2. CAE-test correlation fundamentals

In quantitative CAE–test correlation, the term correlation does not mean that a simulation "looks similar" to a test. It means that a defined set of simulated quantities and a corresponding set of measured quantities agree within an explicitly stated tolerance, over a stated domain, and for a stated purpose. The word quantitative is doing real work here: correlation is a numerical claim, and a numerical claim must be falsifiable. If the acceptance criterion cannot fail, it is not a correlation criterion.

A useful working definition is: CAE–test correlation is the process of establishing, quantifying, and documenting the relationship between simulation outputs and physical measurements for a defined load case, boundary condition set, and response quantity, such that the simulation can be used with known confidence for a specified engineering decision. Three elements in that definition are frequently neglected in practice — the load case, the response quantity, and the decision. Correlation without a decision context degenerates into curve-matching.

What Correlation Is, and What It Is Not

Correlation is often confused with three neighbouring activities. Validation asks whether the model represents the physics adequately for the intended use; correlation is the quantitative evidence that supports that judgement. Verification asks whether the model was solved correctly (mesh convergence, solver tolerances, unit consistency); a verified model can still be poorly correlated. Calibration is the adjustment of uncertain parameters to improve agreement; it is a means to correlation, not correlation itself, and an over-calibrated model can correlate beautifully on the calibration data while predicting poorly elsewhere.

The distinction matters because it determines what you are allowed to conclude. A model calibrated on a single modal test and reported as "correlated" tells you almost nothing about its behaviour under a different boundary condition. The honest statement is narrower: the model reproduces the first six elastic modes of the as-built assembly on soft mounts within 5% in frequency and 0.15 in MAC. That sentence is defensible; "the model is validated" is not.

Objectives of a Correlation Study

Correlation studies in industrial practice pursue a small number of concrete objectives, and it is worth naming them explicitly at the start of any project because they drive the test plan as much as the simulation plan:

• Quantify model error for specific response quantities, so that downstream analyses carry a defensible uncertainty band rather than an unstated one.
• Identify and reduce dominant uncertainty sources — typically boundary conditions, material properties, damping, and assembly stiffness — in order of their contribution to output variance.
• Establish parameter values that are physically meaningful and transferable to sibling configurations, rather than case-specific fudge factors.
• Support a decision: release a design, accept a margin, set a test severity, or authorize extrapolation to a load case that cannot be tested.
• Preserve institutional knowledge in a form that a later engineer can audit — which is why the correlation report, not the correlation plot, is the real deliverable.

Note that "make the FRF match" is not on this list. Matching is instrumental; the objectives are epistemic and decisional.

Where Correlation Sits in the Development Cycle

Correlation is not a single gate at the end of a programme. It appears at several points, with different scope and different rigour, and the most common programme failure is treating a late, single, high-stakes correlation exercise as a substitute for earlier, cheaper ones.

At T1, coupon and joint tests calibrate the material card and connection stiffness before detailed design freezes. This is the cheapest place to discover that a bolted joint's rotational stiffness is an order of magnitude lower than the rigid assumption. At T2, a subsystem modal survey on soft mounts (or in free-free suspension) establishes frequency and mode-shape agreement; this is where boundary-condition errors surface most clearly, because a free-free test removes the mounting uncertainty that dominates later results. At T3, the full system is correlated under a representative service load case, usually through FRFs or operational response. At T4, field data closes the loop and feeds model updates into the next programme.

The V-model placement is deliberate: each correlation touchpoint sits opposite the design stage it informs, and the acceptance metric tightens as the cost of being wrong increases. A 5% frequency tolerance at T2 is not arbitrary — it is roughly the level at which mode ordering becomes reliable and at which a subsequent forced-response prediction is not dominated by modal frequency error.

A Worked Example: Subsystem Modal Correlation

Consider a welded steel bracket assembly, approximately 0.8 m × 0.5 m, mass 24 kg, mounted on four M10 bolts to a rigid base. The test is performed with the bracket suspended on soft bungee cords (rigid-body modes below 3 Hz) and excited with an instrumented hammer at 12 grid points; response is measured with three uniaxial accelerometers roving across the grid. The FE model uses shell elements (S4R, 6 mm target size), a linear-elastic steel card (E = 210 GPa, ν = 0.3, ρ = 7850 kg/m³), and bolted connections represented as CBUSH elements with initial stiffness 1.0 × 10⁶ N/mm in all translational directions.

The first comparison yields the following (values are illustrative, constructed for teaching purposes, not measured data):

The pattern is diagnostic. Modes 1 and 3, both dominated by global bending, are over-stiff by 7–17%, while the torsion mode is nearly correct and the local plate mode is frequency-accurate but shape-poor. A uniform material error would shift all modes proportionally; it does not. The likely culprit is the bolted joint representation: bending modes load the joints in tension/compression and rotation, whereas torsion loads them primarily in shear. Reducing the CBUSH rotational stiffness from rigid to 2.0 × 10⁵ N·mm/rad and the axial stiffness to 4.0 × 10⁵ N/mm brings modes 1 and 3 to within 3% while leaving mode 2 essentially unchanged — consistent with the physical load path.

Mode 4's low MAC (0.79) with correct frequency is a different problem: the shape is wrong, not the stiffness. Common causes are a local mode interacting with a global one, an accelerometer placed near a nodal line, or a mesh that is too coarse to capture the local plate bending. Checking the test grid density and refining the local mesh to 3 mm resolves this class of discrepancy more often than parameter tuning does.

Correlation as a Decision Instrument

The output of a correlation study is not a plot; it is a statement of the form: for load case L, response quantity R, over frequency band B, the model agrees with test within tolerance T, using parameters P, and is therefore fit for decision D. Every element of that statement must be recorded, because the model's fitness is conditional on all of them. A model correlated for a 0–200 Hz bending response on soft mounts is not thereby correlated for a 500 Hz acoustic transmission problem on hard mounts.

This conditionality is why correlation reports should state the domain of validity explicitly, alongside the parameter values that were adjusted and the evidence that they are physically plausible. A calibrated joint stiffness of 4.0 × 10⁵ N/mm is credible if it lies within the range measured on similar joints; a value of 40 N/mm that happens to make the FRF match is a red flag, not a result. The discipline of asking "is this parameter physically defensible?" is what separates correlation from curve-fitting, and it is the single most reliable protection against a model that predicts the past beautifully and the future badly.

Finally, correlation has a schedule dimension. Because each touchpoint informs a decision, a correlation result delivered after the decision has been made has no value beyond documentation. The practical implication is that test and analysis plans must be co-scheduled from the start: the test must measure the quantities the model predicts, at the locations the model resolves, under boundary conditions the model represents. When correlation fails late in a programme, the root cause is frequently not a modelling error but a planning one — the test was designed to qualify a product, not to correlate a model.