Grid independence is evidence, not a checkbox. If the mesh sequence is not geometrically consistent or the quantity of interest is not in an asymptotic trend, a simple “fine mesh changed by less than X%” statement can be misleading.
Problem: the solution changes unpredictably with mesh refinement
Typical symptoms are a drag, pressure drop, temperature or efficiency that improves from coarse to medium and then reverses on the fine grid; a large change in observed order; or a “fine” mesh that gives a worse comparison to experiment.
These patterns can indicate non-asymptotic refinement, different local mesh topology, unresolved separation, iterative error, or simply that the quantity of interest is sensitive to a region that was not refined consistently.
Build a mesh sequence that is actually comparable
- Preserve the same model.
Keep geometry, physics, boundary conditions, turbulence treatment, numerics and report definitions fixed.
- Refine the influential regions consistently.
Changing only one local control can alter the flow topology rather than create a clean global refinement sequence.
- Record effective refinement ratio.
Use cell count only as a screening proxy; when possible, track representative spacing in the region that controls the result.
- Converge each mesh sufficiently.
Iterative error should be small compared with the mesh-to-mesh difference.
Screen a three-grid sequence
Estimate observed order, extrapolated value and GCI from coarse, medium and fine results.
Recognize oscillatory or non-asymptotic behavior
If the quantity of interest changes direction between successive meshes, the sequence is oscillatory. That does not automatically mean the CFD is unusable, but a monotonic Richardson/GCI interpretation is no longer strong evidence by itself.
Oscillatory trend
Investigate whether a separation point, shock, free surface, interface or local recirculation is moving between meshes.
Very high apparent order
Large p values can come from cancellation or small successive differences rather than genuine high-order convergence.
Weak refinement ratio
Meshes that are too similar provide little signal above iterative and modelling noise.
Different mesh topology
A change in prism collapse, wake refinement or feature capture can make the meshes qualitatively different.
Diagnose what is controlling the mesh sensitivity
- Map the difference field.
Compare pressure, velocity, wall shear, heat flux or phase fraction spatially instead of looking only at one scalar.
- Inspect the quantity definition.
Make sure reports integrate the same surfaces, reference values and averaging windows on every mesh.
- Check near-wall consistency.
Refinement can change y+ and wall-treatment behavior, which makes the physics model itself respond differently.
- Separate spatial from temporal sensitivity.
For transient cases, a finer mesh may require a smaller timestep to keep temporal resolution comparable.
- Check conservation.
A mesh with lower discretization error is not credible if mass or energy imbalance remains material.
Turn the study into defensible verification evidence
Document the mesh definitions, refinement ratios, iterative convergence, quantity of interest, observed order, GCI and any non-monotonic behavior. If the trend is not asymptotic, state that limitation rather than hiding it.
For engineering decisions, combine mesh evidence with timestep sensitivity, conservation and validation. A low GCI only addresses one part of numerical uncertainty.
Common mistakes
- Calling the finest mesh “grid independent” because it has the most cells.
- Changing turbulence model, wall treatment or boundary conditions between meshes.
- Comparing meshes that refine different regions or use different prism behavior.
- Ignoring iterative error when mesh-to-mesh differences are small.
- Using a single global cell-count ratio as proof of geometric similarity.
- Reporting only percent change without examining trend direction and local physics.
Related Abecator resources
Need help interpreting a difficult mesh study?
Submit the three mesh definitions, quantity of interest, convergence history and non-confidential plots. We can review whether the sequence supports a credible engineering conclusion.