Archi-Civilarchi-civil.com
Finite Elements · FEM-04  |  Confidence: A
Figures are produced by engine/fem2d.js and pinned in tests/fem2d.test.js.

Home / Articles / Finite Elements FEM-04

Being Valid Is Not Enough

Distorted elements pass the patch test to machine precision at every level — that is, they remain valid. Accuracy nonetheless degrades monotonically. That distinction is the answer to why mesh quality metrics exist at all.

Confidence: A Engine: engine/fem2d.js Tests: tests/fem2d.test.js — 50/50 Prerequisite: FEM-02 Reading: ~15 min Last verified: 2026-09-10 Türkçe: bu yazının Türkçesi

1. Isoparametric mapping

Integrating the stiffness of a quadrilateral element directly over the real geometry is awkward: every element has a different shape. The isoparametric approach inverts this — integration is always carried out over the unit square, and the real element is obtained from that square by a mapping.

x(ξ,η) = Σ Ni(ξ,η)·xi   ·   y(ξ,η) = Σ Ni(ξ,η)·yi Hence the name "isoparametric": geometry and displacement are defined by the same shape functions. The parent element (−1 ≤ ξ,η ≤ 1) is always the same; what changes is the mapping.

2. What the Jacobian carries

The derivative of the mapping is the Jacobian matrix. Its determinant gives how much area in the parent element corresponds to area in the real element:

dAreal = detJ · dξ dη detJ ≤ 0 means the mapping is not invertible — the element is either ordered the wrong way round or folded onto itself. The calculation is then meaningless and our engine rejects the element.

Derivatives of the shape functions with respect to real coordinates are also computed through the inverse of the Jacobian. So the Jacobian carries the quality of the mapping into both the integration and the strain calculation.

3. detJ is constant only for parallelograms

This is counter-intuitive but exact, and directly testable:

ElementdetJ rangeJ ratioConstant?Skew
computing…
 J ratio = min(detJ)/max(detJ). A value of 1 means detJ does not vary within the element.
Findingcomputing…

The meaning: skew alone does not spoil detJ. What spoils it is opposite edges losing their parallelism — that is, trapezoidal distortion and taper. This distinction matters in meshing: a skewed but parallelogram mesh is better than a trapezoidal one.

4. Quality metrics

The metrics you see in the "mesh quality" reports of commercial programs come from this geometry:

These metrics are indicators; they do not give a strict error bound. The real question is: how much degradation produces how much error?

5. Does the patch test survive distortion?

In FEM-02 we saw that the patch test is the standard criterion for an element's validity. Now we displace the interior nodes of a regular mesh progressively and repeat the test. Boundary nodes are held fixed, so the problem stays the same.

DistortionWorst J ratioPatch test errorResult
computing…
 The distortion amount is how far interior nodes are displaced relative to the element size.
Result The patch test passes to machine precision at every level. That is, no matter how distorted it becomes, the Q4 element continues to reproduce a constant strain field exactly. The element formulation retains its validity.

6. But accuracy degrades

Apply the same distortion to a non-constant field — a bending cantilever. The reference from FEM-02: the Timoshenko solution.

DistortionWorst JMean JTip deflectionError
computing…
 16×4 mesh, Q4 with full integration. The mesh density never changes; only the node positions do.
Findingcomputing…

7. Necessary but not sufficient

Reading the two tables together gives the article's central point:

Passing the patch test → the element will converge
Passing the patch test → the element is giving a correct answer on the present mesh: NO The patch test examines a constant strain field. In real problems the field is not constant; there the quality of the element geometry comes into play, and the patch test does not measure it at all.

The practical translation: a program saying "my elements pass the patch test" does not mean your mesh is good. The two measure different things.

8. The direction is unsafe again

A pattern from this series repeats here. In FEM-02 displacement always converged from below; in FEM-03 the peak stress read low. Distortion works in the same direction: the element becomes stiffer and the deflection comes out even smaller.

So the error of a distorted mesh is added on top of the locking that is already present. The two do not cancel; they accumulate.

9. How meaningful is the metric?

Since the Jacobian ratio is an indicator, is it really related to the error? Place the two columns side by side:

Mean J ratioDeflection errorStress error
computing…
 The stress error is computed by comparing against the closed-form bending stress at each element's actual centroid.
A mistake was made in setting up this measurement When this comparison was first constructed, the stress was compared against the theoretical value at the element's nominal centre. But when the nodes move, the element centre moves with them; the reference was being evaluated at the wrong point and the stress error came out artificially low. The measurement was rebuilt using each element's actual centroid. The result is the table above.

In the corrected measurement, stress and deflection errors grow together; at heavy distortion the stress error slightly overtakes. As the Jacobian ratio falls both worsen — so the metric is meaningful, but it does not by itself provide an error bound.

10. What to do in practice

  1. A negative Jacobian is not a warning but a reason to stop. If the mapping is not invertible the result is meaningless; do not "skip the warning".
  2. Do not confuse passing the patch test with mesh quality. The two measure different things: the first your element formulation, the second your mesh.
  3. Build a regular mesh in the critical region. A distorted element where the gradient is large collides both with the slow stress convergence of FEM-03 and with the distortion error here.
  4. Prefer a parallelogram mesh to a trapezoidal one. Skew does not spoil detJ; loss of parallelism does.
  5. Do not use the metrics as absolute thresholds. There is no rule that "a J ratio above 0.7 is fine"; the metric is an indicator, and verification is done by a convergence study.
  6. Know the direction of a distorted mesh's error. Stiffer, that is smaller deflection, that is the unsafe side.

11. Test yourself

  1. What does the "iso" in the isoparametric approach refer to?
  2. For which element shapes does detJ stay constant? Can a skewed element belong to that set?
  3. If an element passes the patch test, can you say the result on that mesh is correct?
  4. In which direction does distortion change the deflection? How does that interact with locking?
  5. What does a negative Jacobian mean and why must the calculation stop?
  6. Can you state the error on a mesh with a Jacobian ratio of 0.6? Why not?
  7. Why was the stress measurement in this article wrong when it was first set up?

References

  1. Isoparametric formulation, Jacobian mapping and mesh quality metrics — standard finite element results, independent of any national code.
  2. The patch test — the standard criterion of element validity; see FEM-02.
  3. Reference solution: Timoshenko beam including shear deformation. Not the exact counterpart of the plane stress elasticity solution; an approximate reference, used as such.
  4. For course modules and MATLAB applications: Kwon, Y.W. & Bang, H., The Finite Element Method using MATLAB. The exposition, derivations and all figures here are our own; the source is cited for topic coverage.

All figures in this article are produced by engine/fem2d.js and separately pinned in tests/fem2d.test.js (50/50). In the distortion procedure only interior nodes are displaced; the boundary geometry is preserved so that the comparison stays on the same problem.

Open the element quality tool → FEM-05: quadratic elements → ← FEM-03: stress convergence Türkçe okuyun →
archi-civil.com — Being Valid Is Not Enough · FEM-04 · Confidence A · Printed:
The patch test examines your element formulation, not your mesh.