Skip to content

rheoMesh — a map as a knowledge-based mesher

English | 日本語

« rheoMesh · development log

State of 2026-11-05 (the status table updated on 2026-11-06). The "compatible mode" was renamed snappyHexMeshLike on 2026-11-06. Every status here is provisional; the numbers are in the development log.

Principle: rules are frozen into settings, and error estimates decide whether they hold

rheoMesh is organised as a knowledge-based mesher. Expert rules of thumb are kept as versioned rule sets; error estimates on solved flows check whether a rule holds; learnt models and an LLM only propose rules; the deterministic mesher reproduces the same mesh from the frozen settings.

The literature supports this split. The knowledge-based meshers of the 1990s (e.g. Dolšak et al., inductive logic programming) learnt per-edge element counts from about ten expert meshes but had no way to check their rules, and did not spread. Commercial meshers (ANSA, Fluent Meshing, STAR-CCM+, Pointwise) keep knowledge as parameter files and replayable procedures, not as inference rules. The best machine-learnt AMR results reach parity with error-estimate-driven AMR. An LLM-driven CFD framework (CFD Forge, 2026) found LLM outputs not repeatable even at temperature 0 and moved every accept/reject decision to deterministic checks.

Five layers

layer role in rheoMesh
knowledge expert rules as rule sets, each with source, verification state and licence numbers from public guidelines (table below), snappyHexMesh's typical values, the owner's DrivAer settings
evidence check rules by error estimates on solved flows Jasak & Gosman's residual error estimate (rheomesh-eree), volume-weighted truncation-error (tau) estimates, AMR
learning learn what the rules miss, from computed results per-cell refinement-level prediction with error estimates as labels
policy choose which rules apply, how strongly, for a case a local LLM proposing only choices from a closed vocabulary and diffs, frozen into a settings file
execution build the mesh deterministically from the frozen settings octree, pillow, extrusion, the snappyHexMeshLike mode, global quality optimisation (CPU → GPU)

Four pillars and five layers

pillar layer knowledge evidence learning policy execution
octree + pillow + extrusion (the rheoMesh route) wall layers, building rules, gap rules layer thickness vs y+; wall-function cells left out of the indicator layer count / first height within the rule set pick the rule set by flow family (external, buildings, internal) deterministic pillow and extrusion; 2:1 steps eased by bands
snappyHexMeshLike mode snappyHexMesh's typical values (nCellsBetweenLevels 3, resolveFeatureAngle 30, maxNonOrtho 65 ...) as the reference rule set compare this set with others by the same error measure the reference set as the initial guess proposals as diffs from this set reproduce each stage from the dictionary (no source read)
global quality optimisation quality limits (checkMesh thresholds, growth rates) weights: skewness > non-orthogonality > size jumps none choose tolerances only global point optimisation with a barrier, fixed-order accumulation; GPU on GH200
AMR knowledge accumulation "in this situation, this level" rules tau and the Jasak residual, damping the jump at level interfaces; no Cd-targeted method level prediction, falling back to rules outside the training range choose rule sets from the accumulated ones, frozen in settings LBM block-forest mean fields as input (no published precedent found)

Expert rules (numbers read in the primary source)

field rule source
aircraft wall layers first cell y+ ≈ 1 (coarse), smaller on finer grids; growth below 1.25; at least 2 constant-spacing layers AIAA workshop gridding guidelines (checked in the SPICES-2019 full text)
aircraft grid families about 3× cells per level; far field about 100 reference chords same
road vehicle (DrivAer) 7 prism layers, first 0.75 mm, total 12 mm (wall functions) DrivAerML paper (CC BY 4.0)
buildings blockage below 3 %; lateral and top at least 5H, outflow at least 10H; at least 10 cells per building side; neighbour size ratio at most 1.3; evaluation height at the 3rd cell or above Tominaga et al., J. Wind Eng. Ind. Aerodyn. 96 (2008) (AIJ guidelines)
snappyHexMesh typical values nCellsBetweenLevels 3, resolveFeatureAngle 30, maxNonOrtho 65, maxInternalSkewness 4 ... OpenFOAM User Guide

Some rules do not map directly onto an octree: a 2:1 level step breaks "neighbour ratio at most 1.3", and AIAA's "about 3× per family level" does not match an octree level (8×); a translation layer is needed. For gaps, curvature and octree level steps hardly any rule backed by a primary source was found: these have to come from rheoMesh's own evidence layer (AMR and error estimates), and the resulting rules and validation data would be an asset neither commercial templates nor ML papers offer.

