The Workflow StackCAE & simulation / Field guide 04
Run a CFD mesh independence study
A stable residual history is useful, but it does not show whether the answer changes with grid resolution. Check the outputs that drive the decision: pressure loss, heat transfer, maximum wall temperature, or another quantity of interest.
1. Lock the case definition
Keep geometry, physical models, boundary conditions, material properties, convergence targets, and reporting surfaces the same across meshes. Record cell count and where resolution changes near walls, interfaces, sharp gradients, and heat transfer surfaces. A changed turbulence model or boundary condition prevents a clean grid comparison.
2. Solve a refinement sequence
Run coarse, medium, and fine meshes with a consistent refinement strategy. Reduce iterative error enough that it does not mask the change caused by mesh refinement. For a transient case, check time-step sensitivity separately. Inspect conservation and local fields as well as the scalar outputs.
Worked comparison
Suppose predicted pressure drops are 8.0, 7.4, and 7.2 kPa on coarse, medium, and fine meshes. The medium-to-fine change relative to the fine result is (7.4 − 7.2)/7.2 ≈ 2.8%. That percentage describes this comparison; it is not a universal acceptance threshold or proof of asymptotic convergence. Repeat the comparison for thermal outputs, which may settle at a different rate.
3. Explain the mesh choice
Set the acceptable numerical influence from the actual decision margin. If two designs differ by only 1% in pressure drop, a 2.8% mesh change can alter their ranking. Document a chosen mesh, the remaining sensitivity, the computational cost, and any unresolved local behavior. If you estimate an observed order or grid convergence index, state the refinement ratio, assumptions, and whether the sequence is in an asymptotic range.
Use the validation checklist tool to keep the numerical study alongside physical validation evidence.
Further reading
NASA’s guide to spatial grid convergence explains refinement studies and the limits of inferring discretization error. Its time-step guidance covers transient cases.