2. Finite element error sources
Every FEA result is an approximation. The engineering question is not whether error exists, but which error dominates and whether it can be reduced by refining the mesh. Confusing a modeling error with a mesh error leads to wasted compute time at best, and to a wrong design decision at worst. This section builds a working taxonomy of error sources and gives a practical procedure for separating mesh-induced error from modeling error.
2.1 A Three-Layer Error Taxonomy
It is useful to decompose the total solution error into three additive contributions:
- Idealization error — the difference between the real physical structure and the mathematical model you chose (geometry simplifications, boundary conditions, material laws, load representation).
- Discretization error — the difference between the exact solution of that mathematical model and the finite element solution on a given mesh. This is the part that mesh refinement attacks.
- Numerical (round-off and solver) error — floating-point accumulation, ill-conditioning, and iterative solver tolerance. Usually small, but it can dominate in contact, near-singular constraints, or badly scaled models.
The critical insight: mesh refinement converges to the solution of your idealized model, not to reality. If the idealization is wrong, a beautifully converged mesh study will converge confidently to the wrong answer.
2.2 Where Discretization Error Comes From
Discretization error has several distinct mechanisms, and they respond differently to refinement:
| Mechanism | Typical symptom | Refinement response |
|---|---|---|
| Poor shape function fit (low-order elements in bending) | Overly stiff response, underestimated deflection | Improves with p-refinement or finer h |
| Stress singularity (re-entrant corner, point load, rigid constraint) | Stress grows without bound as mesh refines | Does not converge — must change model |
| Insufficient through-thickness integration | Shear locking, hourglassing | Improves with element formulation change |
| Poor aspect ratio / distorted elements | Local stiffness errors, noisy stress | Improves with mesh quality repair |
| Inadequate load or BC application region | Local stress concentration artifacts | Improves with Saint-Venant-compliant patch |
Note the second row carefully. A stress singularity is a property of the mathematical model, not the mesh. Refining the mesh makes the peak stress diverge. If you see stress rising monotonically with refinement and never settling, you are looking at a singularity, not a mesh error.
2.3 Where Idealization Error Comes From
Idealization error is introduced before the mesh is ever drawn. Common sources in industrial practice:
- Geometry defeaturing — removing fillets, small holes, or bosses that actually carry load or act as stress risers.
- Dimensional reduction — using shells where the structure is thick, or beams where local bending matters.
- Boundary condition assumptions — fully fixed vs. pinned vs. elastic support; the difference can exceed 50% in deflection.
- Material idealization — linear elastic where plasticity, creep, or hyperelasticity governs.
- Load representation — point loads instead of distributed pressure, or omitting preload and thermal strain.
These errors do not shrink with mesh refinement. They are the reason a converged solution can still be wrong.
2.4 Separating Mesh Error from Modeling Error
The practical test is a controlled convergence study. The procedure below is written for a linear static stress analysis, but the logic transfers to thermal and modal work.
- Pick a scalar quantity of interest (QoI) — e.g., peak von Mises stress at a defined location, or tip deflection. Avoid picking a singular point.
- Run the model on a sequence of meshes with global element size halved each time:
h, h/2, h/4, h/8. Keep element type, formulation, and BCs identical. - Record the QoI and compute the relative change between successive meshes:
Δ = |Q(i+1) − Q(i)| / |Q(i+1)|. - Plot QoI vs. mesh density (or vs. 1/DOF). Look at the shape of the curve.
Example convergence table (illustrative, not measured):
Mesh Element size DOF Tip deflection (mm) Δ (%)
M1 4.0 mm 18,420 2.41 —
M2 2.0 mm 71,300 2.68 10.1
M3 1.0 mm 281,000 2.79 3.9
M4 0.5 mm 1,120,000 2.83 1.4
Interpretation:
- Monotone convergence with shrinking Δ → discretization error is dominant and is being resolved. Extrapolate (Richardson) to estimate the converged value.
- QoI keeps growing, Δ stays roughly constant or grows → singularity or idealization issue. Refining further is wasted effort.
- QoI oscillates non-monotonically → suspect mesh quality, contact chattering, or solver tolerance too loose.
Q∞ ≈ Q_fine + (Q_fine − Q_coarse) / (r^p − 1), where r is the refinement ratio. Use it to bound the discretization error, not to claim a "true" answer.
2.5 A Diagnostic Decision Path
The figure below maps the observed behavior of a QoI under refinement to the most likely error source and the appropriate corrective action. It is the single most useful diagnostic tool when a mesh study does not behave as expected.
2.6 Practical Guidance for Industrial Work
Three habits separate reliable FEA from decorative FEA:
- Define the QoI before meshing. A convergence study on peak stress at a sharp corner will never converge. Choose a QoI at a location where the stress field is smooth, or use a path-averaged quantity.
- Freeze the idealization during the mesh study. If you change BCs and mesh at the same time, you cannot attribute the change to either. Vary one axis at a time.
- Report the discretization error bound. A result of "2.83 mm ± 1.4%" is engineering. A result of "2.83 mm" is a guess with decimal places.
With the error taxonomy and the diagnostic path established, the next step is to quantify discretization error systematically. That requires an understanding of convergence order and how element formulation affects the rate at which the finite element solution approaches the exact solution of the idealized model.