Key takeaway
Reproducing one run does not establish grid convergence. Compare refinements of the same modeled problem and keep numerical sensitivity separate from physical validation.
A stable rerun can preserve a coarse approximation
A simulation archive can reproduce identical results using a fixed environment and still leave an important question unanswered: how much did the numerical result depend on the chosen grid? Repeating the same approximation tests repeatability. Changing discretization while retaining the modeled problem asks about numerical sensitivity. Neither question alone establishes agreement with physical observations.
NASA’s CFD tutorial distinguishes physical-modeling error, iterative convergence error and discretization error. Its grid-convergence discussion concerns how the numerical solution changes as grid spacing decreases. An owner describing simulation records should preserve those distinctions instead of labeling every successfully completed run validated.
Choose one quantity of interest before collecting convenient outputs. A pressure-recovery value, a local force and a maximum temperature are different claims. Identify the exact reported quantity, modeled geometry, boundary conditions and run procedure. An environment record remains useful, but it is only one part of this evidence.
Hold the modeled problem fixed across the refinement
This hypothetical archive records four steady-state CFD runs and a declared dimensionless pressure-recovery output. Relative spacing describes the intended refinement series, not a claimed relationship between total cell counts. Runs C, M and F share model and boundary-condition set S1. Run X changes a boundary condition to S2 while also using a finer grid.
For the S1 candidates, the output difference from C to M is 0.008 and from M to F is 0.002. Those smaller successive changes are a useful observation under the invented numbers. They do not establish that the grids lie in an asymptotic range, provide an error bound or make F physically correct. X is excluded from the isolated S1 grid comparison because it changes another part of the modeled problem.
| Run | Relative spacing | Toy output | Problem settings | Completed records-review decision |
|---|---|---|---|---|
| C | 4 | 0.960 | S1 | Candidate coarse member; stopping evidence required |
| M | 2 | 0.968 | S1 | Candidate medium member |
| F | 1 | 0.970 | S1 | Candidate fine member; decreasing changes observed |
| X | 0.5 | 0.973 | S2 boundary condition | Do not use as the next S1 refinement |
| S1 comparison | 4 to 2 to 1 | Differences 0.008 then 0.002 | Same declared problem | Eligible for specialist convergence review; no accuracy certificate |
Separate the three convergence questions
First, retain the iterative stopping evidence for each candidate run. NASA distinguishes the error left when an iterative method is stopped from the error associated with a discrete mesh. If a run has not reached a suitable iterative state for the task, a difference between grids can mix those effects. A status flag saying finished cannot explain the remaining numerical behavior.
Second, retain how the grid was refined and which output was compared. NASA’s grid-convergence tutorial recommends three levels to estimate convergence order and check asymptotic behavior in its Grid Convergence Index discussion. Merely possessing three files is not that check. The assumptions, refinement relationships and calculations need the responsible CFD specialist’s assessment.
Third, distinguish a limiting numerical value from physical truth. NASA expressly discusses the possibility of error between an asymptotic numerical result and the physical solution. More grid points do not by themselves resolve model choice, inappropriate boundary conditions or every usage error. The proposed owner review must therefore name the limited numerical claim rather than promise that refinement validates a design.
Deliver a numerical-evidence note with an explicit hold
The completed note for this example lists C, M and F as candidate members, excludes X from the S1 comparison and records the two output differences. It requests the grid-construction and iterative-stopping evidence. It holds any Grid Convergence Index, extrapolated value, asymptotic-range or physical-accuracy claim until those analyses are actually supported. No numerical method is chosen or engineering decision approved here.
A compact internal map should connect run identity, geometry/model version, boundary-condition set, mesh identity and refinement, solver/settings, stopping result, quantity of interest and output. Link the original run records rather than recreating a favorable history from a presentation. If different outputs converge differently, preserve that finding instead of retaining only the most stable result.
Inventory can describe which evidence layers exist; Due diligence can clarify the numerical question a recipient wants to investigate. Proprietary geometry, customer models, restricted solver material and sensitive design conditions need their own review before sharing. A fictional comparison can establish the proposed structure. A permissioned introduction does not grant simulation files, software rights or authority to rely on a result.