h5py, h5py.File("16x/test.h5")["fields"][n] returns the (6,1440,480)(6, 1440, 480) sample of row nn without loading the rest of the file. The rows follow the order of the split’s cases in the metadata (column dataset_index), so pairs of configurations need no matching beyond the row index. Every file also stores the case_name of each row. Grid point (i,j)(i,j) lies at x=iL/(nx−1)x=iL/(n_x-1) and y=jH/(ny−1)y=jH/(n_y-1) in mm, with the inlet at i=0i=0. The grids of different configurations share the end points but not the interior points, so two of them cannot be compared without interpolation.

The value 1 of cell-type marks both the obstacle and the two channel walls. A mask of the fluid is cell_type != 1, and obstacle_sdf <= 0 marks the obstacle alone. The signed distance is zero on the outermost obstacle points, so obstacle_sdf < 0 leaves them out. Because the inlet corners are singular, error measures of gradients are best taken without the first 5 mm of the channel, as for the vorticity and dissipation attributes. The normalisation statistics in each file (Data Records) are computed over all grid points, including the zero values inside the obstacle. Statistics of the test or out-of-distribution sets should not be used for normalisation. The cell type is a category and should not be normalised with its statistics.

The attributes in the metadata are the values on the simulation mesh, and the same quantity computed from the arrays differs from them by the resampling loss (Table 10). Values from predicted fields should therefore be compared with values computed in the same way from the label arrays. Such a comparison measures agreement with the stored arrays rather than accuracy against the solution on the mesh, because the two resampling losses need not cancel. The uncertainty of the 19 attributes of every case is in label_uncertainty.parquet (Data Records). Rows with source = openfoam are the uncertainty of the 0.25 mm solution, and rows with source = array describe the convergence of the array values. The column method gives the basis of each estimate: power, low_order or high_order for a fitted estimate, anomalous for the fallback of three times the largest difference, and equal where all four values are equal.

For every run, the metadata records the starting fields, any raised iteration cap, the number of solver runs, the mesh variant and the template, so runs that deviate in any of these respects can be found or left out. Each run records the code revision deployed when it finished. For a run continued over several jobs, this revision also holds for the earlier jobs, because the container image and the attribute, resampling and array code did not change during production. The recorded revision does not hold for the wake velocity and the wall distance (channel 3). Both were computed for every run after the production, and the code at tag data-v1.0.0 (builder revision a21e54b in dataset_info.json) gives the released values. Every geometry can be regenerated from the seed with the pipeline[20]. The archived cases are complete OpenFOAM cases, which can be opened with the pipeline’s container image to compute further quantities or other resamplings.

Data Availability

The arrays, tables, data card and checksums of version 1.0.0 (Data Records) are available on Hugging Face[28]. The archived OpenFOAM case of every run is available on Hugging Face[29]. The label uncertainty is also distributed with the benchmark repository[19].

Code Availability

The pipeline that generated the data[20], tag data-v1.0.0, is available under the MIT licence. It covers geometry sampling, meshing, the OpenFOAM case templates, resampling and the attributes, as well as the production on a SLURM cluster and the dataset audit. Every step that computed the data ran in one container image, ghcr.io/asaydam/ms-cofd-2dsl (digest sha256:ae5bc11fb90901b2d7cab7012fb91a211a268a865f19aef7cd2b3d6e6a2d0364), with OpenFOAM v2512 (ESI), gmsh 4.15.2, VTK 9.7.0 and Python 3.14. Two steps that only unpack and read the run archives ran outside it (Methods, Computation and provenance). The settings are in config/dataset.yaml and the case templates. The per-run exceptions (the raised caps and the relaxed template) are recorded in the metadata, and every run records its code revision. The label uncertainty is computed by tools/uncertainty.py. Tables 9 and 10 are computed by validation/level_errors.py and validation/array_vs_openfoam.py, respectively, and the other checks of Technical Validation by the scripts in validation/, documented in docs/validation.md. The benchmark repository[19] holds readers, interpolation baselines and the label-uncertainty file.

Author Contributions

A.S. and B.A. designed the dataset. A.S. developed the data-generation pipeline and the benchmark code, carried out the simulations and the technical validation, and wrote the manuscript. B.A. proposed the force-based quantities, supervised the project, and reviewed and edited the manuscript.

