4. Surrogate modeling for crashworthiness

In Pareto-based multi-objective optimization of occupant protection, every candidate design point must be evaluated against crashworthiness responses such as HIC15, chest deflection, chest acceleration (3 ms clip), femur axial force, and restraint-system metrics. A single high-fidelity finite element (FE) crash simulation of a full frontal sled or vehicle model may take 4–24 hours on a multi-core workstation, and a multi-objective evolutionary algorithm (MOEA) such as NSGA-II may request thousands of evaluations. Directly coupling the optimizer to the FE solver is therefore computationally infeasible. Surrogate modeling (metamodeling) replaces the expensive FE response with a cheap analytical approximation, enabling the optimizer to explore the design space at negligible cost. This chapter covers the four pillars of a defensible surrogate workflow: design of experiments (DOE), response surface methodology (RSM), Kriging, and validation against occupant safety responses.

Design of Experiments: Choosing Where to Sample

The surrogate is only as good as the sample points that train it. DOE determines the spatial distribution of the n training runs across the k-dimensional design space. For occupant protection studies, typical design variables include seat belt anchor height, pretensioner force limit, airbag vent diameter, seat pan angle, and knee bolster stiffness — often 5–10 variables with physically bounded ranges.

Three DOE families dominate crashworthiness practice:

  • Full factorial and fractional factorial designs are efficient for screening. A 2k full factorial with k = 6 requires 64 runs; a resolution IV fractional factorial (26-2) needs only 16 runs and can still separate main effects from two-factor interactions. Use these in the screening phase to eliminate variables with negligible influence before building the final surrogate.
  • Latin Hypercube Sampling (LHS) provides space-filling coverage with exactly n runs regardless of dimension, making it the default for Kriging training. An optimized LHS (maximin criterion) maximizes the minimum inter-point distance, which improves Kriging conditioning.
  • Central Composite Design (CCD) and Box–Behnken designs are tailored for second-order polynomial RSM. A CCD with k = 5 has 25 factorial points, 2k = 10 axial points, and typically 6 center points — 48 runs total — and supports estimation of all quadratic terms.
For crashworthiness, always include 3–6 replicate center points. The pure-error estimate from replicates lets you separate lack-of-fit from noise, which is essential when validating the surrogate later.

Response Surface Methodology

RSM fits a low-order polynomial to the sampled data. The second-order model is standard:

y(x) = β₀ + Σᵢ βᵢxᵢ + Σᵢ βᵢᵢxᵢ² + Σᵢ<ⱼ βᵢⱼxᵢxⱼ + ε

Coefficients are estimated by ordinary least squares. RSM is transparent, fast to fit, and its coefficients directly reveal sensitivity — a large |βᵢ| signals a dominant main effect, while a significant βᵢⱼ indicates interaction. For occupant responses that are smooth and mildly nonlinear over the design range (e.g., chest deflection versus belt force limit), RSM often achieves R² > 0.95 with 30–50 runs.

RSM's weakness is its rigidity: it cannot capture sharp nonlinearity such as airbag deployment timing thresholds or bottoming-out of a force limiter. When residuals show systematic curvature, escalate to Kriging rather than adding higher-order terms, which destabilize the fit.

Kriging: Interpolation for Nonlinear Crash Responses

Kriging (Gaussian process regression) models the response as a trend plus a correlated stochastic process:

y(x) = f(x)ᵀβ + Z(x)

where f(x)Tβ is a global trend (often a constant or linear polynomial) and Z(x) is a stationary Gaussian process with zero mean and covariance σ²R(θ, x, x′). The correlation kernel R is typically the Gaussian or Matérn function with per-dimension length scales θ. Kriging interpolates the training data exactly and provides a prediction variance σ̂²(x) at every unsampled point — a built-in uncertainty estimate that RSM lacks.

