1. Crash test vs CAE correlation

In vehicle safety development, a physical crash test and its corresponding CAE (Computer-Aided Engineering) simulation are two representations of the same event. The crash test records what a real prototype actually did under a controlled impact; the CAE model predicts what the design should do according to the finite element representation, material laws, and boundary conditions. Correlation is the disciplined process of quantifying how well these two representations agree, explaining the differences, and deciding whether the model is trustworthy enough to support engineering decisions. Without correlation, a simulation is only a plausible animation; with correlation, it becomes a validated predictive instrument.

Why Correlation Exists: Fundamental Objectives

The first objective is model credibility. A full-vehicle frontal impact model may contain 3–8 million elements, dozens of material cards, spot and seam welds, joints, airbags, seat belts with pretensioners and load limiters, and a barrier model. Each of these introduces uncertainty. Correlation against a physical test provides evidence that the aggregated model reproduces the dominant physical mechanisms: longitudinal rail folding, engine cradle kinematics, dash panel intrusion, occupant compartment deceleration, and restraint system loading.

The second objective is design-space confidence. Once a model is correlated within an accepted tolerance for a given load case, engineers may use it to explore variants — gauge changes, material upgrades, geometry revisions, additional reinforcements — without building a prototype for every iteration. The correlated model effectively becomes a virtual test rig whose error bounds are known.

The third objective is diagnosis and root-cause support. When a physical test shows an unexpected result (for example, a rail that folds earlier than predicted, or a door that opens during impact), correlation work reveals whether the discrepancy originates from the model (missing weld, incorrect material strain-rate curve, simplified joint) or from the test (sensor drift, fixture compliance, build variation). This bidirectional insight is often more valuable than a single "good" or "bad" correlation score.

Industrial Context: Where Correlation Sits in the Development Cycle

In a typical OEM program, the safety attribute follows a V-model. On the left side, CAE supports concept and detailed design: crashworthiness targets, load-path definition, restraint tuning. On the right side, physical prototypes are built and tested: sled tests, subsystem impacts, and full-vehicle barrier tests. Correlation activities bridge the two sides, and they occur repeatedly rather than once.

A practical sequence looks like this. First, a pre-test prediction is run and archived before the physical test, so that the prediction is not unconsciously adjusted afterward. Second, the test is executed with a documented sensor layout, barrier condition, and vehicle build record. Third, test data is processed: filtering per SAE J211, channel zeroing, and time-zero alignment at the impact trigger. Fourth, the CAE results are post-processed with matching filtering and extracted at equivalent locations. Fifth, a quantitative comparison is performed, and discrepancies are investigated. Sixth, model updates are made and the updated model is re-validated against the same or additional tests.

The Comparison Problem and the Role of CORA

Comparing two curves by eye is subjective and does not scale. A full-vehicle test may produce 60–150 channels, each with different magnitudes, phases, and frequency content. Engineers therefore need an objective, repeatable metric. CORA (Correlation and Analysis), originally developed within the German automotive industry and standardized through ISO/TS 18571, provides exactly this: a weighted, normalized score that decomposes the disagreement between a reference curve (test) and a comparison curve (CAE) into interpretable components.

CORA evaluates two families of criteria. The corridor criteria assess whether the CAE curve stays within an inner and outer corridor derived from the test signal, capturing magnitude and phase deviation. The cross-correlation criteria assess shape agreement through the progression of a sliding correlation coefficient, capturing how well the overall waveform and its phase align. These are combined into a single overall score, typically reported between 0 and 1, where 1 indicates perfect agreement. The weighting between corridor and cross-correlation contributions is configurable, and the choice of weighting is itself an engineering decision that should be documented.

The score is not a verdict by itself. A high overall score can hide a locally poor region — for example, good agreement during the initial 30 ms but divergence during rebound. CORA is therefore used together with channel-level review, and the overall score is usually reported alongside per-channel scores and the specific criteria values. Common industrial acceptance bands are program-specific, but a frequently cited convention treats an overall score above roughly 0.8 as good agreement, 0.6–0.8 as acceptable with documented justification, and below 0.6 as requiring investigation or model improvement.

What a Correlation Engineer Actually Does

Consider a frontal 64 km/h ODB (offset deformable barrier) case. The test produces a left B-pillar accelerometer channel, filtered to CFC 180. The CAE model has a node at the equivalent B-pillar location. Before any scoring, the engineer confirms that the CAE node is at the same physical position, that the acceleration sign convention matches (positive toward vehicle rear, for example), and that both signals are zeroed at the impact trigger. Only then is CORA run.

Suppose the overall score is 0.72, with a strong corridor score but a weaker cross-correlation score. That pattern typically indicates a phase shift: the CAE pulse arrives slightly earlier or later than the test pulse. The engineer inspects the time history and finds the CAE deceleration peak occurs about 4 ms earlier. Possible causes include a stiffer front rail trigger, a different barrier model stiffness, or a mass distribution difference. The engineer then checks the rail force channel and the barrier force channel to localize the source. If the rail force correlates well but the barrier force does not, the barrier model is the prime suspect. This is the diagnostic value of decomposing the score rather than reading only the overall number.

After a model update — for example, adjusting the barrier material model or correcting a rail trigger geometry — the model is re-run and re-scored. Importantly, the updated model should also be checked against a second, independent test (a different impact speed or a different barrier) to avoid over-fitting to a single event. A model tuned to one test and failing all others is not validated; it is merely fitted.

Practical Guardrails

• Archive the pre-test prediction. Post-test adjustment without a recorded baseline destroys the evidential value of correlation.
• Match the comparison basis exactly. Channel location, filter class, time-zero, and sign convention must be identical before scoring.
• Report per-channel and per-criterion results. An overall score alone can mask localized disagreement.
• Validate against multiple load cases. Single-event agreement is necessary but not sufficient.
• Document weighting and acceptance bands. CORA weighting choices and program thresholds are engineering decisions that must be traceable.

In short, correlation is the bridge that converts a simulation from an illustration into an engineering instrument, and CORA provides the objective, decomposable metric that makes that bridge auditable. The remaining chapters build on this foundation: signal processing details, channel selection, scoring interpretation, and the iterative model-update loop that turns a first-pass model into a validated predictive asset.