4. CAE model setup for correlation

Correlation between CAE predictions and physical test results is decided long before the solver runs. Three modelling decisions dominate the outcome: how the geometry is simplified, how the mesh is discretized and verified, and how boundary conditions (BCs) and loads are represented relative to the test setup. A model that is numerically converged but physically misrepresented will correlate poorly no matter how much effort goes into post-processing. This chapter treats these three decisions as a single, coupled preparation task whose goal is a model that is both computationally tractable and faithful to the test article.

Geometry simplification: keep what the test measures, remove what it does not

CAD geometry from design is built for manufacturing, not for analysis. Fillets, chamfers, weld beads, bolt threads, stamped ribs, and cosmetic features inflate element count and often create poor-quality elements without influencing the quantities being correlated. Simplification is therefore a deliberate filtering process, and the filter criterion is the correlation target, not aesthetics.

A practical workflow is to classify every feature into one of three buckets. Retain features that carry load paths, stiffness, mass, or stress concentration in the region of interest: primary ribs, load-introduction bosses, and any geometry within roughly one characteristic thickness of a strain-gauge location. Idealize features whose behavior can be represented more cheaply: replace a bolted joint with a rigid or CBUSH-type connection, convert a thin stamped panel to shell elements, and substitute a weld bead with a tied contact or a shell offset. Remove features that are demonstrably irrelevant to the target response: small fillets away from the gauge, engraved part numbers, and non-structural trim.

Two cautions matter in correlation work. First, removing mass changes the modal and dynamic response; if the test includes modal hammering or shaker excitation, track the mass delta introduced by simplification and report it. A common rule of thumb is to keep the total mass change below about 1–2 % of the test article mass, and to correct the density or add non-structural mass (NSM) if it exceeds that. Second, simplification must be applied consistently between the CAE model and the as-tested configuration. If the test article had a bracket removed for instrumentation clearance, the CAE model must reflect that, not the nominal design.

Mesh convergence: converge the quantity you intend to correlate

Mesh convergence is frequently treated as a global check, but correlation demands convergence of the specific output being compared. A model can show stable global stiffness while the local strain at a gauge location is still mesh-sensitive. The convergence study should therefore track the correlation quantities directly: gauge strains, accelerations at sensor locations, reaction forces, and the first several natural frequencies.

A defensible procedure uses at least three mesh densities with a consistent refinement ratio (typically 1.5–2.0 in element size) and reports the relative change between successive levels. The convergence criterion is met when the change in the target quantity falls below a stated tolerance, for example 3 % for strain and 1 % for the first bending frequency. Where possible, use h-refinement (smaller elements) in the region of interest and keep the far field coarser; p-refinement (higher-order elements) is an efficient alternative for smooth stress fields.

Element quality interacts with convergence. In thin-walled structures, shell elements with aspect ratios above roughly 5:1 and warping above 5° can produce spurious local results that masquerade as mesh sensitivity. Before concluding that a model is unconverged, check Jacobian, aspect ratio, and warping metrics in the gauge region. Similarly, avoid placing a single element across a thickness where bending dominates; use at least two or three elements through the thickness for solid bending, or switch to shells.

Convergence log (example, illustrative values only)
Level   Element size   Max gauge strain   f1 (Hz)   Delta strain   Delta f1
M1      8 mm           1180 µε            42.1      —              —
M2      5 mm           1245 µε            41.6      +5.5 %         −1.2 %
M3      3 mm           1268 µε            41.4      +1.8 %         −0.5 %
M4      2 mm           1274 µε            41.3      +0.5 %         −0.2 %
Criterion: |Delta| < 3 % strain, < 1 % frequency → M3 accepted

The example above illustrates the decision logic rather than a real result: M3 is accepted because the incremental change from M2 to M3 is within tolerance, and M4 confirms the trend. Note that the frequency converges faster than the local strain, which is typical and is exactly why the local quantity must drive the acceptance decision.

Boundary condition representation: mirror the test, not the idealization

Boundary conditions are the most common source of correlation error because they are invisible in the results but dominate the response. The guiding principle is to represent the test fixture as it actually constrains the article, including its compliance, and to distinguish constrained degrees of freedom from loaded ones.

For a typical vibration or static test, the fixture is not rigid. A shaker table or a bolted test frame introduces stiffness and damping that a fully fixed BC overestimates. Where fixture compliance is significant, represent it with springs or a reduced substructure rather than an encastre condition. For bolted interfaces, a fixed BC applied over the entire contact patch is usually too stiff; a more faithful representation distributes the constraint over the bolt footprint or uses a contact definition with the correct friction coefficient.

Load introduction deserves equal care. A point load applied at a single node creates a stress singularity and unrealistic local strain; distribute the load over the physical pad, grip, or actuator footprint using a rigid spider, a distributed coupling, or a pressure over the contact area. For displacement-controlled tests, apply the measured actuator displacement at the correct location and orientation, and verify that the reaction force predicted by the model matches the test load cell within a stated tolerance before comparing strains.

Coordinate conventions must be fixed and stated once. Define the global axes, the sign convention for strains (tension positive), and the sensor orientation relative to those axes. If a gauge is rosette-based, specify how the rosette axes map to the model coordinate system; a 90° orientation error will produce a sign-flipped correlation that looks like a modelling error but is a bookkeeping error.

Bringing the three decisions together

The three preparation decisions are coupled. Aggressive geometry simplification can hide a load path that the BC representation then misloads; a coarse mesh can mask a BC-induced singularity; an over-stiff fixture can make a converged mesh look unconverged because the local strain keeps changing with refinement near the constraint. The figure below summarizes how the decisions feed the correlation quantities and where each one can corrupt the comparison.

The practical takeaway is a gate sequence. First, simplify geometry and record the mass delta. Second, converge the mesh on the correlation quantities, not just global stiffness. Third, validate the BC representation against a measured reaction force or load-displacement curve. Only when all three gates pass should material parameters be calibrated, because calibration against a misrepresented model will absorb modelling errors into the material constants and destroy predictive capability on the next test.

A useful discipline is to keep a short model-preparation record alongside the model: the simplification log, the convergence table, the BC justification, and the coordinate convention. When correlation later shows a discrepancy, this record lets the engineer distinguish a genuine material or physics issue from a preparation artefact, which is the difference between a calibration that generalizes and one that merely fits a single dataset.