CFD Meshing — Best Practices and Quality Guide

Big difference. In FEA you can tolerate high aspect ratios in the far-field as long as stress gradients are small there — the element is just computing average strains. In CFD, high aspect ratios are actually fine and sometimes essential (you deliberately use very thin, high-aspect-ratio cells in the boundary layer). But what kills CFD accuracy is something FEA doesn't care about much: non-orthogonality — how much the face normal deviates from the vector joining the two cell centers. Non-orthogonal faces create cross-diffusion errors in the discretization of the Laplacian (pressure Poisson and diffusion terms). OpenFOAM's checkMesh reports non-orthogonality as a key metric; above 70° is dangerous, above 85° will likely diverge.

Automatic unstructured tet meshes are fine for getting started, but structured hex meshes can give 2–5x better accuracy per cell in well-aligned flows because hex cells have zero non-orthogonality and the numerical diffusion aligns with the flow direction. The trade-off is meshing complexity — creating a high-quality multi-block structured hex mesh for a complex shape like a car or turbine blade can take days of manual work. For most industrial cases, a hybrid approach wins: structured hex/prism layers in the boundary layer where quality matters most, and unstructured hex (from snappyHexMesh) or polyhedral cells in the bulk. Pure tet is only acceptable if you have very good mesh refinement controls and you run a correction scheme for the large non-orthogonality that tet meshes produce.

The formula is $y_1 = y^+ \nu / u_\tau$, where $u_\tau = \sqrt{\tau_w/\rho}$ is the friction velocity. You estimate the wall shear stress $\tau_w$ from a flat-plate correlation before running: $\tau_w = 0.0594 \rho U^2 Re_L^{-0.2}$ for turbulent flow. For a car at 30 m/s in air: $\nu = 1.5 \times 10^{-5}$, Re ≈ $5 \times 10^6$, so $\tau_w \approx 0.73$ Pa, $u_\tau \approx 0.78$ m/s. For y+ = 1: $y_1 = 1 \times 1.5 \times 10^{-5} / 0.78 \approx 19$ μm. That's a very thin first cell — 0.019 mm — which is typical for automotive external aero.

The formal method is to run at least three mesh levels — coarse, medium, fine — with consistent refinement ratio (typically r = 1.5 or 2 in all three directions, so roughly 3.4× or 8× more cells per level). Extract your key output — drag, Nusselt number, pressure drop — from each mesh. If the solution is in the asymptotic convergence range, Richardson extrapolation gives you an estimate of the "exact" (zero-mesh-size) solution: $f_h = f_{fine} + (f_{fine} - f_{medium})/(r^p - 1)$ where p is the observed convergence order. If your two finest meshes differ by less than 2%, you're usually in good shape to report the fine mesh result. Also report the Grid Convergence Index (GCI) as the uncertainty metric.