3. Evidence hierarchy for design decisions

Virtual buck and sled simulations are only useful to a design team when their outputs can be ranked against physical evidence and mapped onto a decision. A single CAE curve, however well-resolved, is not evidence of a design's safety; it is a prediction whose credibility depends on the fidelity of its inputs, the validation status of its material and joint models, and the degree to which the loading and boundary conditions match the real event. This section establishes a practical evidence hierarchy that lets engineers correlate virtual predictions with buck and sled test data, assign defensible confidence levels, and apply explicit decision thresholds so that early design choices are traceable rather than intuitive.

Why a Hierarchy, Not a Single Number

Design reviews frequently collapse into a comparison of one simulated peak acceleration against one measured peak. That comparison is fragile: it ignores phase, pulse shape, duration, and the sensitivity of the metric to modeling assumptions. A hierarchy instead separates evidence by type and provenance, so that a decision can be justified by the strongest available evidence and the residual uncertainty is visible. The hierarchy used here has five tiers, ordered from weakest to strongest for a given design question:

1. Tier 0 — Analytical estimate. Hand calculations, lumped-mass models, or closed-form pulse approximations. Useful for scoping and sanity checks; never sufficient alone for a structural or occupant-safety decision.
2. Tier 1 — Uncorrelated virtual prediction. A full FE buck or sled model run with nominal inputs but no validation against test data for the relevant configuration. Establishes trends and relative ranking between design variants.
3. Tier 2 — Correlated virtual prediction. The same model, but with material cards, joint stiffnesses, and contact parameters calibrated so that predicted and measured responses agree within stated tolerances on a reference configuration.
4. Tier 3 — Buck test evidence. Physical component-level test (subsystem buck) with instrumented measurements, representing the design's actual hardware but not the full vehicle boundary conditions.
5. Tier 4 — Sled test evidence. Full-system sled test with the production-intent restraint and structure, representing the integrated loading environment.

The hierarchy is not a strict ladder in which higher tiers always dominate. A Tier 2 correlated prediction can be more informative than a Tier 3 buck test if the buck omits a load path that the model captures. The point is that each tier carries a known set of limitations, and the decision record must state which tier the conclusion rests on.

Correlation Metrics That Actually Drive Decisions

Correlation should be quantified on the quantities that the design decision depends on, not on the whole time history indiscriminately. For a frontal sled event, the decision-relevant channels are typically the occupant compartment acceleration pulse, the seat and restraint loads, and the structural intrusion at defined measurement points. Three metrics are sufficient for most early-design work:

• Peak magnitude error: (predicted − measured) / measured, evaluated per channel. A common acceptance band for correlated models is ±10% on primary channels and ±15% on secondary channels.
• Pulse-shape error: a normalized integral difference, e.g., the area between predicted and measured cumulative velocity curves divided by the measured velocity change. This catches phase and duration errors that peak comparison misses.
• Time-of-peak offset: the lag between predicted and measured peak, typically required to be within one sampling interval of the dominant frequency or within 5 ms for occupant pulses.

These metrics are computed on a reference configuration — a specific buck or sled build with known hardware. Only after the model meets the acceptance band on the reference configuration is it promoted to Tier 2 and used to predict a new configuration. This separation is essential: calibrating and predicting on the same data produces a model that fits but does not generalize.

Assigning Confidence Levels

Confidence is a structured judgment, not a feeling. A workable scheme assigns a confidence level to each prediction based on three factors: the tier of the underlying evidence, the correlation quality on the reference configuration, and the sensitivity of the predicted metric to uncertain inputs. The table below is a decision aid, not a substitute for engineering judgment.

Confidence should be recorded per metric, not per model. A model may be High confidence for occupant compartment acceleration and Low confidence for seat-back moment if the seat model was never calibrated. The decision record must carry this granularity, otherwise a single "model validated" statement overstates the evidence.

Decision Thresholds and the Correlation Workflow

A decision threshold converts a predicted metric into an accept/reject/ investigate outcome. Thresholds should be set with margin that reflects the confidence level: a High-confidence prediction may be compared directly against the requirement, while a Medium-confidence prediction requires a margin band (e.g., predicted value must be at least 10% inside the requirement) before acceptance. A Low-confidence prediction cannot clear a threshold on its own; it can only flag a design for test.

The workflow below shows how a prediction moves from an uncorrelated run to a decision, and where the correlation gate and confidence assignment sit. The key structural point is that the correlation gate is a branch, not a step: a model that fails correlation is not discarded, it is demoted and used only for trend ranking.

Worked Example: Applying the Hierarchy to a Seat-Anchor Decision

Consider an early-design decision on whether to raise a seat-anchor attachment by 15 mm to improve belt routing. The requirement is that the occupant compartment peak acceleration must not exceed a stated limit, and the seat-anchor load must remain below the attachment's rated capacity. The team has a Tier 2 correlated model validated on the baseline configuration, with primary channels within ±8% and shape error within band. The predicted change for the raised anchor is a 4% increase in seat-anchor load and a 2% decrease in compartment peak acceleration.

Because the model is High confidence on both metrics, the decision threshold can be applied directly: the predicted anchor load remains below capacity with margin, and the acceleration change is favorable. The decision is to accept the change and confirm with a buck test at the next build. If the model had been Medium confidence, the same prediction would require the anchor load to sit at least 10% inside capacity before acceptance, and the buck test would be mandatory rather than confirmatory. If the model had been Low confidence, the prediction could only rank the raised anchor against the baseline; the decision would be deferred to test.

Recording the Evidence in the Decision Log

Each decision should leave a compact record that a reviewer can audit without re-running the model. The record needs, at minimum: the design question, the metric and threshold, the predicted value, the evidence tier, the correlation status on the reference configuration, the assigned confidence, and the resulting decision with its margin. This is the artifact that makes early design decisions defensible when the sled test data arrives months later and either confirms or contradicts the prediction. When the test contradicts a High-confidence prediction, the correlation gate itself is suspect and must be revisited before the model is used again — that feedback loop, shown as the dashed paths in the figure above, is what keeps the hierarchy honest over the life of the program.