Competing Interests

The authors declare no competing interests.

Funding

This work is supported by the KUIS AI Center of Koç University, Istanbul. We gratefully acknowledge this support.

References

  1. [1]Bonnet, F., Mazari, A. J., Cinnella, P. & Gallinari, P. AirFRANS: High fidelity computational fluid dynamics dataset for approximating Reynolds-averaged Navier–Stokes solutions. In Advances in Neural Information Processing Systems 35, Datasets and Benchmarks Track (2022). https://doi.org/10.48550/arXiv.2212.07564
  2. [2]Elrefaie, M., Morar, F., Dai, A. & Ahmed, F. DrivAerNet++: A large-scale multimodal car dataset with computational fluid dynamics simulations and deep learning benchmarks. In Advances in Neural Information Processing Systems 37, Datasets and Benchmarks Track (2024). https://doi.org/10.48550/arXiv.2406.09624
  3. [3]Tali, R. et al. FlowBench: A large scale benchmark for flow simulation over complex geometries. J. Data-centric Mach. Learn. Res. 2, 1–35 (2025). https://data.mlr.press/assets/pdf/v02-14.pdfarxiv.org/abs/2409.18032
  4. [4]Luo, Y., Chen, Y. & Zhang, Z. CFDBench: A large-scale benchmark for machine learning methods in fluid dynamics. Preprint at https://doi.org/10.48550/arXiv.2310.05963 (2023).
  5. [5]Takamoto, M. et al. PDEBench: An extensive benchmark for scientific machine learning. In Advances in Neural Information Processing Systems 35, Datasets and Benchmarks Track (2022). https://doi.org/10.48550/arXiv.2210.07182
  6. [6]Ohana, R. et al. The Well: A large-scale collection of diverse physics simulations for machine learning. In Advances in Neural Information Processing Systems 37, Datasets and Benchmarks Track (2024). https://doi.org/10.48550/arXiv.2412.00568
  7. [7]Ashton, N. et al. DrivAerML: High-fidelity computational fluid dynamics dataset for road-car external aerodynamics. Preprint at https://doi.org/10.48550/arXiv.2408.11969 (2024).
  8. [8]Kanchi, R., Melanson, B., Somasekharan, N., Pan, S. & He, S. UniFoil: A universal dataset of airfoils in transitional and turbulent regimes for subsonic and transonic flows. In Advances in Neural Information Processing Systems 38, Datasets and Benchmarks Track (2025). https://doi.org/10.48550/arXiv.2505.21124
  9. [9]Yang, Y. et al. SuperWing: a comprehensive transonic wing dataset for data-driven aerodynamic design. Sci. Data 13, 1113 (2026). https://doi.org/10.1038/s41597-026-07769-0
  10. [10]Jakeman, J. D., Barba, L. A., Martins, J. R. R. A. & O’Leary-Roseberry, T. Verification and validation for trustworthy scientific machine learning. Mach. Learn.: Sci. Technol. 7, 025055 (2026). https://doi.org/10.1088/2632-2153/ae59ec
  11. [11]Ye, M., Zhong, Y., Chen, H.-C. & Wan, D. Investigating the impact of numerical uncertainties in CFD-generated training data on neural network models for wind-turbine wake predictions. Ocean Eng. 343, 123200 (2026). https://doi.org/10.1016/j.oceaneng.2025.123200
  12. [12]Koehler, F. & Thuerey, N. Neural emulator superiority: When machine learning for PDEs surpasses its training data. In Advances in Neural Information Processing Systems 38 (2025). https://doi.org/10.48550/arXiv.2510.23111
  13. [13]Shen, Y. & Alonso, J. J. A multi-fidelity double-delta wing dataset and empirical scaling laws for GNN-based aerodynamic field surrogates. In AIAA SCITECH 2026 Forum (AIAA, 2026). https://doi.org/10.2514/6.2026-0686
