cfd-simulation-setup
Skill16 starsChanged 16 days agoMITView on GitHub
Sets up, runs, diagnoses, and reports OpenFOAM CFD simulations from a plain-language request, using the CFD Forge method — registered flow families with acceptance contracts, a closed action vocabulary, and deterministic validators that alone decide ACCEPT / CORRECT_AND_RERUN / REJECT / INCONCLUSIVE. Use when the user asks to set up, run, check, or trust a CFD simulation (OpenFOAM, nozzle, shock, step, bluff body, pipe/duct, airfoil), asks whether a CFD result is converged, mesh-independent or acceptable, or wants a CFD case built from a description.
In this repository34 other files it gives its agents.
- skills/cadsmith-text-to-cad/references/agents.mdAGENTS.md
- agentscad-fdm-dfamSkill
- am-alloy-screeningSkill
- am-literature-ragSkill
- cadsmith-text-to-cadSkill
- design-review-packageSkill
- dfm-reviewSkill
- drawing-gdt-reviewSkill
- engineering-calc-reportSkill
- equation-discoverySkill
- image2gcodeSkill
- lattice-inverse-designSkill
- linkage-synthesisSkill
- llm-3d-print-monitorSkill
- llm-mechanical-designerSkill
- lpbf-defect-predictionSkill
- lpbf-superresSkill
- machine-element-sizingSkill
- material-selectionSkill
- mechanical-teardownSkill
- mechanics-visualsSkill
- meltpool-geometrySkill
- parametric-part-cadSkill
- query2cad-freecadSkill
- quick-feaSkill
- rocketsmithSkill
- sketch-to-cadSkill
- sls-process-optimizationSkill
- stress-field-surrogateSkill
- thermal-fluids-calcSkill
- thermal-porositySkill
- tolerance-stackupSkill
- worked-problem-generatorSkill
- my-new-skillSkill
The file
We split this skill into the rules it gives an agent
6 rules. About 0 are common, 3 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 36
If it's out of scope, stop and report which parameter is outside the envelope.
- Very specific: line 49
Before solving, run python scripts/evidence.py mesh log.checkMesh.
- Very specific: line 70
Read only packet.json. Theory and reference values are stripped out, so you diagnose from the simulation's own evidence.
- Distinctive: line 79
Keep these apart: insufficient run time vs.
- Very specific: line 80
If the evidence can't separate competing hypotheses, propose REQUEST_DIAGNOSTIC.
- Distinctive: line 81
Don't invent measurements.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
CFD Simulation Setup (CFD Forge method)
Based on CFD Forge: A Physics-Constrained Agentic Framework for Autonomous CFD (Oza, Johnson, Barati Farimani, arXiv:2610.14104; code at github.com/aakashoza31/physics-constrained-cfd-agent).
The central rule: you propose, the code decides. You interpret the request, build the case, read the evidence, diagnose, and propose one action from a closed list. scripts/validate.py alone returns the verdict. In the paper, a model judging from the evidence alone accepted 12 of 13 detectable planted faults, and with the deterministic gates in place it accepted none. Never write ACCEPT yourself, never relax a threshold to make a run pass, and never present a number in the report that didn't come from an evidence file.
Requires OpenFOAM Foundation v14 (v11 or later works for foamRun), plus numpy. Everything except the solve itself runs without OpenFOAM.
Workflow
0. Pick or register a flow family
python scripts/contracts.py # list registered families
python scripts/contracts.py forward-step-2d # show a contractRegistered families: cd-nozzle (steady, quasi-1-D reference), forward-step-2d (transient shock, storage-aware mass closure), cube-stationarity (development gate for force histories), and generic-steady-incompressible (a template).
If the request fits no family, register one before running anything. Follow references/new_family.md and agree the thresholds with the user. A threshold chosen after seeing the result isn't an acceptance criterion. If the user won't fix tolerances up front, the best possible outcome is INCONCLUSIVE, and you should say so.
1. Interpret → stated values only
Write stated.json with only the values the request states, plus request_criteria (what the user wants to know: e.g. transient_health, grid_sensitivity, q1d_consistency) and requests_unsupported_physics: true if the request asks for physics the family excludes (viscosity in an Euler family, combustion, and so on). Leave everything else out. Don't guess.
python scripts/build_case.py spec --family forward-step-2d --stated stated.json --out spec.jsonspec.json labels every field stated_by_user or inferred_from_family_defaults. Show the user this table.
2. Scope gate
python scripts/validate.py scope --family forward-step-2d --spec spec.jsonIf it's out of scope, stop and report which parameter is outside the envelope. Offer the nearest admissible case or a new family. Don't stretch the family.
