2. Anatomy of an uncertainty budget

An uncertainty budget is not a single number attached to a simulation result. It is a structured accounting of every identifiable source of doubt, the mechanism by which that doubt travels through the assessment chain, and the resulting confidence that can be attached to a specific claim. In virtual crash assessment, the claim is typically a statement such as "the occupant's chest deflection does not exceed the injury threshold" or "the vehicle structure absorbs the required energy at the specified impact velocity." The uncertainty budget is the evidence that such a claim is defensible.

The anatomy of a budget has three layers that must be kept distinct: sources (where uncertainty originates), propagation paths (how it moves through the model and analysis), and claim mapping (how the accumulated uncertainty translates into a confidence statement about the assessment outcome). Conflating these layers is the most common failure in virtual assessment documentation, because it produces budgets that list many contributors but cannot explain which ones actually matter for the claim.

Layer 1: Sources of uncertainty

Sources in crash simulation fall into recognizable families. Input uncertainty covers material properties (steel grade yield strength scatter, strain-rate sensitivity), geometry (sheet metal thickness tolerances, weld nugget placement), and boundary conditions (barrier friction, impactor alignment). Model-form uncertainty covers constitutive model choice, shell versus solid element representation, hourglass control, and contact algorithm selection. Numerical uncertainty covers mesh discretization, time-step size, and solver tolerances. Experimental uncertainty covers the reference test data used for validation: sensor calibration, filtering, and repeatability. Scenario uncertainty covers the fact that the real-world crash configuration is not exactly the regulatory or design test configuration.

Each source should be characterized by a distribution or interval, not just a label. A material yield strength might be specified as a normal distribution with mean 355 MPa and standard deviation 12 MPa, derived from mill certificates. A thickness tolerance might be a uniform interval of ±0.06 mm. Where data are absent, an expert-elicited interval with a stated confidence level is acceptable, but it must be recorded as such.

Layer 2: Propagation paths

Propagation is where most budgets lose rigor. A source does not affect the claim directly; it affects intermediate quantities, which in turn affect the response metric. For example, yield strength scatter propagates to the force–displacement curve of a crash box, which propagates to the deceleration pulse, which propagates to occupant chest deflection. The budget must trace this chain, because a source with large variance that is strongly attenuated along the path may matter less than a small source close to the claim.

Propagation can be quantified in three practical ways. Analytical sensitivity uses a first-order Taylor expansion: the sensitivity coefficient is the partial derivative of the response with respect to the input, and the variance contribution is the square of that coefficient times the input variance. Sampling methods (Monte Carlo, Latin hypercube) run the model across sampled inputs and estimate the output distribution directly. Surrogate-based propagation fits a response surface or Gaussian process to a small design of experiments and then samples the surrogate cheaply. For crash models, where a single run may take hours, surrogate-based propagation is usually the only feasible route to a full distribution.

A useful discipline: for every source, record the sensitivity coefficient and the variance contribution separately. A source with a large coefficient but tiny variance and a source with a small coefficient but large variance can contribute equally, and only the product reveals which dominates.

Layer 3: Mapping to claim confidence

The final layer converts the propagated output distribution into a statement about the claim. If the claim is a threshold comparison, the relevant quantity is the probability that the response exceeds the threshold, or equivalently the margin between the response distribution and the threshold expressed in standard deviations. A claim that chest deflection remains below 50 mm, when the predicted distribution has mean 42 mm and standard deviation 3 mm, sits at a margin of about 2.7 standard deviations, corresponding to a small but non-zero exceedance probability. That probability, together with the validation evidence supporting the model, is the confidence statement.

This mapping must also account for validation gap. A model validated against a 50 km/h frontal test does not automatically carry the same confidence at 64 km/h or in an oblique configuration. The budget should include a model-form term for extrapolation, and the claim confidence should be reduced accordingly. The mapping is therefore not a pure statistical calculation; it is a structured judgment that combines propagated numerical uncertainty with an explicit assessment of how well the model represents the claimed scenario.

