1. Course overview

Occupant injury risk assessment is the quantitative backbone of vehicle safety engineering. Whether the goal is to certify a restraint system against regulatory load cases, to optimize a seat structure for a broad population of occupants, or to support a real-world crash reconstruction, engineers must answer a deceptively simple question: given what we know, how likely is a specific occupant to sustain an injury of a given severity? The difficulty is that every part of that sentence carries uncertainty — in the occupant, in the crash, in the model, and in the data used to build the model. This course is about quantifying that uncertainty rigorously and using artificial intelligence (AI) to make the quantification tractable at engineering scale.

What Injury Risk Assessment Actually Produces

In practice, injury risk is expressed as a probability conditioned on a scenario. A typical formulation is:

P(injury ≥ AIS k | occupant attributes x, crash scenario s, restraint configuration r)

where AIS is the Abbreviated Injury Scale. The output is not a single number for a single dummy in a single sled test; it is a distribution over outcomes for a population of occupants across a space of crash conditions. That shift — from deterministic pass/fail to probabilistic risk — is what makes uncertainty quantification (UQ) unavoidable. A restraint design that produces AIS 2 chest injury in 5% of a population under a 40 km/h frontal offset impact is a fundamentally different engineering object from one that produces it in 0.5%, even if both pass the same certification test.

Three quantities are usually reported together: a point estimate of risk (e.g., mean probability), an interval or credible band around it (e.g., 90% credible interval), and a sensitivity ranking showing which inputs drive the spread. The third is often the most actionable, because it tells the design team where to spend effort — on occupant variability, on material scatter, or on model form error.

Where the Uncertainty Comes From

Uncertainty in occupant injury prediction is not a single phenomenon. It is useful to separate it into categories, because each category is treated differently and each has a different remedy.

• Occupant variability (aleatory). Anthropometry, posture, material properties of biological tissue, muscle state, and age-related fragility vary across the real population. A 5th-percentile female and a 95th-percentile male are not scaled versions of one another; their injury tolerances differ nonlinearly.
• Scenario variability (aleatory). Impact velocity, principal direction of force, overlap, delta-v, and pulse shape vary from crash to crash. Real-world crashes are not the idealized barrier tests.
• Model form uncertainty (epistemic). Finite element (FE) human body models and ATDs are approximations. Material laws, contact definitions, and injury criteria (e.g., chest deflection, HIC, BrIC) embed assumptions that are wrong to some degree.
• Parameter uncertainty (epistemic). Even within a chosen model, parameters such as foam stiffness, belt stiffness, or friction coefficients are known only within ranges.
• Data and measurement uncertainty (epistemic). PMHS (post-mortem human subject) test data, accident databases, and sensor measurements carry noise, censoring, and sampling bias.

The aleatory/epistemic distinction matters operationally. Aleatory uncertainty cannot be reduced by collecting more data — it is irreducible population spread, and the correct response is to design robustly against it. Epistemic uncertainty can be reduced by better models or more data, and the correct response is to quantify it, propagate it, and report it honestly rather than to ignore it.

Why AI Enters the Picture

The classical UQ workflow for occupant injury is Monte Carlo simulation: sample occupant and scenario parameters, run an FE model for each sample, and build the output distribution. This is conceptually clean and embarrassingly parallel, but it is expensive. A single high-fidelity occupant FE run can take hours; a converged Monte Carlo estimate of a tail probability (say, the 95th percentile of chest deflection) may require thousands of runs. When the design space also includes restraint parameters, the cost multiplies.

AI changes the economics in three distinct ways, and it is worth being precise about which one is being used:

1. Surrogate modeling. A Gaussian process or neural network is trained on a modest set of FE runs to approximate the input–output map. Once trained, evaluating the surrogate is milliseconds, so Monte Carlo, importance sampling, or subset simulation can be run at negligible cost. The surrogate's own approximation error becomes an additional epistemic term that must be propagated, not ignored.
2. Variance reduction and adaptive sampling. AI-driven acquisition functions (e.g., Bayesian optimization, active learning) choose the next FE run where it most reduces uncertainty in the risk estimate, rather than sampling uniformly. This can cut the number of expensive runs by an order of magnitude for tail-focused questions.
3. Learned injury criteria and data fusion. Where physical models are weak — for example, mapping kinematic signals to real-world injury outcomes — machine learning can learn the mapping directly from accident and PMHS data, with uncertainty estimated via ensembles or Bayesian neural networks.

A concrete example makes the scale tangible. Suppose a team wants the 90% credible interval of AIS 2+ chest injury probability for a 40 km/h frontal impact across an occupant population. A direct Monte Carlo with 2,000 FE runs gives a point estimate but a wide interval, because the tail is poorly sampled. Replacing the FE model with a Gaussian process surrogate trained on 150 Latin hypercube samples, then running 100,000 surrogate evaluations, produces a tighter interval — but only if the surrogate's predictive variance is included. If the surrogate is treated as ground truth, the reported interval will be overconfident. This is the central discipline of the course: AI is a tool for making UQ affordable, not a substitute for it.

What This Course Builds Toward

Across the 30 chapters, the course moves from foundations to practice. Early chapters establish the probabilistic vocabulary — random variables, distributions, tail probabilities, and the aleatory/epistemic split — and connect them to standard injury criteria. The middle chapters build the computational machinery: design of experiments, surrogate construction, validation metrics (e.g., leave-one-out error, predictive interval coverage), and sensitivity analysis (Sobol indices, Shapley effects). Later chapters address the harder engineering realities: propagating surrogate error, handling censored injury data, validating against PMHS and accident databases, and communicating risk to design and regulatory stakeholders.

How to Read the Rest of the Course

Each chapter is self-contained enough to be studied on its own, but the material is sequenced so that later chapters assume the vocabulary of earlier ones. When a chapter introduces a numerical example, the parameters are stated explicitly and are illustrative — they are chosen to make the method visible, not to represent a validated vehicle program. Readers are encouraged to reproduce the calculations with their own data; the methods, not the specific numbers, are the transferable content.

The single most important habit to develop from this first chapter is to ask, for every risk number encountered: what is the distribution behind it, what sources of uncertainty were included, and which were left out? A risk estimate that silently omits model form error, or that treats a surrogate as exact, is not conservative — it is misleading. The rest of the course is, in effect, a set of tools for answering that question honestly.