mf-zonal2d¶
英語 | 日本語
« Home · « 多段階(1D ⇄ FVM ⇄ LBM)
Whole-vs-zonal: how wrong is it to model a system as a chain of independent regions? This is the question every multi-fidelity solver assumes the answer to and none of them measures.
Build & run¶
mf_zonal2d [CONFIG] (KEY = VALUE lines; see named_args.hpp)
このページの図を作った計算のコスト(Miyabi-G Grace CPU (72 cores, -stdpar=multicore)): 257.6 s wall。
Validation¶
| Tier | T1 — 自動ゲートあり |
| Gate | ✅ 直接ゲート |
Source & run details¶
| Item | Value |
|---|---|
| Target | orheo-mf-zonal2d |
| Source | native/benchmarks/mf_zonal2d.cpp |
| Lattice | D2Q9 |
| GPU twin | no |
| Modules | channel2d.hpp, multifidelity.hpp, stations.hpp |
| Gallery run | 2026-07-28 @ c9b0572c |
| Args | `` |
| Wall time | 257.63 s |
| CPU time | 18496.06 s |
| Device | Miyabi-G Grace CPU (72 cores, -stdpar=multicore) |
Computation log¶
この計算の実際の出力(先頭と末尾)。
=== whole vs zonal: two slot orifices in a plane channel ===
h=40 opening=20 cells (open fraction 0.50) thickness=2
Re=100 (V h/nu) ulat=0.0400 nu=0.016000 tau=0.5480
legs: L1=800 (20h) L3=2400 (60h) steps=400000 per configuration
[1] SINGLE orifice -- the region response a 1D network stores
single Nx= 3202 V=0.04076 K=+1.9687 dp=+1.635e-03 mass=0.000% dev(grad 1.049/1.009, alpha 0.999/0.999)
[2] WHOLE (both resolved) vs ZONAL (chained) at each separation
zonal model = K_orifice + friction(s - th) + K_orifice,
assembled through multifidelity.hpp's series solver.
s=2h Nx= 3284 V=0.04076 K=+4.0532 dp=+3.366e-03 mass=0.000% dev(grad 1.054/1.009, alpha 0.999/0.999)
s/h=2 K_whole=+4.0532 K_zonal=+4.3968 (2K_orif=+3.9375 + K_gap=+0.4593) decomposition error 8.48%
s=5h Nx= 3404 V=0.04076 K=+5.1722 dp=+4.296e-03 mass=0.000% dev(grad 1.057/1.009, alpha 0.999/0.999)
...
What this means for the 1D tier: the question is not only
"are the components far enough apart" but "how long is the
leg that never develops". Interference costs 0.36 in K at
2 gaps and is gone by 5; the leg costs +0.0203 in K for
EVERY gap of separation and does not stop. A separation
criterion written only in terms of L/D gets this backwards
for long ducts.
RESULT h=40 openf=0.50 Re=100 K_single=1.968743 s_far=20.0 err_far=0.041812 err_near=0.084776 interference=-0.358709 leg_slope=0.020300 monotone=1