2. Active learning fundamentals for crash simulation

Crash simulation campaigns are expensive: a single nonlinear explicit dynamics run of a full vehicle model may consume hundreds of CPU-hours and produce gigabytes of d3plot and binout output. Running a dense factorial or Latin hypercube design over ten or more design variables is rarely affordable. Active learning (AL) reframes the campaign as a sequential decision problem: instead of pre-committing to a fixed DOE, the surrogate model itself proposes which simulation to run next, so that each expensive LS-DYNA job buys the maximum possible reduction in predictive uncertainty. This chapter establishes the AL machinery — query strategies, acquisition functions, uncertainty quantification, batch selection, and stopping criteria — and adapts each element to the specific pathologies of explicit crash dynamics.

Why Passive DOE Breaks Down in Crash Work

Crash responses are non-stationary and discontinuous. Intrusion, peak acceleration, and HIC are governed by discrete events: a spot weld failing, a hinge folding, a rocker panel bottoming out. Between two nearby points in the design space, the response may jump by 30% because a single connection changed its failure sequence. A space-filling design spends most of its budget resolving smooth regions and under-samples the narrow bands where the physics actually switches. Active learning concentrates samples where the surrogate is wrong or uncertain, which in crash work usually means near these event boundaries.

The AL loop has five recurring components. The surrogate (typically a Gaussian process, or a deep ensemble / MC-dropout network for high-dimensional inputs) provides both a prediction and a variance. The query strategy decides whether to exploit (sample where the prediction is extreme) or explore (sample where the variance is high). The acquisition function scores every candidate point. Batch selection picks k points to run in parallel on the cluster. The stopping criterion halts the campaign when further runs no longer improve the decision the campaign is meant to support.

Uncertainty Quantification: Two Sources That Must Be Separated

In crash AL, uncertainty is not one quantity. Aleatoric uncertainty comes from run-to-run scatter: material property tolerances, weld nugget variability, solver time-step noise. It does not shrink as you add samples. Epistemic uncertainty comes from lack of data and does shrink. Acquisition functions that ignore this split will keep sampling noisy regions forever. A practical decomposition for a GP surrogate is:

sigma_total^2(x) = sigma_epistemic^2(x) + sigma_noise^2
# GP posterior:  mu(x), sigma_epistemic(x) from kernel
# sigma_noise estimated from repeat runs at 3-5 anchor points
# Only sigma_epistemic drives exploration; sigma_noise sets the floor

For deep ensembles, epistemic uncertainty is the variance across ensemble members, while aleatoric uncertainty is the mean of per-member predicted variances. Estimate the noise floor empirically: run the same model three times with identical inputs but perturbed initial velocities within the tolerance band, and record the spread in peak intrusion. If that spread is 2 mm, no acquisition function should chase differences smaller than 2 mm.

Acquisition Functions for Crash Responses

The choice of acquisition function encodes what the campaign is optimizing. Four are standard, and each maps to a different crash objective.

EI is the workhorse when a single scalar objective dominates. For crash, however, the objective is often a constraint set: intrusion must stay below a limit and the design must remain manufacturable. Here a constrained EI is more appropriate — multiply EI by the probability that all constraints are satisfied. If the constraint is a hard regulatory threshold, PoF alone may be the right driver early in the campaign, switching to EI once a feasible region is identified.

Batch Selection Under Cluster Constraints

Crash jobs run in parallel, so the AL loop must propose a batch of k designs per iteration, not one. Naively taking the top-k acquisition scores produces a cluster of near-identical points — a failure mode called batch collapse. Two remedies work well in practice:

• Local penalization: after selecting point x₁, down-weight the acquisition surface near it by a penalty proportional to the surrogate length-scale. Re-score and pick x₂, and so on.
• Diversity-constrained selection: enforce a minimum Euclidean distance in normalized design space, e.g. 0.05 per variable, between any two batch members.

For a 12-variable crash model (thicknesses, material grades, weld pitch), a batch of k = 8 with local penalization typically covers the design space far better than k = 8 independent EI maxima. The batch size should match the number of concurrent solver licenses; running k = 8 when only 4 licenses exist just queues jobs and delays the loop.

Stopping Criteria Tailored to Explicit Dynamics

Stopping is where most AL campaigns silently fail. Three criteria are defensible for crash work:

1. Convergence of the decision, not the surrogate. Stop when the best feasible design has not changed by more than a tolerance (e.g. 1% in peak intrusion) over three consecutive iterations.
2. Exhausted improvement budget. Stop when the maximum EI over the candidate pool falls below a fraction of the observed response noise — sampling below the noise floor is wasted money.
3. Resource cap. A hard limit on total CPU-hours or number of runs, set before the campaign starts.

Worked Example: Side-Impact Intrusion Campaign

Consider a simplified side-impact model with five design variables: B-pillar thickness (1.2–2.0 mm), rocker thickness (1.4–2.2 mm), door beam gauge (1.0–1.8 mm), material grade (DP590–DP780), and weld pitch (20–40 mm). The objective is to minimize maximum door intrusion at 50 km/h MDB impact while keeping mass below a limit. The campaign is seeded with 12 Latin hypercube runs, then runs 10 AL iterations with batch k = 4, using constrained EI and a GP surrogate with a Matérn 5/2 kernel.

After seeding, the surrogate's epistemic uncertainty is highest near the weld-pitch extremes, where the failure sequence changes. The first AL batch selects four points clustered at weld pitch 22–26 mm and B-pillar thickness 1.5–1.7 mm — the region where intrusion transitions from beam bending to weld rupture. By iteration 6, the best feasible design stabilizes at 1.62 mm B-pillar, 1.85 mm rocker, 1.45 mm beam, DP700, 28 mm pitch, with predicted intrusion 142 mm and mass 1.8 kg under the limit. Iteration 7's maximum EI drops below the 2 mm noise floor, triggering criterion 2. Total runs: 52, versus an estimated 180 for a comparable-resolution full factorial.

Practical Guardrails

Three failure modes recur in crash AL campaigns. First, noise-floor neglect: the acquisition function chases differences smaller than the solver's own scatter, producing a campaign that never converges. Always estimate the noise floor before iteration 1. Second, batch collapse: without penalization or a diversity constraint, the batch wastes licenses on near-duplicate designs. Third, decision drift: the stopping criterion is tied to surrogate accuracy rather than the engineering decision, so the campaign halts with a surrogate that looks good globally but mispredicts the event boundary. Tie every stopping rule to the design decision the campaign exists to support.

With these five components — surrogate, query strategy, acquisition function, batch selection, and decision-tied stopping — the AL loop becomes a controllable engineering process rather than an open-ended search. The next chapter applies this framework to the specific surrogate architectures that handle crash response discontinuities.