Skip to content

rheoMesh — an AMR-native octree mesher generating both body-fitted prism-layer meshes for FVM and block forests for LBM

English | Japanese

« Home

Navigation: design · results and comparisons · development log and detailed history · AMR literature survey · map

Summary: rheoMesh is the mesh generator of OpenRheo. It reads the snappyHexMeshDict and the geometry files as they are, builds a layered FVM mesh (an OpenFOAM polyMesh) with threads, and also writes an LBM block forest from the same octree. Source: mesh/ (Apache-2.0). Status as of 2026-11-21; all DrivAer figures and numbers are provisional.

Procedure of the snappyHexMeshLike mode. castellate makes the refined mesh outside the geometry, snap moves boundary points onto the geometry, shrink pushes the wall points back by the layer thickness, and extrude fills the gap with layers

Procedure of the snappyHexMeshLike mode (castellate → snap → shrink → extrude)

  • Hex-dominant octree, with conformal polyhedral transition cells and prism layers grown from the surface. No OpenFOAM library is linked.
  • No domain decomposition. ISO C++ standard parallel algorithms, threads in one process; the same input gives the same mesh bit for bit at any thread count and on any machine.
  • Clean-room development. Written only from the specification and the papers and public documents it cites; snappyHexMesh and cfMesh were used as black boxes whose output alone was examined.

The two modes

snappyHexMeshLike mode (main line) native method (layers in two passes)
How to choose it RHEOMESH_MODE=snappyHexMeshLike (alias shm) no RHEOMESH_MODE; settings in a recipe file
Procedure as snappyHexMesh (castellate → snap → shrink → extrude) remove cells near the surface from the core, close the space with the pillow and project it onto the wall
Status (2026-11-22) default rule set shm-defaults-v2611.6 frozen (2026-11-20); only its layer field was ported to snappyHexMeshLike as S1

rheoMesh at a glance

Centre-plane sections, left: snappyHexMesh, right: rheoMesh in snappyHexMeshLike mode. Inside the body, the number of layers per wall face seen from the side (red 0, yellow 1, green 2, blue 3). Below, enlarged sections of A the underbody pocket, B the rear window, C the diffuser edge and D the rear underbody protrusion

DrivAer centre-plane sections and wall layer counts (left: snappyHexMesh, right: rheoMesh). The layers stay continuous around the wheels.

DrivAer centre plane
DrivAer centre plane
DrivAer boundary layers
DrivAer boundary layers
Wheel boundary layers
Tyre / wheel boundary layers
Underbody detail
Underbody detail
motorBike front
motorBike front
2D schematic of the LBM block forest made from the same octree
LBM block forest (same octree)

Headlines

All 88 OpenFOAM tutorials (2026-11-21)

All tutorials with a snappyHexMeshDict in OpenFOAM v2606 (65) and OpenFOAM-13 (23) were run on Miyabi-C. 37 newly pass Mesh OK versus v2611.5, with 0 regressions.

Mesher / version Mesh OK (all 88) OpenFOAM-13 (23) OpenFOAM v2606 (65)
snappyHexMesh (2026-11-20) 57 / 88 16 / 23 41 / 65
rheoMesh v2611.5 (2026-11-20) 45 / 88 8 / 23 37 / 65
rheoMesh integration build v2611.6 (2026-11-21) 82 / 88 (82/83 runnable = 98.8 %) 23 / 23 (100 %) 59 / 65 (90.8 %)
rheoMesh latest (42f96bce, 2026-11-22) 83 / 88 (83/83 runnable = 100 %) 23 / 23 (100 %) 60 / 65 (92.3 %)

On the 45 cases both pass, rheoMesh is faster on 40 (median speedup 1.98×). With max_aspect = 950 enabled by default, the sole remaining runnable case airfoilWithLayers (20 layers, expansionRatio 1.5) now passes Mesh OK in 0.93 s (max aspect ratio 203.6), reaching 100 % across all 83 runnable tutorials (the other 5 reference an unreplaced placeholder stlName.stl and cannot be run by snappyHexMesh either). Full table in the detailed history.

Public GitHub snappyHexMeshDict cases (2026-11-22)

Beyond the official tutorials, public GitHub snappyHexMeshDict cases (external car aero, 21-inch UAV propeller at 3 resolutions, 4-blade quadcopter drone, floating boat hull, urban buildings, and Ahmed body RANS 8.21 M / LES 16.06 M cells) were compared on one Miyabi-C node. All 10 runnable cases pass Mesh OK (0 faces over 70°), run 4.1×–9.3× faster than 32-rank parallel snappyHexMesh, and reduce max non-orthogonality from ~65° to 42°–60°.

