Mesh convergence in finite element analysis: when is your model good enough?

Every finite element result is an approximation. The mesh takes continuous physics and discretizes it into a finite number of elements, each with a simple interpolation function. Coarser meshes run faster but miss stress concentrations, steep thermal gradients, and local buckling modes. Finer meshes capture more of the physics but cost more computation, and eventually hit a point of diminishing returns where further refinement changes the result by less than the uncertainty in the input data.

A mesh convergence study is the systematic process of demonstrating that the mesh is fine enough for the result to be trusted.

The procedure

Start with a coarse mesh. Run the analysis. Record the quantity of interest: peak von Mises stress at a fillet, total reaction force at a support, maximum displacement at a loading point. Whatever the question demands.

Refine the mesh globally. Halve the element size everywhere. Re-run. Record the same quantity. Compare. If the result changed by more than a few percent, refine again.

After two or three global refinements, the critical regions become apparent. Switch to local refinement: bias the mesh so that elements are small near stress concentrations, contact surfaces, and geometric discontinuities, but remain coarse in regions where the gradients are gentle. This is where the analyst earns their keep. Knowing where to refine requires understanding the physics, not just running the software.

Continue until successive refinements change the quantity of interest by less than a defined tolerance. For structural assessments, 1 to 2 percent is a common threshold. For fatigue life predictions, where the result is exponentially sensitive to stress, tighter convergence may be necessary.

Common pitfalls

The most frequent mistake is converging on the wrong quantity. Global metrics like total strain energy or support reaction force converge quickly because they integrate over the entire domain. Local metrics like peak stress or contact pressure converge slowly because they depend on the mesh density in a small region. A model can show converged displacements while peak stress is still climbing with every refinement.

Sharp re-entrant corners produce stress singularities: the stress increases without bound as the mesh is refined. No amount of mesh refinement will produce a converged peak stress at a true geometric singularity. The correct approach is either to model the actual fillet radius (which eliminates the singularity), use a sub-modeling technique, or apply an analytical correction such as a stress intensity factor.

Over-refining regions that do not matter is a less obvious but equally costly mistake. A mesh with 500,000 elements that are uniformly small everywhere runs ten times slower than a mesh with 80,000 elements that are small only where they need to be, and produces no better answer at the critical locations.

The mesh is not the model

A converged mesh with incorrect boundary conditions, wrong material properties, or an inappropriate constitutive model gives a precise wrong answer. Convergence proves that the numerical approximation has stabilized. It does not prove that the model represents reality.

Every structural integrity assessment we produce includes a documented mesh convergence study: the quantity of interest, the mesh sizes tested, and the convergence curve. It is the minimum standard for a result we are willing to sign.