2. Fundamentals of probability theory for crashworthiness

Crash CAE is often described as deterministic: given a mesh, a material card, and an initial velocity, the explicit solver returns a single deformation field. In engineering practice, however, that single result is one realization drawn from a population of physically possible outcomes. Sheet metal thickness varies within coil tolerances, spot welds miss their nominal positions, strain-rate parameters scatter between material batches, and the occupant's seating posture differs from the certification dummy. Probability theory provides the language for describing this scatter, propagating it through the model, and making defensible statements about the probability that a restraint or structure meets its performance target. This section establishes the vocabulary and the computational mechanics that the rest of the course builds on.

Random variables and their representation in crash models

A random variable (RV) is a measurable function that maps each outcome in a sample space to a real number. In crashworthiness we almost always work with continuous RVs: gauge thickness t, yield stress σ_y, longitudinal stiffness k, peak deceleration a_max, intrusion d, or the Head Injury Criterion HIC. Each is written with an uppercase symbol when treated as random (T, Σ_y, A_max) and lowercase when a specific realization is meant (t = 1.42 mm). This convention matters when writing limit-state functions: g(X) = d_allow − D(X) is a random variable, while g(x) evaluated at a design point is a number.

Two representation choices dominate. The cumulative distribution function (CDF), F_X(x) = P(X ≤ x), is the most general description and is what reliability methods ultimately evaluate. The probability density function (PDF), f_X(x) = dF_X/dx, is more intuitive for continuous variables and is what most sampling schemes draw from. For a thickness tolerance of 1.40 ± 0.06 mm, a normal or truncated normal PDF with μ = 1.40 and σ = 0.02 (i.e., ±3σ covering the tolerance band) is a common and defensible assumption; a lognormal is preferable when the variable must remain strictly positive and is right-skewed, as with fatigue-relevant weld nugget diameter.

Distributions encountered in crashworthiness

The choice of distribution is an engineering decision, not a formality, because it controls the tails that dominate reliability estimates. The table below summarizes the families that appear most often in body-in-white and restraint work, together with the parameterization and the physical justification.

When only a tolerance band is available, a useful rule is to set the standard deviation to one-third of the half-width for a normal fit, or to one-sixth of the full range for a uniform fit. These are heuristics, not laws; they should be documented and, where possible, replaced by measured batch statistics.

Expectation, variance, and covariance

The expectation E[X] = ∫ x f_X(x) dx is the probability-weighted mean and is the quantity a nominal CAE run approximates. The variance Var(X) = E[(X − E[X])²] measures dispersion; its square root is the standard deviation, which carries the same units as the variable and is therefore the natural scale for tolerance statements. For a linear combination Y = Σ a_i X_i, the variance is

Var(Y) = Σ a_i² Var(X_i) + 2 Σ_{i<j} a_i a_j Cov(X_i, X_j)

The covariance Cov(X_i, X_j) = E[(X_i − μ_i)(X_j − μ_j)] captures linear dependence, and the correlation coefficient ρ_ij = Cov/(σ_iσ_j) normalizes it to [−1, 1]. In crash models, correlation is not a nuisance: thicknesses of panels cut from the same coil are positively correlated, and yield stress and ultimate stress of the same steel grade are strongly correlated (ρ ≈ 0.8–0.95). Ignoring positive correlation in a sum of contributions typically underestimates the variance of the response, which is unconservative for reliability.

Worked example: variance of peak deceleration from two correlated inputs

Consider a simplified frontal sled model in which peak deceleration is approximated by a linear response surface, A_max = 42.0 + 18.0(t − 1.40) − 0.35(σ_y − 350), with t in mm and σ_y in MPa. The coefficients are illustrative and would normally be fitted from a designed set of runs. Suppose t ~ N(1.40, 0.02²), σ_y ~ N(350, 15²), and ρ_t,σ = 0.3. Then

Var(A_max) = (18.0 × 0.02)² + (0.35 × 15)² + 2(18.0)(−0.35)(0.3)(0.02)(15)
           = 0.1296 + 27.5625 − 1.1340
           ≈ 26.56  →  SD ≈ 5.15 g

The negative covariance term arises because the two coefficients have opposite signs; the correlation partially cancels the contributions. Had the correlation been ignored, the variance would have been 27.69 and the standard deviation 5.26 g — a 2% difference here, but the sign and magnitude of such errors grow quickly when several correlated inputs act together. The expected peak deceleration is 42.0 g, and a one-sided 95th-percentile estimate is 42.0 + 1.645 × 5.15 ≈ 50.5 g, which can be compared directly against a restraint design target.

From moments to decisions

Expectation and variance are summaries, not the whole story. Two designs can share the same mean and standard deviation yet differ markedly in the upper tail that governs injury risk. This is why the course later introduces full-distribution methods (Monte Carlo, First- and Second-Order Reliability Methods) rather than relying on a mean-plus-sigma rule. Nevertheless, the moment framework is the correct starting point: it exposes which inputs dominate the variance budget, it makes correlation visible, and it gives a first-order estimate of the probability of exceeding a limit. In the example above, the yield-stress term contributes roughly 99% of the variance, so any effort to reduce scatter in peak deceleration should begin with material characterization rather than with tighter gauge control — a conclusion that is difficult to reach by inspecting nominal runs alone.

With these definitions in place — random variables, their distributions, and the first two moments including covariance — the next step is to connect them to the deterministic solver through sampling or response surfaces, which is the subject of the following sections.