2. Surrogate modeling fundamentals for crash CAE

Full-vehicle and component crash simulations remain among the most computationally expensive workloads in automotive CAE. A single frontal offset impact model with a detailed body-in-white, closures, restraints, and a validated material card set can require 8–24 wall-clock hours on 64–256 cores. Design space exploration, robustness studies, and multidisciplinary optimization routinely demand hundreds to thousands of such runs, which is economically infeasible with direct LS-DYNA or PAM-CRASH execution. Surrogate modeling — also called metamodeling — replaces the expensive solver with a cheap analytical approximation trained on a limited design of experiments (DoE). This chapter establishes the mathematical foundations of the four dominant surrogate families used in crash CAE and clarifies where machine learning extends, rather than replaces, classical approximation theory.

Problem Formulation and the DoE Backbone

Let the crash model be a black-box function y = f(x), where x ∈ R^d collects design variables (gauge thicknesses, material yield strengths, rib positions, spot-weld pitch) and y ∈ R^m collects responses (peak intrusion, HIC₁₅, chest deflection, mass, section forces). A surrogate ŷ = g(x) is constructed from n sampled pairs {(xi, yi)}. The sampling plan is not incidental: for crash responses, which are highly nonlinear and often discontinuous in x, a space-filling design such as Latin Hypercube Sampling (LHS) or a maximin-optimal design outperforms factorial or central composite designs. A practical rule of thumb is n ≈ 10·d for kriging and n ≈ 3·d for polynomial response surfaces, though crash problems with buckling and fracture frequently require 2–3× these counts to capture regime changes.

Polynomial Response Surfaces

Response Surface Methodology (RSM) fits a low-order polynomial, typically quadratic:

ŷ(x) = β₀ + Σ βᵢxᵢ + Σ βᵢᵢxᵢ² + ΣΣ βᵢⱼxᵢxⱼ

Coefficients are estimated by ordinary least squares. RSM is transparent, cheap to evaluate, and its coefficients expose sensitivity directly — a large βᵢᵢ signals strong curvature, while a large βᵢⱼ reveals interaction between variables. Its weakness is rigidity: a quadratic cannot represent the sharp force plateau of a crash pulse or the bifurcation between bending and axial collapse. RSM is therefore best reserved for smooth, mildly nonlinear responses such as mass or low-speed stiffness, or as a screening tool before committing to a more flexible model.

Kriging (Gaussian Process Regression)

Kriging treats the response as a realization of a Gaussian random field:

ŷ(x) = μ + Z(x),  Cov[Z(xᵢ), Z(xⱼ)] = σ² · R(θ, xᵢ, xⱼ)

The correlation kernel R is usually a squared-exponential or Matérn function with per-dimension length scales θ. Kriging is an exact interpolator: predictions pass through every training point, and it returns a variance estimate that quantifies local uncertainty. This makes it the workhorse for Efficient Global Optimization (EGO), where the expected improvement criterion balances exploitation of low predicted values against exploration of high-variance regions. In crash CAE, kriging with an anisotropic Matérn kernel (ν = 5/2) is a common default because it tolerates the moderate non-smoothness introduced by contact and element erosion. The cost is cubic scaling in n for hyperparameter fitting, which becomes limiting beyond roughly 1,000 samples.

Polynomial Chaos Expansion

Polynomial Chaos Expansion (PCE) represents the response as a spectral sum of orthogonal polynomials Ψα in the input random variables:

ŷ(x) = Σ<sub>α∈A</sub> c<sub>α</sub> Ψ<sub>α</sub>(ξ),  ξ ~ prescribed input distribution

For uniform inputs, Legendre polynomials are used; for Gaussian inputs, Hermite polynomials. PCE is uniquely suited to uncertainty quantification: once the coefficients cα are known, the mean, variance, and Sobol sensitivity indices follow analytically from the coefficient magnitudes. In crash safety, PCE is applied to propagate manufacturing scatter (sheet thickness tolerance, material property variation) into occupant injury metrics, providing a probabilistic distribution of HIC or intrusion without Monte Carlo re-simulation. Sparse PCE via least-angle regression mitigates the curse of dimensionality when d exceeds roughly 10.

Where Machine Learning Enters

Classical surrogates assume a fixed functional form. Machine learning relaxes this: neural networks, gradient-boosted trees, and support vector regression learn the mapping directly from data. For crash CAE, the most consequential development is the deep neural network surrogate trained on field outputs — full displacement or stress fields on the mesh — rather than scalar responses. Architectures such as PointNet, graph neural networks operating on the FE mesh, and convolutional decoders can predict an entire deformed shape in milliseconds, enabling near-real-time design feedback. However, ML surrogates demand larger training sets (often 500–5,000 simulations), offer weaker extrapolation guarantees, and provide no native uncertainty estimate unless combined with ensembles or Bayesian layers.

Accelerating Design Iterations

The operational payoff is the replacement of the inner simulation loop. A typical workflow proceeds as follows: (1) define variables and bounds; (2) generate an LHS DoE of 60–200 runs; (3) execute crash simulations in batch; (4) fit and cross-validate the surrogate using leave-one-out or k-fold RMSE and R²; (5) run optimization or Monte Carlo on the surrogate at negligible cost; (6) verify the top candidates with full simulation. The critical discipline is step 6 — a surrogate is a hypothesis, not a result. Acceptance criteria commonly require R² > 0.9 and normalized RMSE < 5% on a holdout set before the surrogate is trusted for design decisions.

Practical Failure Modes and Diagnostics

Surrogates fail in characteristic ways. If cross-validated R² is high but the surrogate performs poorly on the optimization optimum, the DoE likely under-sampled the feasible boundary — add infill points near the constraint. If kriging predictions oscillate between training points, the length scale θ is too small; refit with a Matérn kernel or apply nugget regularization. If PCE coefficients do not converge as polynomial order increases, the response contains discontinuities that spectral methods cannot resolve, and a piecewise or ML approach is warranted. Finally, if an ML field surrogate produces physically implausible deformed shapes, the training set lacks diversity in deformation mode — augment with samples that trigger axial crush, bending, and mixed modes explicitly.

The remainder of this course builds on these foundations: subsequent chapters apply kriging and PCE to specific crashworthiness responses, demonstrate ML field surrogates on full-vehicle models, and establish validation protocols that keep surrogate-driven design decisions defensible.