forest-ahmed-dynamic¶
英語 | 日本語
Dynamic (feature-following) AMR on the Ahmed body: the fine patch TRACKS the turbulent wake shed off the bluff base, refining downstream where the vortex structures are and coarsening once the wake has advected past -- while a PROTECTION FLOOR (docs/design/dynamic-amr-design.md sec4.5) keeps every moving coarse/fine interface out of the near-body / near-wake region so the hydrodynamic force is never corrupted by a regrid-synchronous transient (the dynamic reincarnation of the lv3/lv4 force bug).
概要: 後流追従の動的 AMR。格子が組み替わる際に抗力係数が跳ねないことを確認する定性ケース。
Geometry¶

Fields¶
この計算が最後の出力ステップで書き出した場をそのまま描いたもの。3D ケースは中央断面。


Plots¶
この計算が書き出した QoI の時系列・分布。参照値がある場合は水平線で重ねている。


Build & run¶
cmake -S native -B native/build && cmake --build native/build -j
./native/build/orheo-forest-ahmed-dynamic
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)): 119.4 s wall。
Validation¶
| Tier | T3 — デモ(定量ゲートなし) |
| QoI | Cd の regrid 安定性 |
| Reference | ERCOFTAC case082(定性) |
| Gate | — ゲートなし |
Source & run details¶
| Item | Value |
|---|---|
| Target | orheo-forest-ahmed-dynamic |
| Source | native/benchmarks/forest_ahmed_dynamic.cpp |
| Lattice | — |
| GPU twin | yes (orheo-forest-ahmed-dynamic-gpu) |
| Modules | ahmed_solver.hpp, dynamic_amr.hpp, forest_regrid.hpp, forest_sensor_forest.hpp, forest_step.hpp |
| Gallery run | 2026-07-27 @ 7628c0cf |
| Args | . 8 900 2 60 1200 0.05 0.50 5 |
| Wall time | 119.37 s |
| CPU time | 7250.96 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)。
=== forest Ahmed DYNAMIC AMR ===
nh=8 dx0=0.03600m blocks=32x8x6 (kNb=4) tau=0.5010 nu=3.333e-04
max_level=2 N_regrid=60 steps=900 re=1.2e+03 ulat=0.050 cs=0.500
collide=RR wmodel=0
floor: near-wake x[-0.252,0.173]->L2, corridor x[0.173,2.400], wake |y|<=0.294 z<=0.458
HI=0.30 LO=0.10 b_buf=1 H_coarsen=2 force_every=5 Aref=1383.1
[dynamic] step 0/900 t=0.00 Cd=37.1237 Cl=11.7046 nfine=5456 leaves=9334 regrids=0 0s
[dynamic] step 500/900 t=0.20 Cd=2.0233 Cl=14.8089 nfine=8064 leaves=12036 regrids=8 19s
[dynamic] step 900/900 t=0.35 Cd=2.0841 Cl=13.6392 nfine=8096 leaves=12064 regrids=15 34s
[static_fine] step 0/900 t=0.00 Cd=37.1237 Cl=11.7046 nfine=17360 leaves=21010 regrids=0 0s
[static_fine] step 500/900 t=0.20 Cd=1.5137 Cl=17.3844 nfine=17360 leaves=21010 regrids=0 36s
[static_fine] step 900/900 t=0.35 Cd=1.6385 Cl=11.6067 nfine=17360 leaves=21010 regrids=0 64s
[static_coarse] step 0/900 t=0.00 Cd=37.1237 Cl=11.7046 nfine=5456 leaves=9334 regrids=0 0s
...
================ SUMMARY ================
WAKE-TRACKING budget: dynamic avg fine-blocks=7991 vs static-fine=17360 => 46.0% of the uniform-fine patch
FORCE accuracy: Cd dynamic=1.7074 static-fine=1.6968 static-coarse=1.7799 (dyn Cl_rms=8.8384, sfine Cl_rms=8.4616)
|Cd_dyn - Cd_sfine| = 0.0106 (Cd scatter dyn base_rms=3.99e-01)
sec4.5 FORCE GATE: Cd'-jump-at-regrid RMS=4.405e-01 vs turbulent background RMS=3.993e-01 => ratio=1.10
regrid events sampled=15. CLEAN (no coherent Cd kick at regrid)
CSVs written to .