2. Uncertainty taxonomy for crash safety AI

Crash safety prediction sits at the intersection of physics-based simulation and data-driven learning. A neural surrogate trained on finite element (FE) results, a Gaussian process mapping intrusion to occupant injury risk, or a Bayesian calibration of a material card all face the same fundamental question: when the model says something, how much should we trust it? Answering that question requires more than a single "confidence" number. It requires a taxonomy that separates uncertainty by its origin, because different origins demand different treatments — more data helps some, better physics helps others, and nothing helps a third category except acknowledging it.

This chapter establishes that taxonomy. We distinguish aleatoric from epistemic uncertainty, then decompose the practical sources that dominate crash safety AI: model form error, numerical noise, and material variability. Each is defined, illustrated with crash-specific examples, and connected to the modeling decisions it should drive.

Aleatoric vs. Epistemic: The First Cut

Aleatoric uncertainty (from alea, dice) is irreducible randomness inherent to the system or its measurement. Even with a perfect model and infinite data, the outcome varies. In crash safety, aleatoric uncertainty appears when two nominally identical vehicles hit the same barrier at the same speed and produce different peak accelerations — because spot-weld nugget diameters vary by tenths of a millimeter, because the occupant's posture differs by a few degrees, because the barrier face has microscopic roughness differences. You cannot eliminate this by collecting more data; you can only characterize its distribution.

Epistemic uncertainty (from episteme, knowledge) is reducible ignorance. It stems from limited data, imperfect model structure, or unknown parameters. A surrogate model trained on 200 simulations has epistemic uncertainty about regions of the design space it has never seen. A constitutive model that omits strain-rate sensitivity has epistemic uncertainty about high-speed events. Collect more relevant data, improve the model form, or narrow the parameter bounds — and this uncertainty shrinks.

The practical distinction matters because the two require opposite responses. Aleatoric uncertainty calls for robust design: choose a design whose performance distribution has acceptable tails. Epistemic uncertainty calls for information gathering: run more simulations, add experiments, or refine the model. Confusing them leads to wasted effort — running a million Monte Carlo samples to "reduce" aleatoric uncertainty that was never going to shrink, or accepting a design as safe when the real problem was a sparse training set.

Model Form Error

Model form error (also called model discrepancy or structural uncertainty) is the gap between the best possible prediction from your model class and physical reality — even with perfect parameters. A classic crash safety example: a lumped-mass occupant model cannot represent belt slip and airbag interaction simultaneously, no matter how carefully you tune its stiffness and damping. The error is structural, not parametric.

In AI for crash safety, model form error has two layers. The first is inherited from the simulation that generated training data: if your FE solver uses a simplified spot-weld representation, every label carries that bias. The second is introduced by the ML model itself: a neural network with a fixed architecture may be unable to represent the true input–output relationship, especially near discontinuities like the onset of buckling or fracture.

Quantifying model form error typically requires a discrepancy function. In a Bayesian calibration framework, you model the physical observation as:

y_obs(x) = y_sim(x, θ) + δ(x) + ε

where:
  y_sim(x, θ)  = simulator output at input x with calibration parameters θ
  δ(x)         = model discrepancy function (often a Gaussian process)
  ε            = observation noise

The discrepancy term δ(x) absorbs systematic model error. In crash safety, δ(x) is often structured — for example, a bias that grows with impact velocity because the model under-predicts strain-rate effects. Ignoring δ(x) and forcing the simulator to fit data by tuning θ alone produces calibrated parameters that are physically wrong but numerically convenient, a well-documented pitfall.

Numerical Noise

Numerical noise is uncertainty introduced by the computational solution process itself. It is distinct from model form error: the model form may be adequate, but the numerical solution is imperfect. Sources in crash simulation include:

• Time-step discretization: Explicit solvers use a critical time step; results shift slightly with step size.
• Mesh sensitivity: Refining a mesh changes predicted intrusion by a few percent even when the solution is "converged."
• Solver tolerances: Contact penalty stiffness, hourglass control, and convergence criteria introduce non-deterministic-looking variation.
• Floating-point accumulation: In long explicit runs, round-off accumulates, especially in parallel decompositions.

