4. Validation experiment design

A CAE prediction is only as credible as the experiment it is compared against. In occupant safety work, that experiment is almost always either a sled test (a body-in-white or buck mounted on a deceleration sled, driven by a prescribed pulse) or a full-scale crash test (barrier or car-to-car impact with the complete vehicle). The two serve different validation purposes: the sled isolates the restraint system and occupant response under a controlled, repeatable pulse, while the full-scale test validates the entire chain — structural crush, pulse generation, restraint deployment timing, and occupant kinematics. Designing the validation experiment means deciding what to measure, where, at what sample rate, how to filter, and how to quantify agreement with the LS-DYNA prediction.

Data requirements: what the correlation actually needs

For a sled validation, the minimum dataset is the sled acceleration pulse (measured on the sled carriage, not the buck), the belt force at each retractor and anchor, belt payout/webbing displacement, airbag pressure or deployment timing where applicable, and the full occupant response: head acceleration (triaxial), chest acceleration and chest deflection (IR-TRACC or chest band), pelvis acceleration, and, for the Hybrid III, upper-neck and femur loads. For a full-scale frontal test, add vehicle acceleration at the B-pillar base and rear sill, intrusion measurements at the toe pan, dash, and steering column, and seat track and belt anchor displacements. The rule of thumb: if the LS-DYNA model outputs a channel, the test should measure it, and if the test measures it, the model should output it at a comparable location and in the same coordinate system.

Coordinate conventions must be fixed before anything else. LS-DYNA occupant models typically use the SAE J211/ISO 8727 vehicle axis system: +X forward, +Y to the left, +Z upward, with rotations about those axes positive per the right-hand rule. Every accelerometer and load cell in the test must be documented with its mounting orientation and sign convention, and the test data must be transformed into the model's frame before correlation. A sign flip on chest acceleration is one of the most common and most embarrassing correlation failures.

Instrumentation and sample rate

Occupant accelerometers are typically piezoresistive or piezoelectric units in the ±500 g to ±2000 g range for head and chest, and ±200 g to ±500 g for pelvis and vehicle structure. Belt load cells are tension/compression cells rated to roughly 15–25 kN. Chest deflection uses the IR-TRACC (Infra-Red Telescoping Rod for Assessment of Chest Compression) or a chest band, sampled at the same rate as the accelerometers. All channels should be recorded at a minimum of 10 kHz (20 kHz preferred for head acceleration and belt forces, where high-frequency content matters), with anti-alias filtering applied in hardware before digitization. The data acquisition system must be time-synchronized across all channels — a common event trigger (e.g., the sled fire signal or a breakwire at bumper contact) with a pre-trigger buffer of at least 100 ms.

Filtering per SAE J211

SAE J211/1 defines the channel frequency classes and the corresponding low-pass filter characteristics used to process crash data. The standard classes are CFC 60, CFC 180, CFC 600, and CFC 1000, where the number is the cut-off frequency in Hz at which the filter attenuates by 3 dB (the −3 dB point). The filter is a 4-pole Butterworth, phaseless (zero-phase) filter applied forward and backward so that no phase shift is introduced. The practical consequence: the same physical signal filtered at CFC 180 and CFC 600 will give different peak values and different timing, so test and simulation must be filtered identically before any comparison.

Typical assignments: CFC 60 for vehicle structure and sled pulse (long-duration, low-frequency events); CFC 180 for chest acceleration, pelvis acceleration, and belt forces; CFC 600 for head acceleration and chest deflection; CFC 1000 for high-rate events such as airbag pressure and some neck loads. The filter must be applied to the raw (hardware anti-aliased) data, and the filter's transient response at the start and end of the record must be discarded — typically the first and last 10–20 ms. In LS-DYNA post-processing, the same CFC filter is applied to the nodal or element output using the same class, and the time base is aligned to the test trigger.

Correlation metrics

Visual overlay is necessary but not sufficient. The standard quantitative metrics used in occupant safety correlation are:

• Peak value and time-of-peak error — the most intuitive and the most sensitive to timing misalignment. Report both absolute and percentage error.
• Peak-to-peak amplitude ratio — simulation peak divided by test peak, per channel.
• ISO/SAE correlation factor (often called the ISO 18571 or objective rating method) — a composite score combining magnitude, phase, and shape agreement, typically reported on a 0–1 scale where >0.8 is good and >0.9 is excellent. It is computed per channel and then aggregated.
• Normalized integral square error (NISE) or root-mean-square error normalized by the test peak — useful for shape agreement over the full pulse.
• Cross-correlation time lag — reveals systematic timing offsets (e.g., a restraint firing 5 ms late in the model).

A practical correlation report tabulates, for each channel: CFC class, test peak, sim peak, peak error %, time-of-peak error (ms), and the ISO 18571 score. Channels are then grouped by physical subsystem (structure, belt, airbag, occupant) so that a low score can be traced to a specific modeling assumption rather than averaged away.

Practical correlation workflow and failure symptoms

A reproducible workflow looks like this. First, export the LS-DYNA channels at the same sample rate as the test (or resample both to a common rate, e.g., 10 kHz) and apply the identical CFC filter class to each. Second, align the time bases using a common event — the sled fire signal or the bumper breakwire — and verify alignment by checking that the sled pulse onset matches within 1–2 ms. Third, compute the per-channel metrics and populate the correlation table. Fourth, inspect the channels with the lowest ISO 18571 scores and trace them to the responsible subsystem.

Typical failure symptoms and their likely causes: a systematic time lag across all occupant channels usually points to a restraint firing time offset (retractor pretensioner or airbag trigger) or a pulse onset misalignment; a peak amplitude that is too high in simulation but correct in timing often indicates an over-stiff belt or seat cushion, or a missing energy-absorbing feature; a chest deflection that is too low in simulation frequently traces to a belt stiffness or geometry issue rather than the chest model itself; and a head acceleration that is too high only at CFC 600 suggests a contact stiffness or mesh-density problem in the head/airbag interface. Each of these is diagnosable only if the filter class and coordinate convention are consistent between test and model — which is why those two decisions belong at the very start of the validation experiment design, not at the end of the correlation report.