4. Correlation and validation fundamentals
In virtual crash testing, a simulation is only useful once we can demonstrate that it reproduces the physical event it claims to represent. Correlation is the process of quantifying the agreement between a physical test and its corresponding simulation; validation is the formal judgment that this agreement is sufficient for the intended purpose. The two are related but distinct: correlation produces evidence, validation interprets that evidence against acceptance criteria. This chapter establishes the metrics, the objective rating methods, and the validation hierarchy that make that judgment defensible.
Test-to-simulation comparison metrics
Comparison begins with time-history signals — acceleration at the B-pillar base, occupant head acceleration, seat-belt force, intrusion at the toe pan. Before any metric is computed, the signals must be aligned and conditioned. Three preprocessing decisions dominate the outcome:
- Time-zero alignment. Define a common reference, typically the first contact between barrier and vehicle, and shift both signals to it. A 2 ms misalignment can inflate a peak-difference metric by several percent.
- Filtering. Apply the same SAE J211 channel filter class to both signals (CFC 60 for structural acceleration, CFC 180 or 1000 for occupant head acceleration). Filtering only one side is a common and fatal error.
- Sampling and interpolation. Resample the simulation output onto the test sampling grid, or both onto a common grid, before computing pointwise metrics.
With conditioned signals, the standard metrics are:
| Metric | Definition | What it reveals |
|---|---|---|
| Peak ratio | Simulation peak ÷ test peak | Amplitude fidelity; insensitive to timing |
| Peak timing error | tsim,peak − ttest,peak | Phase/stiffness mismatch |
| Integrated value (e.g., Δv, HIC, 3 ms clip) | Derived injury or kinematic measure from each signal | Whether the physical outcome is reproduced |
| ISO/TS 18571 rating | Weighted combination of phase, magnitude, and shape sub-metrics | Overall signal agreement on a 0–1 scale |
| CORrelation and Analysis (CORA) | Weighted corridor, phase, and shape rating | Overall agreement with tunable weighting |
Peak ratio and timing error are intuitive but individually misleading: a signal can match peaks while being entirely out of phase. This is why composite objective rating methods were developed.
Objective rating methods
ISO/TS 18571 and CORA both decompose the comparison into sub-metrics and combine them with weights. The ISO/TS 18571 rating for a single channel is a weighted sum of an amplitude (magnitude) term, a phase term, and a shape term, each normalized to [0, 1], with 1 indicating perfect agreement. CORA uses a similar decomposition but allows the analyst to set the weighting and the evaluation window, which makes it flexible but also means the weighting must be documented and fixed before evaluation — otherwise the rating becomes a tuning target rather than an assessment.
A practical workflow for a single channel:
- Condition both signals (alignment, filter class, resampling).
- Define the evaluation window — for example, 0 to 120 ms after impact — and exclude pre-trigger noise.
- Compute the sub-metrics and the composite rating.
- Record the rating alongside the weighting scheme and window in the correlation report.
Ratings are not pass/fail by themselves. They become acceptance criteria only when a threshold is agreed in advance for a defined channel set and a defined purpose. A common engineering convention is a composite rating of 0.7–0.8 as "acceptable" for structural channels and higher for occupant injury channels, but these numbers must be justified by the program, not assumed.
The validation hierarchy
Validation is not a single event at the full-vehicle level. It proceeds through a hierarchy in which each level isolates a smaller physical subsystem, so that a discrepancy can be attributed to a specific model assumption rather than buried in a full-vehicle result.
The hierarchy is not merely organizational; it enforces attribution. If a component-level force–displacement curve already diverges from its test, no amount of full-vehicle tuning will produce a physically meaningful correlation — it will only produce a model that matches one test by accident. The rule is to validate each level before trusting the next, and to trace any full-vehicle mismatch downward to the level where it originates.
From correlation to validation judgment
A validation report should state, for each level and each channel: the conditioned signals, the metric values, the composite rating, the pre-agreed threshold, and the pass/fail conclusion. Where a channel fails, the report should identify the likely model assumption — material card, contact stiffness, joint representation — rather than simply recording a low score. This is what separates validation from a numerical exercise: the output is a defensible statement of applicability, not a single number.
With metrics defined, ratings computed under fixed criteria, and the hierarchy respected, the simulation carries evidence of its own fidelity. That evidence is the foundation for the applicability and acceptance boundaries developed in the following chapters.