4. CAE model verification planning

An airbag CAE model is not "verified" in the abstract. It is verified for a purpose. A model built to rank restraint concepts during an early architecture study has a completely different verification burden than a model used to sign a production release for FMVSS 208 or Euro NCAP. Verification planning is the discipline of writing down, before any correlation run begins, what the model is for, how accurate it must be, what numeric thresholds define "accurate enough," and which physical tests will serve as the reference truth. Without that plan, correlation degenerates into an open-ended fitting exercise, and the release decision becomes a matter of opinion rather than evidence.

Step 1 — Define the model purpose and the decision it supports

Start by naming the engineering decision the model must support. Typical purposes in an airbag program include: (a) concept screening of vent size and tether length, (b) restraint-system tuning against a sled pulse, (c) OOP (out-of-position) risk assessment, (d) occupant injury prediction for a specific load case, and (e) final compliance evidence for a regulatory or consumer test. Each purpose implies a different required fidelity, a different set of outputs, and a different acceptance threshold.

A useful practice is to write the purpose as a single sentence with an explicit output variable and a decision rule, for example: "Predict peak chest deflection (mm) for the 50th-percentile male Hybrid III in the 56 km/h frontal ODB load case, to decide whether the restraint tuning margin exceeds 15% of the injury criterion." This forces the planner to commit to the output quantity, the occupant, the load case, and the decision margin before any simulation is run.

Step 2 — Derive required fidelity from the purpose

Fidelity is not a single knob. It decomposes into at least four independent dimensions, and each must be justified by the purpose:

• Geometric fidelity — Is the folded airbag represented by a mapped fold, a simplified uniform-pressure volume, or a corpuscular particle method (CPM) gas model? Concept screening may accept a uniform-pressure (UP) volume; OOP and deployment-kinematics studies generally require a mapped or CPM representation.
• Material and constitutive fidelity — Fabric permeability, coating friction, seam stiffness, and strain-rate dependence. A release-grade model needs coupon-level validation of the fabric model, not just a nominal modulus.
• Load-case fidelity — Sled pulse, barrier geometry, seat position, belt routing, and pre-crash braking. The pulse must match the physical test within a stated tolerance (commonly ±5% on velocity change and ±10% on peak acceleration over the primary pulse window).
• Numerical fidelity — Mesh size in the bag and occupant, time-step control, contact thickness, and hourglass energy. A common rule is to keep hourglass energy below 5% of internal energy and total energy error below 5%.

The plan should state, for each dimension, the chosen level and the reason. "We used a UP volume because the purpose is concept screening and the decision margin is 15%" is a defensible statement. "We used a UP volume because it was faster" is not.

Step 3 — Set acceptance thresholds before correlation

Thresholds must be numeric, tied to the output variable, and agreed with the release authority before the correlation matrix is executed. Three families of metrics are standard in occupant safety CAE:

Thresholds should be tighter for release-grade models and looser for concept models, but they must be written down. A practical convention is to define a primary channel (the one the decision depends on) with the tightest threshold, and secondary channels with relaxed thresholds. If the primary channel passes but a secondary channel fails, the plan should state whether that blocks release or triggers a documented waiver.

Step 4 — Establish traceability to physical tests

Every CAE output that supports a release decision must trace to a physical test with a known configuration. Traceability has three layers:

1. Test-to-model mapping — A table linking each physical test ID (e.g., SLED-2024-017) to the corresponding CAE deck, load case, and occupant model. Include the test date, facility, and any deviations from the nominal configuration.
2. Input traceability — The sled pulse, belt geometry, seat position, and bag pressure trace used in the model must be the measured values from that specific test, not a nominal curve. Store the measured pulse as a CSV and reference it by filename and checksum.
3. Output traceability — Each CAE channel (head accel, chest deflection, belt force) must map to a specific test channel and sensor location. Document the sensor ID, sampling rate, and filter class (e.g., CFC 600 for accelerometers per SAE J211).

Traceability is what allows a reviewer to reproduce the correlation independently. Without it, the correlation is anecdotal.

Step 5 — Assemble the verification plan

The plan is a controlled document, typically 5–15 pages, containing: purpose statement, output variables, fidelity decisions with justification, threshold table, test-to-model mapping table, correlation matrix (which tests are used for calibration vs. validation), and a sign-off block. A minimal but complete example is shown below.

VERIFICATION PLAN — Airbag Model AB-2024-FRONT
Purpose: Predict peak chest deflection (mm), 50th Hybrid III, 56 km/h ODB,
         decision margin 15% of FMVSS 208 criterion.
Primary channel: CHEST_DEFL (CORA ≥ 0.80, peak error ≤ 10%, Δt_peak ≤ 5 ms)
Secondary channels: HEAD_ACCEL, BELT_FORCE (CORA ≥ 0.70)
Fidelity: mapped fold, CPM gas, coupon-validated fabric, 8 mm bag mesh
Calibration tests: SLED-2024-011, SLED-2024-014
Validation tests:  SLED-2024-017, SLED-2024-021
Pulse tolerance: Δv ≤ 5%, peak accel ≤ 10%
Energy check: hourglass ≤ 5% IE, total energy error ≤ 5%
Sign-off: CAE Lead, Test Lead, Release Authority

Common failure modes in verification planning

Three mistakes recur. First, thresholds set after correlation — the model is tuned until it looks good, then a threshold is chosen to match. This invalidates the verification. Second, calibration and validation tests mixed — using the same test to tune and to validate gives an optimistic CORA that does not generalize. Keep them separate; a common split is 60% calibration, 40% validation. Third, missing pulse traceability — using a nominal sled pulse instead of the measured one introduces a systematic error that no amount of material tuning can remove.

The diagram above shows the planning sequence and the decision gate that follows correlation. The critical property is that steps 1–4 are completed and signed off before the correlation matrix is executed. That ordering is what separates verification from curve-fitting, and it is what makes the eventual release defensible to a regulator, a customer, or an internal review board.