3. Build the case (values change, formulations don't)
python scripts/build_case.py forward-step --spec spec.json --case cases/m2 # copies $FOAM_TUTORIALS recipe, hashes it
python scripts/build_case.py nozzle --spec spec.json --case cases/nozIf patches_missing_in_template_fields in the builder output is non-empty, the template's 0/ files don't match the mesh patches. Fix that before solving. Boundary-condition types, numerics, and the physical model come from the family recipe. Don't hand-edit fvSchemes, fvSolution, or BC types. For a new family, write its builder first (see references/new_family.md).
4. Mesh gate, then solve
cd cases/m2 && ./Allrun.agent # blockMesh, checkMesh, foamRun with logsBefore solving, run python scripts/evidence.py mesh log.checkMesh. If mesh_ok is false, or a 2-D family has fewer than 2 solution directions, stop.
5. Extract evidence (deterministic; never typed by hand)
E=scripts/evidence.py
python $E log cases/m2/log.foamRun > ev_log.json
python $E mesh cases/m2/log.checkMesh > ev_mesh.json
python $E minmax cases/m2/postProcessing/fieldMinMax/0/fieldMinMax.dat > ev_mm.json
python $E closure mass_records.csv > ev_mass.json # transient families
python $E front line_y0p1.csv --step-x 0.6 > ev_front.json
python $E drift monitor.dat --col 1 --window 2e-3 > ev_drift.json # steady families
python $E growth forces.dat --col 3 --window 20 > ev_growth.json # bluff-body development
python $E merge --family forward-step-2d --criteria transient_health \
--state t=4,end_time=4,iteration=1,refinements=0,max_co=0.2 --out evidence.json ev_*.jsonreferences/function_objects.md has the controlDict function objects that write these records, including the per-step mass and boundary-flux records. Reference or theory comparisons are computed with scripts/theory.py and added to evidence after diagnosis.
6. Diagnose from the blind packet, and propose ONE action
python scripts/validate.py blind --evidence evidence.json --out packet.jsonRead only packet.json. Theory and reference values are stripped out, so you diagnose from the simulation's own evidence. Write proposal.json:
{"diagnosis": "INSUFFICIENT_RUN_TIME", "action": "CONTINUE_RUN",
"evidence_cited": ["monitor_drift 2.0% > 0.2%", "mass mismatch decreasing"],
"competing_hypotheses": [{"h": "local_underresolution", "for": "...", "against": "checkMesh clean"}],
"confidence": "high"}Diagnosis rules (from the paper's system instruction):
- Solver completion is not evidence of stationarity. A settled mean is not evidence of a developed flow; check growth of fluctuating components too.
- Keep these apart: insufficient run time vs. local under-resolution vs. numerical failure vs. boundary anomaly.
- If the evidence can't separate competing hypotheses, propose
REQUEST_DIAGNOSTIC. - Don't invent measurements. Cite packet keys.
- Choose the action from
permitted_actionsonly.
7. Gates decide
python scripts/validate.py action --evidence evidence.json --proposal proposal.json
python scripts/validate.py verdict --evidence evidence.json- Action REFUSED → don't execute it. You may propose once more, citing the refusal reason. A refused ACCEPT ends as
STOPPED_ACTION_REFUSED, which is INCONCLUSIVE. - Action APPROVED → execute exactly that action (continue, bounded extension, refine ×2 as a fresh solve), increment
iteration, and go back to step 5. The budget is 4 iterations. - The terminal verdict is whatever
verdictreturns. Record it unchanged.
8. Report
Use references/report_template.md: the request, the provenance table, case/mesh, the evidence table straight from evidence.json, each proposal with its ruling, the verdict with reasons, the reference comparison, and limitations. Use normalized units where the family is normalized; the paper's model wrongly attached SI units to normalized quantities. Keep every case file, log, proposal, and ruling so the run can be audited.
Things the gates catch that look fine
- Mesh-sensitivity requests with no registered cross-grid tolerance → INCONCLUSIVE, even when both grids pass.
CONTINUE_RUNafter the horizon is reached is refused; anEXTEND_END_TIMEbeyond the registered horizon is refused.- Refinement is refused unless the current solution is healthy (mesh or topology faults aren't resolution problems).
- Unequal inlet and outlet mass flux in a transient run isn't a conservation failure if the change in stored mass balances it.
- The OpenFOAM start-up line
sigFpe : Enabling floating point exception trappingisn't a fatal error. The extractor ignores it; the paper's validator once didn't.
The contracts protect against the model's errors, not against errors in their own diagnostics. If a verdict looks wrong, inspect the extractor, fix it, and re-analyse without re-solving. Report that you did.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

