Validate and Calibrate Airflow Simulation
Use this guide before an airflow field or coupled physics result is accepted as evidence for an engineering comparison. Validation should answer two separate questions: whether the solver behaves correctly for known tests, and whether the selected model and inputs represent the project question well enough for the intended decision.
Prerequisites
| Requirement | Why it matters |
|---|---|
| Defined decision and metric | Validation tolerances need to follow the decision being reviewed. |
| Fixed input versions | Geometry, sources, boundaries, and object properties must remain traceable. |
| Engineering owner | An accountable owner selects evidence, tolerances, and escalation criteria. |
| Reproducible run record | Solver, grid, runtime, hardware, and post-processing settings must be available for review. |
Evidence ladder
Validation layers
| Layer | Review question | Example evidence |
|---|---|---|
| Numerical checks | Does the implementation preserve its stated invariants? | Repeatable tests, stable fields, unit and coordinate checks. |
| Conservation | Is source momentum accounted for in an appropriate control region? | Integrated momentum or force budget with a stated tolerance. |
| Convergence | Does the decision metric stabilize as the grid or run window changes? | Coarse, medium, and fine comparisons at agreed review locations. |
| Public benchmark | Does the method reproduce a documented reference case within the intended use range? | Velocity and pressure profiles compared with experimental data or an established solver. |
| Project calibration | Do source assumptions and object response align with approved evidence? | Measurements, equipment data, or controlled observations with provenance. |
| Runtime coupling | Does the same reviewed field produce credible force, torque, and object behavior? | PhysX or Newton run records, statistics, and failure analysis. |
Define the decision before the tolerance
Validation criteria should follow the decision. An early source-placement study may need stable direction and relative ranking. A moving-object study may need force, torque, orientation, or throughput distributions. A final safety, thermal, contamination, or regulated decision requires discipline-specific methods and acceptance criteria.
Record the metric, review location, operating range, required tolerance, and decision owner before tuning the model.
Run conservation and convergence checks
Use control surfaces or force budgets that match the model boundary. State where conservation is expected and where walls, openings, or imposed boundary conditions change the budget.
For convergence, compare at least the metrics used in the decision. A visually similar flow field is not enough. Review profiles, integrated force or torque, scenario ranking, and any near-wall quantity that affects the conclusion. If the result changes materially with resolution, keep it exploratory or refine the model.
Use public benchmarks correctly
Choose a benchmark with comparable flow structure and published experimental evidence. Record geometry, boundary conditions, nondimensional parameters, grid, runtime, hardware, and post-processing method. Compare multiple quantities when possible rather than reporting one favorable value.
A benchmark result applies to the tested configuration. When reporting speed or accuracy, include the configuration, hardware, measured quantities, and intended scope.
Calibrate without hiding assumptions
Calibration may adjust an effective source strength, loss factor, surface-response coefficient, or another project parameter. Preserve:
- the uncalibrated input and its source;
- the calibrated value and method;
- the evidence used for fitting;
- a separate validation observation where available;
- the operating range in which the calibration may be reused;
- the reviewer and approval date.
Create a new calibration version when geometry, source hardware, operating range, object properties, or boundary assumptions change.
Validate physics-runtime coupling
Review the airflow field independently before diagnosing object motion. Then confirm object geometry, surface sampling, mass, inertia, center of mass, collision representation, time step, and force or torque limits.
PhysX and Newton are separate runtime implementations. Contact handling, collision behavior, and numerical stepping can differ. Validate each runtime against the decision metrics and use distributions or aggregate outcomes where individual trajectories are sensitive.
Decision record
Every accepted study should retain:
| Record | Content |
|---|---|
| Question | Decision, alternatives, and intended use. |
| Inputs | Scene, geometry, sources, object properties, and provenance. |
| Model | Solver path, resolution, boundaries, runtime, and versions. |
| Evidence | Conservation, convergence, benchmark, calibration, and runtime results. |
| Limitations | Known weak regions, excluded effects, and valid operating range. |
| Decision | Accepted, accepted with conditions, revise, or escalate. |
| Ownership | Engineering owner, reviewer, and review date. |
Escalate the study when needed
Escalate to a dedicated engineering CFD, thermal, structural, contamination, or field-test workflow when the decision depends on effects outside the reviewed model, near-wall detail has not converged, calibration evidence is insufficient, or the consequence of error exceeds the approved use boundary.