Public GitHub snappyHexMeshDict benchmark. Left: wall time for 32-rank parallel snappyHexMesh versus 1-node threaded rheoMesh (4.1x to 9.3x faster); right: max non-orthogonality across all 10 Mesh OK cases up to 16.1M cells

airshaper_motorbike (external aero, 5.2× faster) propeller_21in_medium (UAV propeller, 6.7× faster)
airshaper_motorbike mesh section comparison: rheoMesh vs snappyHexMesh propeller_21in_medium mesh section comparison: rheoMesh vs snappyHexMesh
ahmed_tutorials (Ahmed body 1.64 M cells, 5.4× faster) boat_hull_floating (ship hull, 9.3× faster)
ahmed_tutorials mesh section comparison: snappyHexMesh vs rheoMesh boat_hull_floating mesh section comparison: snappyHexMesh vs rheoMesh

DrivAer 3 M cells: versus snappyHexMesh (3 layers, updated 2026-11-22)

snappyHexMesh (3 layers, 11-10) rheoMesh (3 layers, 11-10) rheoMesh latest (3 layers, 11-22)
Time with layers 37.4 s 24.0 s 27.3 s
Time without layers 30.7 s 8.2 s 11.6 s (maxNonO 50.3°)
Cells (with layers) 3,068,471 3,122,614 3,241,521
Faces with non-orthogonality over 60° 52 39 37 (max 64.40°, 0 over 70°)
Faces with skewness over 3 172 53 27 (max 3.812)
Pocket volume between wall and geometry 841 cm³ 149 cm³ —
Faces with layers: body 92.6 % 98.8 % 98.9 % (99.4 % by area)
Faces with layers: wheels (front / rear) 58.1 % / 54.9 % 94.0 % / 94.4 % 93.4 % / 93.5 % (96.3 % / 96.1 % by area)

DrivAer 3-layer mesh comparison: left snappyHexMesh, right rheoMesh (layer counts on body/wheels and zoom-ins A–D)

Times are the mesher process only on one node (snappyHexMesh: 384 MPI ranks; rheoMesh: one threaded process). On the same 3-layer setup, the latest build (f61a73b) completes in 27.3 s with Mesh OK, cutting faces with skewness > 3 from 53 → 27 (6.4× fewer than snappyHexMesh) and non-orthogonality > 60° from 39 → 37 (99.9 % ground, 98.6 % total wall faces / 99.9 % by area extruded). Measurement details are in the detailed history.

DrivAer 22.6 M cells: 2-node MPI performance (Miyabi-C, 2026-11-22)

Item Result
Total time (16 ranks × 14 threads) 112.09 → 99.69 s (sub-100 s, bit-identical)
Layer stage (wall bands + sparse shrink / morph) 45.67 → 30.94 s (−14.7 s, 1.48×, bit-identical)
MPI reproducibility 0/1/2/4/8-rank comparison 37/37 PASS (RHEOMESH_MPI_LBAND=1, bit-identical to serial)
Direct solver-ready decomposition --decompose N; 0.6 s for 6.3 M cells into 224 parts (19 s with decomposePar)
LBM export rheomesh-lbm writes the octree block forest as .orl (2.03 M cells for the half car, --wall-distance supported)

With the layer field (S1, opt-in), the 22.64 M-cell DrivAer is Mesh OK and gives Cd = 0.2431 with 448-rank simpleFoam (sting-free reference 0.244). Grid convergence (L1–L4), the v2611.4 / v2611.5 history and the assessment of the frozen method are in the detailed history.

Quick start

rheoMesh is a standalone CMake project in mesh/ of the OpenRheo repository (build and tests in mesh/README.md). Build and generate meshes in a compute-node job; on the login node only fetch sources.

git clone https://gitlab.com/masaz/openrheo.git
git clone -b Nagoya-University-Information-Technology-Center-20261218 \
    https://gitlab.com/OpenCAE/Supercomputer-OpenFOAM-Training.git
(cd Supercomputer-OpenFOAM-Training/Flow2/drivaerFastback/run && ./getGeometry)   # DrivAer geometry

# compute-node job (build + ctest + N5 mesh; give your own queue and group)
qsub -q <queue> -l select=1 -W group_list=$G \
    -v MESH=$PWD/openrheo/mesh,TRI=$PWD/Supercomputer-OpenFOAM-Training/Flow2/drivaerFastback/run/base/constant/triSurface,NS=5,WRITE=1,KEEP=1,OUT=$PWD/rheomesh_runs,RHEOMESH_SETTINGS=$PWD/openrheo/mesh/recipes/core.recipe \
    openrheo/mesh/tools/miyabi/job_castellate_c.sh

N5 gives about 2.86 M cells in about 12 s on Miyabi-C and passes Mesh OK. The GPU (GH200) build variants and job details are in "Building and running a sample on Miyabi" of the detailed history.

Design (structure, design principles, quality gates) · Results and comparisons (dated comparisons and figures) · Development log and detailed history (chronological log, comparison tables of other mesh generators, grid convergence) · AMR literature survey · Map