FEA Stress Singularity: Mesh Convergence and Peak Stress Explained

How to spot a stress singularity in your FEA results

Developmech | Engineering Analysis & Simulation

The number that doubles when you “improve” your model

You run a finite element analysis on a loaded steel bracket. The peak von Mises stress at the inner corner reads 280 MPa — comfortably below yield. To be thorough, you refine the mesh, expecting the result to settle down. Instead, the peak climbs to 740 MPa. You refine again, and it climbs higher still.

Which number is correct? The instinct is to trust the finer mesh — smaller elements, more accuracy, surely? In this case that instinct is wrong. Neither number is the “true” stress, because the corner you are looking at does not have a true stress at all. You have hit a stress singularity, and chasing it with a finer mesh only makes the illusion more convincing.

The same sharp-cornered bracket meshed two ways. The peak stress more than doubles from coarse to fine — not because the model got better, but because the corner is singular.

What a singularity actually is

In the real world, stress is always finite. A physical part made of real material redistributes load through local yielding, microscopic radii and surface texture long before stress reaches infinity. The mathematics inside your solver knows none of that. It works on an idealised, perfectly sharp, perfectly elastic continuum — and in that idealised world, certain geometric and loading features produce theoretically infinite stress.

A stress singularity is exactly that: a point where the continuum solution predicts infinite stress. Because the value is infinite, the finite element method can never converge to it. Every time you halve the element size the discretised peak climbs, and it will keep climbing forever. The result is not a property of your part; it is an artefact of an idealisation in your model.

This is the single most important distinction in interpreting FEA results: a real stress concentration converges to a finite value as the mesh is refined; a singularity does not.

The three usual suspects

Singularities do not appear at random. They live at specific idealisations, and once you know where to look they become easy to anticipate:

Sharp re-entrant corners. A perfectly sharp internal corner — a fillet modelled with zero radius — is the classic case. Real corners always have some radius, however small. Model that radius and the stress converges; leave it sharp and it runs to infinity.
Concentrated point loads. A force applied to a single node acts over zero area, which means infinite pressure. The solver dutifully reports a spike that grows without bound under refinement.
Single-node constraints. The same logic applies to boundary conditions. Fixing or supporting a model at a single node creates a reaction over zero area — another singularity, this time hiding in your supports rather than your loads.

A modelled fillet (left) gives a smooth, convergent stress field. The same bend with a zero-radius corner (right) produces a singular hot-spot with no physical meaning.

A point load on a single node (left) drives stress toward infinity. Spreading the same force over a realistic contact area — here a washer (right) — gives a stable, physical result.

The definitive test: a mesh convergence study

Contour plots are seductive and misleading — a red hot-spot looks alarming whether it is real or not. The only reliable way to tell the difference is to refine the mesh in steps and watch the peak stress at the location of interest.

Run the same model at several element sizes and plot peak stress against element size. Two behaviours emerge. At a genuine stress concentration the curve rises and then flattens, settling on a stable value — that plateau is your physical stress. At a singularity the curve never settles; it keeps climbing as the elements shrink, announcing that there is no finite answer to converge to.

The convergence study is the tell. The green curve (a real feature) stabilises; the red curve (a sharp corner) diverges without bound. If your peak will not converge, it is not a stress.

A decision framework you can apply in two minutes

When a peak stress looks suspicious, three quick questions resolve almost every case. Does the peak change with mesh refinement? If it stabilises, it is real — full stop. If it keeps rising, is the hot-spot confined to a single node or element, or is it spread across several? A real concentration occupies a region; a singularity collapses to a point. And finally, is that point sitting on an idealisation — a sharp corner, a point load or a single-node support? If the answer to all three is yes, you are looking at a singularity, and the reported number should never make it into a design calculation.

A practical triage for any suspicious peak stress.

What to do about it

A singularity is not a dead end — it is a signal that the model needs to better reflect reality, or that you need a stress measure that ignores the idealisation. Depending on the situation:

Add a realistic fillet radius. The simplest fix. If the real part has a radius, model it; the stress will converge and you can design against the real peak.
Distribute loads and constraints. Replace point loads with pressure over a contact area, and single-node supports with constraints spread across a realistic region. This removes load- and BC-driven singularities at the source.
Use submodelling. When a fine local mesh would be too expensive globally, solve the coarse global model, then drive a refined submodel of the critical region from its displacement field.
Adopt a mesh-insensitive stress measure. For welds and fatigue, hot-spot (structural) stress methods and stress linearisation are designed to give stable, code-compliant results even in the presence of local singular behaviour.
Lean on St. Venant’s principle. Local idealisations only corrupt the field nearby. If your point of interest is a characteristic dimension or more away from the singularity, the spike at the idealisation can simply be ignored.

Why this matters

A singularity treated as a real stress leads to one of two costly errors: over-design, where parts are needlessly thickened or scrapped to chase a number that does not exist, or — far worse — a loss of trust in FEA itself, where sound analysis is dismissed because someone once “proved” a healthy part was failing. The discipline of separating real stress from numerical artefact is what distinguishes analysis that informs decisions from analysis that merely produces colourful pictures.

At Developmech, convergence-checked, validated FEA is the baseline — not an upsell. If you need stress analysis you can actually design against, we are glad to help.