For AI models, numerical noise in training labels is particularly insidious. If the same nominal simulation run twice with different random seeds (e.g., for stochastic material sampling or parallel reduction order) yields outputs differing by 2%, a model trained to predict those outputs will learn noise as signal unless the noise is characterized and either averaged out or modeled explicitly. A practical diagnostic: run the same input through the simulation pipeline N times with only solver seed changed, compute the standard deviation of the output, and compare it to the model's reported uncertainty. If the model's uncertainty is smaller than the numerical noise floor, the model is overconfident.

Material Variability

Material variability is the aleatoric uncertainty in material properties and their spatial distribution. In crash safety, it is often the dominant uncertainty source. Consider a stamped high-strength steel B-pillar: yield strength varies coil-to-coil by ±5–8%, thickness by ±3–5%, and the forming process introduces thickness thinning and residual stresses that vary part-to-part. Spot welds have nugget diameter distributions. Composites have fiber orientation and void content variability.

These variations propagate through the crash event nonlinearly. A 5% yield strength reduction does not produce a 5% intrusion increase; it may produce a 15% increase if it triggers a different buckling mode. This nonlinear amplification is why material variability cannot be treated as a simple error bar added at the end — it must be propagated through the model.

In AI-based crash safety, material variability enters in two ways. First, as input features: if the model takes yield strength and thickness as inputs, variability is represented explicitly. Second, as unmodeled randomness: if the model takes only geometry and impact speed, material variability becomes part of the aleatoric noise the model must absorb. The choice has consequences. Explicit representation enables robust design optimization; implicit representation simplifies the input space but inflates the irreducible uncertainty.

Why the Taxonomy Matters in Practice

Consider a concrete engineering scenario. A team trains a Gaussian process surrogate to predict peak chest deflection from impact speed and restraint parameters. The model reports a predictive standard deviation of 2.1 mm. Is that good? The answer depends entirely on decomposition.

Suppose the team runs a repeatability study: the same FE model, same inputs, five different solver seeds. The output standard deviation is 1.8 mm. That means 1.8 mm of the reported 2.1 mm is numerical noise — the model has learned almost nothing beyond the noise floor. The remaining 0.3 mm is the model's actual epistemic uncertainty, which is small. The correct action is not to collect more training data (epistemic is already low) but to reduce numerical noise by tightening solver tolerances or averaging multiple runs per label.

Now suppose instead the repeatability study gives 0.2 mm, but the model's predictive standard deviation is 2.1 mm. Here the noise floor is low, and the 2.1 mm is genuine epistemic uncertainty — the model is unsure because the training set is sparse. The correct action is to add simulations in the high-uncertainty region, not to tighten the solver.

This decomposition — noise floor vs. model uncertainty — is the single most practical diagnostic in uncertainty-aware crash safety AI. It prevents the two most common errors: over-engineering the solver when the problem is data, and collecting data when the problem is numerical hygiene.

Connecting Sources to Model Design

The taxonomy also guides architecture choices. If material variability dominates, the model should take material parameters as explicit inputs and be trained on a distribution of material properties — not on a single nominal card. If model form error dominates, a pure data-driven model may be preferable to a physics-informed one, or the discrepancy term must be modeled explicitly. If numerical noise dominates, the training labels should be denoised (e.g., by averaging multiple solver runs) before the model sees them.

In the chapters that follow, we will build on this taxonomy to implement specific uncertainty quantification methods: Bayesian neural networks for epistemic uncertainty, heteroscedastic regression for aleatoric uncertainty, and Gaussian process discrepancy models for model form error. Each method targets a specific branch of the tree above. Choosing the right method begins with correctly diagnosing which branch dominates your problem.