Cross-code comparison¶
English | Japanese
A table of which features and validation cases each code has is in LBM feature matrix.
Cross-code comparison — OpenRheo vs the external LBM galleries¶
Generated by scripts/compare_external.py. Each row runs OpenRheo's own
benchmark and places its result beside what an external code computed for
the same problem — but both are judged against the ORIGINAL published
reference, not against each other: two LBM codes agreeing proves nothing
if they share a modelling error.
Licensing. OpenLB (GPLv2), Palabos (AGPLv3), waLBerla (GPLv3) and FluidX3D (non-commercial) are incompatible with OpenRheo's Apache-2.0. No line of their source is read or ported. Only facts are used: which cases exist, the numbers their runs printed, and published benchmark values. Every OpenRheo solver here is clean-room from the cited paper.
| case | QoI | OpenRheo | published reference | rel. err | external code |
|---|---|---|---|---|---|
✅ cavity_ghia |
min of the vertical-centreline u_x/U_lid | -0.376 | -0.3829 — Ghia, Ghia & Shin (1982), Re=1000 | 1.80% | openlb -0.3693 †, palabos -0.3755 † |
✅ couette |
L2 error of u(z) vs the exact linear profile | 1.556e-07 | analytic Couette profile | 0.00% | walberla (not run) |
✅ cylinder_st_cd |
Cd_max (DFG 2D-2, Re=100, unsteady) | 3.474 | 3.23 — Schaefer & Turek (1996) DFG 2D-2, Cd_max = 3.22-3.24 | 7.55% | openlb 5.821, openlb 5.591 |
✅ cylinder_st_strouhal |
Strouhal number (DFG 2D-2, Re=100) | 0.2982 | 0.3 — Schaefer & Turek (1996) DFG 2D-2, St = 0.295-0.305 | 0.60% | — |
✅ poiseuille_eoc |
L2 error of u(z) vs the exact parabola | 0.0003181 | analytic Hagen-Poiseuille parabola | 0.03% | openlb 0.001177 |
✅ sphere_drag |
drag coefficient of a sphere at Re=100 | 1.111 | 1.09 — Clift-Gauvin correlation, Re=100 | 1.95% | walberla (not run) |
✅ tgv3d |
peak dissipation rate eps(t), Re=1600 | 0.0124 | 0.0125 — Brachet (1983) / van Rees (2011) DNS Re=1600, peak eps ~ 1.25e-2 | 0.80% | openlb 0.009248 |
✅ thermal_cavity |
Nusselt number at the hot wall, Ra=1e4 | 2.244 | 2.243 — de Vahl Davis (1983), Ra=1e4 Nu = 2.243 | 0.05% | openlb 1.104 † (different Ra) |
Notes per case:
- cavity_ghia — driven runs: OpenLB -0.3693, Palabos -0.3755 (this session's earlier sweep)
- couette — waLBerla ships the case but it was not run here
- cylinder_st_cd — OpenLB ran the OTHER benchmark (2D-1, Re=20): its 5.82 is vs the 5.58 2D-1 reference, NOT comparable to our 2D-2 number
- cylinder_st_strouhal — OpenLB's steady 2D-1 run has no shedding frequency to compare
- poiseuille_eoc — OpenLB prints the same L2 norm
- sphere_drag — waLBerla ships the case but it was not run here
- tgv3d — OpenLB Turbulence_Dissipation_Rate.dat (last row, not necessarily the peak)
- thermal_cavity — OpenLB ran Ra=1e6-class settings (Nu 1.104 at its own Ra) -- not the same Ra, quoted for context only
Legend: ✅ within tolerance of the published value · ⚠️ ran but outside
tolerance · ❌ did not produce the QoI · † = not obtained in this comparison
run (an earlier sweep, or a run at other settings). These numbers were not
produced under identical conditions, so do not read them as a ranking. Logs: _agent_out/xcompare/*.log (written by scripts/compare_external.py when run).
About the license (clean-room)¶
OpenRheo is Apache-2.0. OpenLB (GPLv2) / Palabos (AGPLv3) / waLBerla (GPLv3) / FluidX3D (non-commercial) are all license-incompatible, and none of their source code has been read or ported. What is used is only the fact of which cases exist and the reference values published in each case's original paper. Every solver in this gallery is implemented clean-room from the original paper, and is always judged against the original literature (never by matching the output of another code).