aloeme/rotating-cube-cfd

Parametric rotating-cube CFD dataset

Unsteady laminar flow around a finite, rotating cube, simulated with OpenFOAM ESI v2512 (pimpleFoam, AMI sliding-mesh) over a 2-D parameter grid. Built to train/evaluate neural operators on oscillating/rotating-structure flows, extending the 3-D rotating-cube case of Gao, Cheng & Jaiman (φ-GNN) from a single simulation to a 25-case sweep.

Parameter grid (25 cases)

• Re ∈ {50, 100, 150, 200, 300} — varied by viscosity only (nu = U·c/Re, with U=c=1).
• psi ∈ {0, 15, 30, 45, 60}° — inflow elevation out of the x–y plane, U_inf = (cos psi, 0, sin psi); in-plane azimuth beta = 0 (a symmetry, fixed).
• Constant everywhere: |U_inf|=1, cube edge c=1, physical spin omega=0.25 rad/s (omega* = omega·c/U = 0.25), dt* = 5e-3, sampling every 8 steps (dt*_sample = 0.04), to t* = 200. The cube rotates about +z; rotation sense fixed.

psi is the physically meaningful axis: tilting the inflow toward the spin axis adds an axial through-flow that is not reducible to a frame rotation (oblique/yawed rotating flow, helical wakes). beta is degenerate (frame rotation about z) and held at 0.

Repository layout

data/<case>/<case>.tar.zst.partNNN  # ~4 GB split parts of the per-case archive
data/<case>/<case>.tar.zst.sha256   # checksum of the reassembled archive
data/<case>/PARTS_COMPLETE          # marker: all parts for this case are uploaded
visualization/<case>/               # 3 mp4s per case: vorticity, Umag (|U|), p (z=0 slice)
code/                               # generator + solver + visualization scripts
README.md                           # this card

Archives are split into ~4 GB parts (the run server's uplink can't reliably commit a single 55 GB file). Reassemble a case, verify, and extract:

cat data/Re100_psi30/Re100_psi30.tar.zst.part* > Re100_psi30.tar.zst
sha256sum -c data/Re100_psi30/Re100_psi30.tar.zst.sha256   # (path in the .sha256 is the bare name)
zstd -d Re100_psi30.tar.zst -c | tar -x                    # -> Re100_psi30/

Each <case>.tar.zst extracts to an OpenFOAM case directory decomposed across 12 MPI ranks (processor0..11/). Storage is lean by design: the rigid-rotation mesh is stored once (processor*/constant/polyMesh, with the rotating cellZone), and only U and p are written per snapshot (binary), 4991 snapshots per case over t* = 0..200.

Reconstructing a sample

The mesh is the t=0 configuration; the inner cylinder (r < c) co-rotates rigidly. To get physical node positions at snapshot time t, rotate the rotating cellZone points by theta = omega·t about +z (omega = 0.25). U is stored in the absolute frame. A worked z=0-slice example (with this rotation applied) is in code/visualize_sample.py.

Force coefficients (stationary average, t* ∈ [150, 200])

Aref = c² = 1, magUInf = 1; Cd drag (x), Cl lift (y, Magnus), Cs axial (z).

Re \ psi	0°	15°	30°	45°	60°
Cd 50	2.334	2.270	2.094	1.790	1.321
100	1.718	1.690	1.620	1.428	1.038
150	1.513	1.497	1.407	1.289	0.916
200	1.423	1.413	1.307	1.075	0.843
300	1.358	1.354	1.297	1.096	0.641
Cs 50	0.000	-0.653	-1.163	-1.535	-1.811
100	0.000	-0.503	-0.844	-1.109	-1.308
150	0.000	-0.463	-0.824	-0.991	-1.155
200	0.000	-0.450	-0.827	-1.087	-1.095
300	0.000	-0.430	-0.816	-1.032	-1.149

Trends (both axes carry distinct physics): Cd falls with Re and with psi; Cs (axial force) grows 0 → ~-1.8 with psi; in-plane shedding amplitude shrinks with psi and grows with Re; Cl < 0 throughout (Magnus lift from the fixed +z spin).

Validation & caveats

• Static-cube validation (low-blockage, separate mesh): Cd = 0.978 vs paper 0.935, empirical bracket [0.854, 1.122] — pass.
• Laminar-validity ceiling: highest at the high-Re / low-psi corner (Re=300, psi ≤ 15), where shedding fluctuation peaks; treat that corner as the resolution/validity edge of the current mesh. Higher psi monotonically stabilises the wake.
• Mesh: ~294k hex cells, semi-structured (O-grid in x–y, banded in z); cube finite in z (|z| < c/2). Numerics: 2nd-order implicit time, Gauss linear, PIMPLE, 1 non-orthogonal corrector, residual tol 1e-8.

Citation

Reproduces/extends: R. Gao, Z. Cheng, R. K. Jaiman, A Mesh-Adaptive Hypergraph Neural Network for Unsteady Flow Around Oscillating and Rotating Structures.