  14. [14]Xu, W., Grande Gutiérrez, N. & McComb, C. MegaFlow2D: A parametric dataset for machine learning super-resolution in computational fluid dynamics simulations. In Cyber-Physical Systems and Internet of Things Week 2023 (Workshops) 100–104 (ACM, 2023). https://doi.org/10.1145/3576914.3587552
  15. [15]Setinek, P., Galletti, G. & Brandstetter, J. Towards multi-fidelity scaling laws of neural surrogates in CFD. Preprint at https://doi.org/10.48550/arXiv.2511.01830 (2025).
  16. [16]Setinek, P., Paischer, F., Thurey, N., Brandstetter, J. & Koehler, F. Compute optimal data scaling for neural surrogates via multi-fidelity training. Preprint at https://doi.org/10.48550/arXiv.2610.17221 (2026).
  17. [17]Chung, W. T. et al. Turbulence in focus: Benchmarking scaling behavior of 3D volumetric super-resolution with BLASTNet 2.0 data. In Advances in Neural Information Processing Systems 36 (2023). https://doi.org/10.48550/arXiv.2309.13457
  18. [18]Ribeiro, M. D., Rehman, A., Ahmed, S. & Dengel, A. DeepCFD: Efficient steady-state laminar flow approximation with deep convolutional neural networks. Preprint at https://doi.org/10.48550/arXiv.2004.08826 (2020).
  19. [19]Saydam, A. & Akgün, B. ms-cofd-2dsl-bench: tasks, metrics and baselines, tag v1.0.0. GitHub https://github.com/ASaydam/ms-cofd-2dsl-bench/tree/v1.0.0 (2026).
  20. [20]Saydam, A. & Akgün, B. ms-cofd-2dsl: generation pipeline, tag data-v1.0.0. GitHub https://github.com/ASaydam/ms-cofd-2dsl/tree/data-v1.0.0 (2026).
  21. [21]Geuzaine, C. & Remacle, J.-F. Gmsh: A 3-D finite element mesh generator with built-in pre- and post-processing facilities. Int. J. Numer. Methods Eng. 79, 1309–1331 (2009). https://doi.org/10.1002/nme.2579
  22. [22]Weller, H. G., Tabor, G., Jasak, H. & Fureby, C. A tensorial approach to computational continuum mechanics using object-oriented techniques. Comput. Phys. 12, 620–631 (1998). https://doi.org/10.1063/1.168744
  23. [23]Van Doormaal, J. P. & Raithby, G. D. Enhancements of the SIMPLE method for predicting incompressible fluid flows. Numer. Heat Transf. 7, 147–163 (1984). https://doi.org/10.1080/01495728408961817
  24. [24]OpenCFD Ltd. OpenFOAM v2512, function normFactor in src/OpenFOAM/matrices/lduMatrix/lduMatrix/lduMatrixSolver.C. https://gitlab.com/openfoam/core/openfoam (2025).
  25. [25]Schroeder, W., Martin, K. & Lorensen, B. The Visualization Toolkit: An Object-Oriented Approach to 3D Graphics 4th edn (Kitware, 2006).
  26. [26]Eça, L. & Hoekstra, M. A procedure for the estimation of the numerical uncertainty of CFD calculations based on grid refinement studies. J. Comput. Phys. 262, 104–130 (2014). https://doi.org/10.1016/j.jcp.2014.01.006
  27. [27]Roache, P. J. Perspective: A method for uniform reporting of grid refinement studies. J. Fluids Eng. 116, 405–413 (1994). https://doi.org/10.1115/1.2910291
  28. [28]Saydam, A. & Akgün, B. ms-cofd-2dsl (Revision b312ddef). Hugging Face https://doi.org/10.57967/hf/11554 (2026).
  29. [29]Saydam, A. & Akgün, B. ms-cofd-2dsl-raw (Revision 1f12a7b). Hugging Face https://doi.org/10.57967/hf/11553 (2026).
  30. [30]Schäfer, M., Turek, S., Durst, F., Krause, E. & Rannacher, R. Benchmark computations of laminar flow around a cylinder. In Flow Simulation with High-Performance Computers II, Notes on Numerical Fluid Mechanics 547–566 (Vieweg+Teubner, 1996). https://doi.org/10.1007/978-3-322-89849-4_39

Paper details

vlnc:W_9kd888z4
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Claude, ChatGPT

View statement
Looking for related papers…
Contents