How to perform a CFD mesh-independence study

What a mesh-independence study proves

A mesh study evaluates how strongly the numerical result changes as the spatial discretization is refined. It is part of solution verification: determining whether the chosen mesh resolves the computed quantity sufficiently for the intended conclusion.

Plan the mesh study before running the final cases

The study should be connected to the paper's research question. Decide in advance which outputs matter, which mesh features will be refined and how convergence will be judged.

• Identify the primary quantity or quantities of interest.
• Define coarse, medium and fine mesh strategies.
• Keep geometry, physics, boundary conditions and numerical schemes unchanged.
• Define convergence requirements for each mesh.
• Use consistent report definitions, sampling locations and averaging windows.
• Budget enough computation to add a fourth mesh if the trend is unclear.

The working mesh used in the full parameter study does not always need to be the finest. It must be fine enough that remaining spatial uncertainty does not change the study's engineering conclusions.

Use three systematically related meshes

Two meshes provide one difference. Three meshes provide two differences, which makes it possible to examine whether the result is approaching a limit and to estimate observed convergence order.

The meshes must represent the same model. Refinement should be systematic enough that differences can be interpreted as spatial-resolution effects rather than unrelated meshing decisions.

Keep these features comparable

• Domain boundaries and geometric defeaturing
• Surface-size logic and volume-refinement regions
• Prism-layer strategy and wall-treatment intent
• Gap and curvature resolution
• Mesh topology where practical
• Interfaces, porous regions and rotating regions

Define a meaningful refinement ratio

For geometrically similar structured grids, refinement ratio can be based directly on characteristic cell spacing. For general three-dimensional unstructured meshes, a representative spacing is often estimated from domain volume and cell count.

The fine-to-medium refinement ratio is defined using the coarser spacing divided by the finer spacing, so it is greater than one:

A ratio too close to one produces small differences that may be overwhelmed by incomplete iterative convergence or sampling noise. A very large ratio may move the coarse mesh outside the asymptotic range. Use a controlled, documented ratio suitable for the geometry and available resources.

Choose quantities that support the paper's conclusions

Mesh independence should be evaluated using the outputs that determine the engineering conclusion—not only residuals, maximum cell quality or a convenient global average.

Use more than one quantity when the paper makes several important claims. A mesh may be adequate for pressure drop but inadequate for peak wall temperature or local separation.

Converge every mesh before comparing them

A mesh comparison is invalid if the differences mainly reflect incomplete iterative or statistical convergence. Each mesh should meet the same convergence and averaging criteria.

• Residuals reach a stable level consistent with the solver and physics.
• Primary engineering monitors stop drifting.
• Mass and energy imbalance are acceptably small.
• Transient statistics use comparable physical time and sampling windows.
• Periodic cases cover enough cycles to establish repeatability.
• Initialization effects are excluded from averages.

Compare the mesh results transparently

Present the three mesh definitions and outputs together. Avoid using only percentage differences without showing the underlying values.

A useful comparison table contains

• Coarse, medium and fine mesh names
• Cell counts and refinement ratios
• Relevant mesh controls and wall resolution
• Values of every primary quantity of interest
• Relative differences between successive meshes
• Observed convergence type
• Extrapolated value and GCI where applicable

Plot the quantity against representative spacing or cell count. The curve communicates whether changes are decreasing regularly, oscillating or remaining unresolved.

Calculate observed order, Richardson extrapolation and GCI

For three meshes with a common refinement ratio and monotonic convergence, the observed order can be estimated from successive solution differences.

The fine-grid Richardson-extrapolated value is:

A commonly used fine-grid Grid Convergence Index is:

When refinement ratios are unequal, the observed order generally requires an iterative solution rather than the simplified equal-ratio expression. Use the actual mesh spacings and do not quietly substitute cell-count ratios.

What to do with non-monotonic mesh convergence

Non-monotonic behaviour occurs when the medium result does not lie between the coarse and fine results or successive changes reverse direction. This may indicate oscillatory convergence, competing errors or inconsistent mesh families.

Investigate before adding more decimals

• Confirm all three cases reached comparable iterative or statistical convergence.
• Check that report definitions and sampling locations are identical.
• Review whether prism layers, wall treatment or y+ regime changed.
• Check whether local refinement moved a discontinuity across the sampling location.
• Assess whether geometry representation changed with the surface mesh.
• Check timestep sensitivity if spatial and temporal errors interact.
• Add a fourth mesh or redesign the refinement sequence.

Do not remove an inconvenient mesh solely to manufacture a smooth trend. Report the behaviour honestly and explain the additional investigation.

Global convergence can hide local mesh dependence

A stable global quantity does not guarantee that local gradients, hotspots, separation, jets or interfaces are adequately resolved. The required evidence depends on the paper's claims.

Use targeted local refinement while maintaining a traceable mesh family. If different regions have different refinement needs, document the strategy and show that the quantities used in the paper are insensitive to further relevant refinement.

How to report the study in a journal paper

Methods section: describe the design

State how the three meshes were generated, which controls changed, how representative spacing and refinement ratios were defined, how each solution was converged and which quantities were evaluated.

Results section: show the evidence

Present a table of mesh characteristics and outputs, a convergence plot, successive differences, observed order, extrapolated value and GCI where the assumptions are appropriate. Explain the selected production mesh and why its remaining uncertainty is acceptable for the paper's conclusions.

Replace placeholders with actual evidence and modify the statement if convergence is non-monotonic. Never copy a claim of “mesh independence” that the data do not support.

Common reviewer objections—and how to prevent them

• “Only two meshes were tested.”

  Add a third systematic level so convergence behaviour and uncertainty can be assessed.

• “The refinement method is unclear.”

  Report actual global and local controls, representative spacing, refinement ratios and wall-mesh changes.

• “Cell counts changed, but critical regions did not.”

  Show refinement in the regions that influence the quantity of interest.

• “Residuals are shown, but solution convergence is not.”

  Add engineering monitors, conservation and statistical/periodic convergence evidence.

• “The selected variable is not relevant.”

  Evaluate the outputs used to make the paper's central claims.

• “Mesh independence is called validation.”

  Separate numerical verification from comparison against physical reference evidence.

• “The trend is non-monotonic.”

  Investigate mesh consistency and add evidence rather than forcing a standard convergence claim.

Complete mesh-independence checklist

• Define the paper's primary quantities of interest.
• Create at least three systematically related meshes.
• Document global and local refinement, prism layers and y+ intent.
• Keep geometry, physics, boundaries and schemes consistent.
• Fully converge or statistically resolve every mesh.
• Check conservation and numerical plausibility.
• Calculate representative spacings and refinement ratios.
• Compare raw values and successive relative differences.
• Classify convergence as monotonic, oscillatory or divergent.
• Calculate observed order, extrapolation and GCI only when appropriate.
• Check local fields where the paper makes local claims.
• Select the production mesh based on evidence and study objectives.
• Report the method, data, uncertainty and limitations transparently.