This variance is the key to adaptive sampling: after fitting an initial Kriging model on, say, 40 LHS points, compute the expected improvement (EI) or the maximum prediction variance across a dense candidate grid, run the FE simulation at the most informative point, refit, and repeat. In practice, 15–25 adaptive infill points typically reduce the maximum prediction error by an order of magnitude compared with a static 60-point LHS. For occupant safety, infill points frequently cluster near constraint boundaries — for example, where HIC15 approaches 700 or chest deflection approaches 63 mm — because these regions govern feasibility and are where the optimizer needs accuracy most.

The figure below summarizes the complete surrogate workflow, from variable screening through adaptive infill to validation and handoff to the MOEA.

1. Variable Screening 2^(k-p) fractional factorial, 16–32 runs → keep 5–8 factors 2. Training DOE Maximin LHS, n = 40 + 4 center replicates FE runs in batch 3. Fit Surrogate RSM (quadratic) or Kriging (Matérn 5/2) per response 4. Adaptive Infill Max EI / max σ̂² 15–25 new FE runs near constraint bounds refit until convergence 5. Validation Hold-out set (10 pts) R², RMSE, max error LOO cross-validation 6. MOEA Handoff NSGA-II on surrogate 10⁴–10⁵ evals Pareto front extraction 7. FE Confirmation Re-run 3–5 Pareto points in full FE verify HIC, deflection Occupant safety responses tracked per surrogate: HIC₁₅ (limit 700) Chest deflection (63 mm) Chest accel 3 ms (60 g) Femur force (10 kN) Neck Nᵢⱼ (1.0) Acceptance gate: R² ≥ 0.90 and max hold-out error ≤ 5% of the regulatory limit for every response before the surrogate is released to the optimizer. If any response fails, add infill points in the failing region and refit — do not proceed to MOEA with an unvalidated surrogate.

Validation Against Occupant Safety Responses

A surrogate that fits its training data well can still fail catastrophically at unsampled points. Validation must therefore use data the model has never seen. The standard protocol has three layers:

  1. Hold-out validation. Reserve 10–15% of the FE runs (e.g., 8 of 60) before fitting. Compute R², RMSE, and — most importantly for safety — the maximum absolute error on the hold-out set. A model with R² = 0.97 but a single 12% error at a near-limit point is unsafe for constraint handling.
  2. Leave-one-out cross-validation (LOOCV). For Kriging, LOOCV is cheap because the prediction at each point can be computed analytically from the correlation matrix. Report the LOOCV RMSE and plot predicted-versus-actual scatter; systematic curvature in the residual plot indicates the trend function is inadequate.
  3. Constraint-boundary check. Because the MOEA will push designs against the HIC and deflection limits, evaluate the surrogate on a dense grid within ±10% of each constraint boundary and compare against targeted FE runs. This is the check that most often exposes a surrogate that is accurate on average but wrong exactly where feasibility is decided.
Never validate a surrogate only on its training set. An interpolating Kriging model will report near-zero training error by construction, which says nothing about predictive accuracy. Always report hold-out or LOOCV metrics alongside the training fit.

Practical Example: Belt Force Limiter and Anchor Height

Consider a two-variable study: belt force limit F ∈ [2.5, 4.5] kN and anchor height h ∈ [0, 60] mm, with responses HIC15 and chest deflection. A 16-run maximin LHS plus 4 center replicates (20 FE runs) is executed. A quadratic RSM fit yields R² = 0.94 for HIC15 but only 0.81 for chest deflection, with residuals showing a clear interaction between F and h. A Kriging model with a Matérn 5/2 kernel and per-dimension length scales raises the LOOCV R² for chest deflection to 0.96. Adaptive infill adds 12 points concentrated where chest deflection approaches 63 mm, and the hold-out maximum error drops from 8.4% to 2.9% of the limit. Only then is the surrogate handed to NSGA-II. The final Pareto front is confirmed by re-running three selected designs in the full FE model, where predicted versus simulated HIC15 agrees within 3.1%.

The lesson is consistent across occupant safety applications: RSM is adequate for smooth, weakly interacting responses and for screening; Kriging with adaptive infill is the workhorse for nonlinear responses and constraint-boundary accuracy; and validation on held-out and boundary-focused data is non-negotiable before any surrogate drives a multi-objective optimizer.

← Previous 📖 Contents Next →