4. Surrogate modeling for crash simulation

Explicit finite element (FE) crash solvers such as LS-DYNA, Radioss, or Abaqus/Explicit are the reference tools for predicting structural response under impact. A single full-vehicle frontal or side impact model can contain 5–15 million elements and require 6–48 wall-clock hours on a 64–256 core cluster. When a design team needs to explore hundreds of thickness, material, or topology variants, or when an optimizer must evaluate thousands of candidates, running the FE solver for every query becomes the bottleneck. Surrogate modeling — also called metamodeling or response surface modeling — replaces the expensive solver with a cheap mathematical approximation trained on a limited set of FE runs. This chapter explains when a surrogate is the right engineering choice, how to build one that is trustworthy for crash-relevant outputs, and how to validate it before it is allowed to drive a design decision.

When a Surrogate Is the Right Tool

A surrogate is justified when three conditions hold simultaneously. First, the design space is well bounded: the parameters that vary (e.g., B-pillar thickness, material grade, bead depth) are known and their ranges are fixed. Second, the output of interest is a scalar or low-dimensional quantity — peak intrusion, HIC, peak acceleration, mass, or a binary pass/fail against a regulatory threshold. Third, the FE model itself is stable and validated; a surrogate trained on a noisy or unconverged solver is worthless. If any of these fails, invest in solver throughput or model simplification instead.

Typical crash-surrogate use cases include:

• Design space exploration — mapping how intrusion and mass trade off across a thickness sweep before committing to detailed FE runs.
• Optimization loops — genetic algorithms or Bayesian optimizers that need 10³–10⁵ evaluations, far beyond FE budget.
• Real-time or on-vehicle estimation — reduced-order models embedded in a restraint control unit or a digital twin.
• Sensitivity and robustness studies — Monte Carlo sampling over manufacturing tolerances.

Conversely, a surrogate is a poor fit when the output is a full-field contour (stress distribution over the whole body), when the design space includes discrete topology changes that the training set never saw, or when the required accuracy is tighter than the FE discretization error itself.

Choosing the Surrogate Family

Crash responses are strongly nonlinear and often non-smooth: a structure may switch from stable folding to global buckling between two nearby parameter values. This rules out simple linear regression. The practical families are:

For a first crash surrogate with fewer than 200 FE runs, Kriging is usually the best default because it returns both a prediction and a variance, which lets the optimizer balance exploitation against exploration. Neural networks only become competitive once the sample set is large enough to justify them.

Building the Training Set: Design of Experiments

The quality of a surrogate is bounded by the quality of its training samples. Random sampling wastes runs in already-well-covered regions. Use a space-filling design instead:

1. Define each parameter with a lower and upper bound, e.g. B-pillar thickness t ∈ [1.2, 2.4] mm, material yield σy ∈ [340, 780] MPa, bead depth d ∈ [0, 12] mm.
2. Generate an initial design with Latin Hypercube Sampling (LHS) or a maximin criterion. For three parameters, 40–60 samples is a reasonable starting budget.
3. Run the FE solver for each sample and record the scalar outputs. Keep the raw d3plot or equivalent in case the surrogate must be retrained.
4. Optionally add a few samples near expected discontinuities (e.g., the thickness where the pillar transitions from folding to bending) to reduce interpolation error there.

A minimal LHS generation in Python looks like this:

import numpy as np
from scipy.stats import qmc

# 3 parameters: thickness, yield, bead depth
bounds = np.array([[1.2, 2.4], [340.0, 780.0], [0.0, 12.0]])
sampler = qmc.LatinHypercube(d=3, seed=42)
u = sampler.random(n=50)                 # 50 samples in [0,1]^3
X = qmc.scale(u, bounds[:, 0], bounds[:, 1])
np.savetxt("doe_lhs.csv", X, delimiter=",", header="t,sy,d", comments="")

Each row of doe_lhs.csv becomes one LS-DYNA keyword deck. After the runs complete, the output vector y (peak intrusion, HIC, mass) is paired with X to form the training set.

Fitting and Validating the Surrogate

Never trust a surrogate that has only been checked on its training data. Use k-fold cross-validation: split the 50 samples into 5 folds, train on 40, predict the held-out 10, and repeat. The relevant error metric for crash work is not R² alone but the normalized root-mean-square error (NRMSE) and the maximum absolute error at the design-relevant region:

NRMSE = sqrt( mean( (y_pred - y_true)^2 ) ) / (max(y_true) - min(y_true))

As a rule of thumb for regulatory-relevant outputs such as peak intrusion, require NRMSE < 5% and maximum error < 10% of the allowable threshold. If the surrogate fails, add samples in the region of largest residual rather than retraining blindly.

Worked Example: B-Pillar Intrusion Surrogate

Consider a side-impact model where the objective is to minimize peak B-pillar intrusion Imax while keeping mass m below a target. Three parameters vary: thickness t, yield σy, and bead depth d. Fifty LHS samples are run through LS-DYNA; each run takes about 4 hours on 96 cores, so the total cost is roughly 200 core-hours. A Kriging model is fitted to Imax and m separately.

Five-fold cross-validation gives NRMSE = 3.8% for intrusion and 1.1% for mass. The maximum error on intrusion is 6.2 mm against a 150 mm threshold, which is acceptable. A multi-objective genetic algorithm then evaluates 20,000 candidate designs against the surrogate in under a minute, returning a Pareto front. The three most promising candidates are re-run in the full FE solver; the surrogate-predicted intrusion is within 4 mm of the FE result, confirming the surrogate is fit to drive the trade study.

Operational Guardrails

• Version the surrogate with its training data. A surrogate is only valid for the parameter ranges and FE model revision it was trained on. Store the DOE table, the solver version, and the fitted model together.
• Re-validate after any FE model change. A mesh refinement or a new material card invalidates the old surrogate.
• Keep a hold-out set. Reserve 10–20% of runs that are never used for fitting, only for final acceptance.
• Log every surrogate-driven decision. When an optimizer selects a design based on the surrogate, record the predicted value and the eventual FE confirmation so the surrogate's field accuracy can be audited over time.

The surrogate does not replace the crash solver — it replaces the thousands of solver calls that would otherwise be needed to search a design space. The solver remains the source of truth; the surrogate is the fast, cheap filter that tells the engineer where to spend the next expensive run.