Uncertainty budget anatomy: sources, propagation paths, and claim confidence mapping Three-layer diagram showing uncertainty sources on the left, propagation through intermediate response quantities in the center, and the mapping to a threshold-based claim confidence on the right. Uncertainty budget anatomy Sources → propagation paths → claim confidence LAYER 1 · SOURCES LAYER 2 · PROPAGATION LAYER 3 · CLAIM MAPPING Input: material scatter σ_y = 355 ± 12 MPa (normal) Input: geometry tolerance thickness ±0.06 mm (uniform) Model form shell vs solid, hourglass control Numerical mesh size, time step, tolerance Experimental reference sensor calibration, repeatability Scenario real crash ≠ test configuration Crash-box force–displacement sensitivity S₁ = ∂F/∂σ_y Vehicle deceleration pulse sensitivity S₂ = ∂a/∂F Occupant chest deflection sensitivity S₃ = ∂d/∂a Response distribution μ = 42 mm, σ = 3 mm (surrogate) Threshold comparison Claim: d < 50 mm Margin = (50 − 42)/3 ≈ 2.7 σ P(exceed) small, non-zero Validation gap Tested: 50 km/h frontal Claimed: 64 km/h oblique Extrapolation term added Confidence reduced Claim confidence statement Defensible with stated margin

Worked example: from source to claim

Consider a frontal crash box whose force–displacement response feeds an occupant chest-deflection claim. The material yield strength is characterized as normal with mean 355 MPa and standard deviation 12 MPa. A one-at-a-time sensitivity study on the crash box gives a sensitivity coefficient of 0.8 kN per MPa of yield strength at the plateau force. The variance contribution from this source alone is therefore (0.8 × 12)² = 92.2 kN², or a standard deviation of about 9.6 kN in plateau force. If the plateau force itself has a mean of 180 kN, this is roughly 5.3% coefficient of variation.

That force uncertainty propagates to the deceleration pulse. Suppose the vehicle mass is 1500 kg and the pulse is approximated as a rectangular plateau; the deceleration is F/m, so a 9.6 kN force uncertainty becomes 6.4 m/s² of deceleration uncertainty. If the occupant model has a chest-deflection sensitivity of about 0.4 mm per m/s² of deceleration, the contribution to chest deflection is 2.6 mm. Combined with other sources (geometry, model form, numerical), the total chest-deflection standard deviation might reach 3 mm, which is the value used in the claim mapping above.

This example illustrates the key discipline: the budget is built from sensitivity coefficients and variances, not from qualitative labels. It also shows why a source that looks small at the material level (12 MPa on 355 MPa, about 3.4%) can become a meaningful fraction of the claim margin after two propagation steps.

Common failure modes in budget construction

  • Double counting. Treating mesh refinement and solver tolerance as independent when both are controlled by the same time-step setting inflates the total variance. Sources must be checked for correlation before variances are summed.
  • Ignoring model-form uncertainty. Budgets that include only input scatter and numerical error systematically understate confidence, because the largest contributor in crash simulation is often the constitutive and contact modeling choice.
  • Claim drift. A budget built for a 50 km/h frontal claim is silently reused for a 64 km/h oblique claim. The propagation path changes, and the validation gap term must be re-evaluated.
  • Unstated distributions. Recording a source as "±10%" without stating whether that is a uniform bound, a standard deviation, or a 95% interval makes the variance contribution non-reproducible.
A budget that lists sources but does not quantify sensitivity coefficients cannot support a confidence claim. The mapping from simulation inputs to claim confidence requires numbers at every link in the chain, and any link left qualitative becomes an unquantified risk that must be disclosed in the assessment report.

What a complete budget contains

A defensible budget for a virtual crash assessment claim should contain, at minimum: a source register with distribution type and parameters; a propagation map identifying intermediate quantities and sensitivity coefficients; the combined response distribution at the claim metric; the threshold and computed margin; and an explicit validation-gap assessment with its effect on confidence. The budget is complete when a reviewer can trace any single source through the propagation map to its contribution at the claim, and can see how the claim confidence would change if that source were removed or halved. That traceability, rather than the number of sources listed, is the measure of budget quality.

← Previous 📖 Contents Next →