4. Data-driven surrogate modeling for impact

Full-vehicle and component crash simulations remain among the most computationally expensive workloads in automotive CAE. A single LS-DYNA frontal offset run on a detailed body-in-white model can consume 8–24 wall-clock hours on 64–128 cores, and design-of-experiments (DOE) studies that vary thickness, material grade, and geometry across dozens of factors quickly become intractable. Surrogate modeling — also called metamodeling — replaces the expensive solver with a cheap mathematical approximation trained on a limited set of simulation runs. This chapter covers the three most widely used surrogate families for crash metrics: polynomial response surfaces, kriging, and Gaussian process regression, and explains when each is appropriate.

Why Crash Metrics Need Special Treatment

Crash responses are not smooth, well-behaved functions of design variables. Consider three representative metrics:

• Peak longitudinal deceleration at the B-pillar — highly sensitive to local buckling sequence, often discontinuous.
• Intrusion at the pedal box or toe pan — governed by progressive folding, with step-like changes when a hinge triggers.
• HIC (Head Injury Criterion) — depends on airbag timing and contact, introducing near-discontinuous behavior.

These responses exhibit nonlinearity, interaction effects, and in some regions genuine non-smoothness. A surrogate that assumes global quadratic behavior will fail near buckling transitions. This is the central motivation for moving beyond classical response surfaces toward kriging and GPR.

Polynomial Response Surface Methodology (RSM)

RSM fits a low-order polynomial to sampled data. A second-order model in k variables takes the form:

y(x) = β₀ + Σ βᵢxᵢ + Σ βᵢᵢxᵢ² + ΣΣ βᵢⱼxᵢxⱼ + ε

The coefficients are estimated by ordinary least squares. For a 5-variable crash study, a full quadratic model requires 21 coefficients, so a face-centered central composite design with 27–43 runs is typical. RSM is fast to fit, easy to interpret (coefficients indicate factor importance), and adequate when the response is genuinely smooth — for example, mass or global stiffness metrics.

However, RSM cannot represent the sharp transitions that dominate intrusion and acceleration peaks. Its prediction error near a buckling threshold can exceed 20%, which is unacceptable when the design margin is only 10–15%.

Kriging and Gaussian Process Regression

Kriging, developed in geostatistics and formalized as Gaussian process regression in machine learning, treats the unknown response as a realization of a stochastic process:

y(x) = μ(x) + Z(x)

where μ(x) is a deterministic trend (often a constant or low-order polynomial) and Z(x) is a stationary Gaussian process with zero mean and covariance:

Cov[Z(xᵢ), Z(xⱼ)] = σ² · R(θ, xᵢ, xⱼ)

The correlation function R is typically a Gaussian or Matérn kernel. For crash applications, the Matérn ν=5/2 kernel is often preferred because it produces smoother but less over-confident interpolants than the squared-exponential kernel:

R(θ, d) = (1 + √5·d/θ + 5d²/(3θ²)) · exp(−√5·d/θ)

where d is the scaled Euclidean distance between two design points and θ is the length-scale hyperparameter. Hyperparameters are estimated by maximum likelihood or cross-validation.

The key advantage for crash metrics is that kriging is an exact interpolator: predictions pass through every training point, and the model provides a variance estimate at each prediction location. This variance is invaluable for adaptive sampling — you can add simulation runs precisely where the surrogate is most uncertain, which is exactly where buckling transitions tend to occur.

Practical Workflow and Example

Consider a simplified example: a front rail section with three design variables — gauge t (1.2–2.0 mm), yield strength σy (340–780 MPa), and trigger depth d (5–15 mm). The crash metric is peak force F from a 40% offset impact.

1. Sampling: Use a Latin hypercube design with 30 initial points across the three-variable box.
2. Simulation: Run each point in LS-DYNA, extract peak force from the cross-section force history.
3. Fit: Train a kriging model with Matérn 5/2 kernel. Optimize θ by maximizing the concentrated log-likelihood.
4. Validate: Perform leave-one-out cross-validation. Compute R² and normalized RMSE. For crash metrics, R² > 0.9 and NRMSE < 5% are reasonable targets.
5. Refine: Identify regions with high prediction variance, add 5–10 infill points, and refit.

In practice, the initial 30-point kriging model often achieves R² ≈ 0.85–0.92 for peak force, while a quadratic RSM on the same data reaches only R² ≈ 0.70–0.80. The gap widens for intrusion metrics, where kriging's flexibility captures the step-like behavior that polynomials cannot.

Comparison and Selection Guidance

For crash metrics with discontinuous or highly nonlinear behavior, kriging/GPR is the default choice. RSM remains useful as a screening tool or when the response is known to be smooth. A common hybrid strategy is to use RSM for initial factor screening, then switch to kriging for the final optimization on the reduced variable set.

Failure Modes and Troubleshooting

Two failure modes are common. First, ill-conditioned correlation matrices occur when two design points are nearly identical; the matrix becomes singular and hyperparameter optimization diverges. The fix is to enforce a minimum distance in the DOE or add a small nugget (regularization) to the diagonal. Second, overfitting with noisy data — if the simulation itself has run-to-run variation (e.g., from contact or hourglass control), an interpolating kriging model will fit noise. Adding a nugget term or using a regression kriging variant addresses this.

Visual Summary

The workflow above shows the iterative nature of surrogate construction: initial DOE, simulation, fitting, validation, and adaptive refinement. The comparison panel summarizes the trade-offs that drive method selection. In the next chapter, we will examine how these surrogates integrate into multi-objective optimization and reliability assessment for crashworthiness design.