Mesh Quality Metrics and What They Do Not Tell You — Troubleshooting Guide

What Mesh Quality Metrics Actually Measure

Here are the main quality metrics, paired with what really goes wrong when each one is bad.

The essential point is where these sit: they are necessary conditions confirming that the discretization is not broken, not a measure of whether the solution is right. Miss that distinction and you walk straight into the maze of Symptoms 1 and 2 below.

Symptom 1: Every Quality Check Passes, Yet the Result Is Wrong or Will Not Converge

Symptom 2: A Handful of Bad Elements Simply Will Not Go Away — How Hard Should You Fight?

Driving the worst value to zero on a mesh of several million elements is a cost-benefit decision.

1. Look at where they are — if the bad elements sit in the region you evaluate (a stress concentration, the boundary layer), fixing them is mandatory. In a remote, unimportant region they are usually acceptable
2. Look at how many and how they are distributed — a few isolated elements, or a sheet or band of them? A connected cluster points to a systematic problem in the geometry or the size field, and individual repairs will not remove it
3. Measure the effect on the solution — when in doubt, compare results against a version with that region remeshed. If the quantity of interest does not move, those bad elements do no real harm — and this demonstration is the strongest justification for accepting them
4. Know your solver's tolerance — FV codes are sensitive to non-orthogonality and skewness, explicit schemes to the smallest element size (the time step), second-order FEM to the Jacobian. Prioritize against the weak spot of the solver you actually use

Symptom 3: The Same Mesh Reports Different Quality Values in Different Tools

This is normal, because the definition of each metric differs from tool to tool. Skewness alone circulates in several definitions — normalized equiangular skew, normalized volume ratio, angle-based — and it is not unusual for the same element to score 0.6 in one and 0.9 in another. The Jacobian likewise changes depending on whether it is the min/max ratio of the corner values or an evaluation at the Gauss points. Two rules apply in practice: ① state a quality criterion together with "in which tool, under which definition" (write the defining formula into your in-house standard), and ② align the mesher's pass criteria with the solver's own check criteria (so you never get the twist of "mesher passes, solver warns"). Give up on absolute comparison of the numbers and manage relative values within a single definition, and the confusion disappears.

Symptom 4: The Aspect Ratio Warning on the Boundary Layer Mesh Will Not Clear

This is the classic warning you are meant to leave alone. The solution in a wall boundary layer varies steeply along the wall normal and gently along the flow, so flat elements with aspect ratios in the hundreds or thousands are in fact optimal (meeting a y+ requirement with isotropic elements gives an unrealistic element count). The correct way to handle it in quality management is to ① separate the boundary layer zone from the quality statistics and manage it not by aspect ratio but by continuity in the normal direction (collapsed or punched-through layers) and growth rate, and ② count its warnings separately from the aspect ratio warnings in the general regions. There are exceptions, though: flat elements near a separation point or crossing a shock become an error source through directional mismatch — judge those case by case on the principle that "anisotropy is only justified when it is aligned with the anisotropy of the solution".

Symptom 5: The Automatic Mesher Mass-Produces Low-Quality Elements

When no amount of tuning in the mesher settings fixes it, the root cause is almost always in the geometry data.

• Sliver faces and tiny edges — a CAD-derived face 0.01 mm wide forces elements of that size. Geometry clean-up (face merging, edge suppression) is the real cure
• Mismatched contact and shared faces — small gaps or interpenetrations between parts create bands of thin elements. Impose shared topology on the assembly
• Abrupt change in the size field — the ratio of local refinement to the global size is too large and the transition breaks down. Insert intermediate sizes that respect the growth rate
• Excessive detail such as fillets and threads — suppress features the analysis does not need (defeaturing). "Eight out of ten mesh problems are geometry-handling problems" is a rule of thumb shared across companies

Designing a Practical Quality Gate — Automate Detection, Contextualize Judgement

Here is the recommended structure for systematizing quality management. ① Automatic gate (mandatory): zero negative Jacobians, a worst-case skewness threshold, layer continuity in the boundary layer — any mesh that breaks these is rejected unconditionally. ② Statistical report: output histograms of the metrics and the location (coordinates) of the worst values, so that a review can see where the bad elements are. ③ Contextual judgement: strict quality in the evaluated region, looser far away — zone-specific criteria rather than one uniform bar. ④ Connection to the convergence study: passing the gate is the starting line, and quality assurance is completed by confirming mesh convergence of the quantity of interest (GCI). With those four stages you avoid both "the criteria are so strict that no work gets done" and "they are so loose that we have an accident".

Checklist

• Zero negative Jacobians and inverted elements (this one is unconditional)
• Checked the location of the worst elements (fix if in the evaluated region, confirm the effect if far away)
• Documented the definition of each quality metric (which tool, which formula)
• Boundary layer flat elements managed separately from the general criteria
• Bands and sheets of low quality attacked at the root with geometry clean-up
• Mesh convergence study carried out after the quality gate was passed

Related: GCI troubleshooting, h-refinement, checking and diagnosing solver convergence.