Womersley pulsatile pipe — speed and accuracy of LBM and FVM¶
English | Japanese
What was compared¶
Pulsatile flow in a circular pipe (pressure gradient F = A cos ωt) is compared with the exact solution (Womersley 1955). With the same problem and the same checks (a single-harmonic fit of the wall shear and a periodicity check), LBM was compared with OpenFOAM's pimpleFoam (header of womersley_pipe3d.cpp).

- LBM: D3Q19, Guo forcing, Bouzidi walls, TRT (Λ = 1/2). A 3D periodic pipe. α = 4 uses N = 48 (R = 22), α = 12 uses N = 142 (R = 69).
- FVM: pimpleFoam, axisymmetric wedge mesh (40 radial cells for α = 4, 80 cells for α = 12).
- Both run 0.5 α² periods (8 for α = 4, 72 for α = 12) before measuring.

Results¶
First measurement (LBM on A100 and FVM on 1 Cascade Lake core, from the source header):
| α | Method | Amplitude error | Phase error | Time |
|---|---|---|---|---|
| 4 | pimpleFoam | −0.024 % | +0.170° | 11.29 s |
| 4 | LBM | −0.51 % | +0.19° | 5.55 s |
| 12 | pimpleFoam | +0.019 % | +0.034° | 440.31 s |
| 12 | LBM | −0.13 % | +0.11° | 78.49 s |
LBM is 2.0 times faster at α = 4 and 5.6 times faster at α = 12, but its amplitude error is 21 times and 6.8 times larger. In terms of cost for the same accuracy, FVM wins at both α. This case favours FVM.
In the re-measurement on GH200 (LBM) and 1 Grace core (FVM), all errors reproduced to the same digits and only the times changed: LBM 3.58 s / 39.78 s, FVM 8.36 s / 328.71 s. The speed ratio becomes about 2.3 times at α = 4 and about 8.3 times at α = 12. The error ratios (21 times, 6.8 times) do not change.

How to read this¶
- The larger α is, the smaller the gap, but it does not disappear. FVM can exploit axisymmetry and needs 160 cells against 6,112 cells for the 3D LBM at α = 4 (1/38 of them).
- LBM is suited to unsteady flow in geometries where symmetry cannot be exploited, free surfaces, and moving boundaries (the conclusion of the header).
- Grid guideline: to match the shape of the cross-sectional profile, R/dx ≳ 5.7 α (the basis for N = 142 at α = 12).
Open points¶
- "2.0–5.6 times" and "Cascade Lake" in the source header have not been corrected since the re-measurement on GH200.
- The re-measured FVM times are the solver run time for α = 4 and the job elapsed time for α = 12, so the basis is not consistent (on elapsed time, α = 4 is 2.5 times).