2. Characterizing crash simulation discrepancies

Before a DMAIC team can eliminate a discrepancy between a crash simulation and a physical test, it must first describe that discrepancy in terms precise enough to be measured, reproduced, and traced to a cause. "The model doesn't match the test" is not a characterization; it is a complaint. Characterization converts an observed mismatch into a structured statement of what differs, where in space and time it differs, how much it differs relative to a defined tolerance, and which design decision that difference could corrupt. This chapter establishes the vocabulary and the measurement discipline that the rest of the DMAIC investigation depends on.

Classifying discrepancy types

Crash CAE discrepancies fall into a small number of recurring families, and naming the family correctly is the first analytical act because each family points toward a different root-cause space.

• Kinematic discrepancies — the model and test disagree on motion: intrusion velocity, occupant compartment displacement, steering-wheel or pedal-box travel, door opening, or the timing of a rebound. These are usually read from accelerometer-derived displacement or from high-speed video tracking.
• Force and acceleration discrepancies — peak magnitude, pulse shape, rise time, and duration differ on a barrier load cell, seat-belt load limiter, or B-pillar accelerometer. A common signature is a model pulse that is too stiff and too short, indicating an over-stiff structural path or a missing energy-absorption mechanism.
• Energy discrepancies — the partition of initial kinetic energy into structural deformation, friction, and rebound differs. Internal energy curves that plateau at different levels, or hourglass energy exceeding roughly 5% of internal energy, indicate the model is dissipating or storing energy incorrectly.
• Failure-mode discrepancies — the model predicts a different mechanism: a weld separates in the test but holds in the model, or a component buckles in the model where the test shows membrane stretching. These are the most consequential because they change the physical story, not just the numbers.
• Timing discrepancies — events occur in the right order but at the wrong time, such as airbag deployment relative to occupant forward excursion, or a secondary impact arriving early.

In practice a single discrepancy often spans several families. A late occupant excursion may be a kinematic symptom of a force discrepancy (restraint loads too low) rooted in a failure-mode discrepancy (a seat-track weld that the model keeps intact). Recording the family chain explicitly prevents the team from chasing the symptom.

Where discrepancies originate

Sources can be organized by where they enter the simulation chain, which also determines how cheaply they can be tested.

The last row deserves emphasis. A discrepancy is a property of the comparison, not of the model alone. Before attributing a mismatch to CAE, confirm that the test signal has been filtered to the same class and channel frequency as the model output, that the accelerometer is not mounted on a resonant bracket, and that the physical build matches the released CAD and material batch. Many "model errors" dissolve once the test-side processing is aligned.

Quantifying the discrepancy

Characterization requires a metric and a tolerance agreed upon before the investigation begins, so that improvement can be measured objectively. A practical approach uses channel-level metrics with explicit acceptance bands:

Metric definitions (example acceptance bands, project-specific):
  Peak force error        |F_sim - F_test| / F_test        <= 10%
  Pulse duration error    |t_sim - t_test| / t_test        <= 8%
  Intrusion at t=80 ms    |d_sim - d_test|                <= 15 mm
  Energy plateau error    |E_sim - E_test| / E_test        <= 7%
  Hourglass / internal    E_hg / E_int                     <= 5%
  Correlation score       ISO/TS 18571 or CORA rating     >= 0.80

These numbers are illustrative and must be set from the program's own safety margins and regulatory targets. The key discipline is that each metric is computed on a named channel, at a named time window, against a named reference signal. A discrepancy statement that cannot be reduced to such a row is not yet characterized.

From discrepancy to design impact

The reason characterization matters is that CAE drives design decisions: gauge selection, material grade, weld layout, crush-initiation features, and restraint tuning. A discrepancy is only "significant" if it can flip one of those decisions. The figure below maps the measurement chain from test and simulation signals through discrepancy classification to the design decision at risk.

Consider a worked example. A frontal 64 km/h ODB simulation shows a B-pillar acceleration peak of 78 g against a test value of 62 g, with the pulse 6 ms shorter. The intrusion at 80 ms is 12 mm low, and the model's internal energy plateaus 9% above the test-derived estimate. Classified by type, this is a force discrepancy with a kinematic consequence and an energy signature. The source attribution narrows quickly: the pulse is too stiff and too short, energy is over-absorbed, and the intrusion is under-predicted — a pattern consistent with an over-stiff crush zone, for example a missing bead or a material card without strain-rate sensitivity. The design decision at risk is the gauge of the front rail: if the model over-predicts stiffness, the team may down-gauge the rail and under-protect the occupant in the real vehicle. That is the impact statement the DMAIC team must carry forward.

Impact on CAE-driven design decisions

Not every discrepancy carries equal weight. A useful triage is to ask three questions for each characterized discrepancy: Does it exceed the agreed tolerance? Does it change the rank order of design candidates? Does it move a safety margin across a regulatory or internal threshold? A discrepancy that fails all three can be logged and monitored. One that passes the second question is the most dangerous, because it silently corrupts trade studies even when absolute values look acceptable. One that passes the third must be resolved before any design freeze.

Deliverable: discrepancy characterization record

The output of this phase is a single record per discrepancy that the team can hand to the root-cause analysis. A minimal, directly usable form is shown below; the fields are chosen because each one is consumed by a later DMAIC step.

With this record in hand, the investigation has a defined target: reduce the named metric on the named channel below the named tolerance, and confirm that the design ranking is restored. Everything that follows in the DMAIC cycle — cause-and-effect analysis, hypothesis testing, corrective action — operates on this characterized discrepancy rather than on a vague impression of mismatch.