Status (2026-11-06)

pillar stage status
octree + pillow + extrusion layer thickness against the cell just outside the layer (switch); the dictionary's inGroups written (default); local refinement around bad faces and cell merging tried motorBike's layered pass 2 keeps one negative cell. One level more around the bad faces makes motorBike worse (after two rounds: failing faces 89 → 345, cells +105 %), and merging does not raise the time step. On the DrivAer rear, the LBM forest taken as a staircase mesh, merged and snapped is Mesh OK (2.44 M cells, max non-orthogonality 64.9°)
snappyHexMeshLike mode C0 (dictionary and nine quality criteria), C1 (castellation), S1 (snap), layers; from snappyHexMesh's outputs and logs: the layer-count ramp, the refinement band width, surface feature edges with the pocket rule, the post-layer repair and more (2026-11-06) On the course DrivAer (N = 5) 1.9 times (layers) and 5.6 times (no layers) faster than snappyHexMesh, Mesh OK; fewer faces over 50° and over skewness 3 and closer feature lines than snappyHexMesh. The sum of pockets and cuts into the solid is not reduced. motorBike S1 numbers in the development log
global quality optimisation G0 (kernel), G1 (replacing today's shape optimisation, behind a switch), G2 (convergence) Cutting today's shape optimisation from 20 to 3 sweeps brings motorBike from about 6.3 s to 2.33 s with no loss of quality. The global optimiser is an optional polish (max non-orthogonality 1-3° lower on DrivAer and cavitatingBullet). With the switch off the meshes are bitwise unchanged
AMR knowledge accumulation K0 (tools), K1 (5 cases, 3 cycles) 4 cases completed 3 cycles with Mesh OK (injectorPipe's cycle-1 mesh failed). Best cycle-0 features: lift 4.1-5.6 (cavity 2.58). EREE becomes the primary label for external and building flows
(shared) one GH200 node GPU simpleFoam and CPU rheoMesh together mutual slowdown 0.9-3.6 %; at motorBike size the meshing (17.5 s) takes longer than a coarse flow field and does not hide
(shared) the course DrivAer (N = 5) snappyHexMesh, snappyHexMeshLike (no layers and layers) and the native method (layers) from one dictionary; development version 38d99e21 Flow2 Type I, one node (384 cores), faster of two runs, snappyHexMesh as blockMesh + decomposePar + its process time. Layers: snappyHexMesh 41.4 s, snappyHexMeshLike 21.5 s, native 47.4 s; no layers 33.4 s and 5.9 s. All Mesh OK. Wall without snappyHexMeshLike layers where snappyHexMesh has them: 0.15 % of the body. Cd: snappyHexMesh 0.2564, snappyHexMeshLike 0.2697, native 0.2952 (TUM experiment 0.243-0.252); the native gap is under investigation (higher wake k, first-cell y⁺ 23 against snappyHexMesh's 37)

Next steps

  1. A rule-set file format with source, verification state and licence per rule; a first set from the checked numbers above and snappyHexMesh's typical values only.
  2. Check the octree translation rules (2:1 steps vs ratio 1.3, grid families, ground layers) on a single building: separation position against the band width.
  3. Evidence layer: tau estimates with the parent cells as the 2h grid, and the Jasak residual, written out and fed to AMR. The AMR knowledge study continues on more cases with EREE as the primary label.
  4. A regression run before the shape optimisation's default becomes 3 sweeps. GPU global optimisation: fixed accumulation order; same result across thread counts and CPU/GPU.
  5. Candidates being looked into (no results yet): joining the wall to the octree (the owner's choice: tile-wise refinement → gathering into the layer band → global optimisation of the whole shell), flow-informed layer sizing (potentialFoam with an integral boundary layer, then a panel method), a coarse LBM mean field as the AMR's cycle-0 field.
  6. LLM policy layer: closed-vocabulary choices and diffs only; record the differences of the settings file over repeated identical inputs.

Cd is not a target at any stage.

Licences

Numbers are cited as facts with their sources; texts are not copied. The ERCOFTAC Knowledge Base Wiki reportedly reserves AI/ML-training and text-and-data-mining rights, so it is used only for reading and validation comparisons, not as training data. Datasets used for learning (e.g. DrivAerML, CC BY-SA 4.0) are used only after their share-alike terms are checked.

Main references

  • P. J. Frey, P.-L. George, Mesh Generation: Application to Finite Elements, 2nd ed., 2008
  • J. F. Thompson, B. K. Soni, N. P. Weatherill (eds.), Handbook of Grid Generation, CRC Press, 1999
  • H. Jasak, PhD thesis, Imperial College London, 1996; H. Jasak, A. D. Gosman, Numer. Heat Transfer B 39 (2001) 1-19; Computers & Fluids 32 (2003) 223-248
  • A. Syrakos et al., truncation-error-based refinement (J. Comput. Phys. 2012); gradient schemes and skewness (Phys. Fluids 29, 2017)
  • B. Dolšak, I. Bratko, A. Jezernik, AI EDAM 12 (1998) 95-106
  • Y. Tominaga et al., J. Wind Eng. Ind. Aerodyn. 96 (2008) 1749-1761
  • J.-S. Camier et al., J. Comput. Phys. 474 (2023)
  • MeshingNet (arXiv:2004.07016), ASMR++ (arXiv:2406.08440), CFD Forge (arXiv:2610.14104), DrivAerML (arXiv:2408.11969)