2. Bayesian optimization overview

Bayesian optimization (BO) is a sample-efficient, model-based strategy for optimizing expensive black-box functions. In vehicle crashworthiness design, a single high-fidelity finite element (FE) simulation — for example, a full-vehicle frontal offset impact model with roughly 2–5 million elements — may require 4–12 hours of wall-clock time on a 64-core cluster. Gradient information is unavailable or unreliable because the response (e.g., peak intrusion, HIC, or mass) is computed through contact, plasticity, and failure algorithms that are non-differentiable. Under these conditions, gradient-based optimizers and naive design-of-experiments (DOE) sweeps are impractical. BO addresses this by building a cheap probabilistic surrogate of the objective and using it to decide, one evaluation at a time, where the next simulation should be run.

The three components of a BO loop

A BO algorithm consists of three tightly coupled elements:

• A probabilistic surrogate model that approximates the true response f(x) and, critically, quantifies its own uncertainty at unobserved points.
• An acquisition function α(x) that scores candidate designs by combining predicted mean and predicted uncertainty into a single scalar utility.
• An optimization sub-problem that maximizes α(x) over the design space to select the next evaluation point.

The loop proceeds as follows: fit the surrogate to all evaluated designs; maximize the acquisition function to propose x_next; run the expensive simulation at x_next; augment the dataset; repeat until the evaluation budget is exhausted or a convergence criterion is met.

Probabilistic surrogate modeling with Gaussian processes

The most widely used surrogate in BO is the Gaussian process (GP). A GP defines a distribution over functions, fully specified by a mean function m(x) and a covariance (kernel) function k(x, x'). Given n observed design–response pairs D = {(x_i, y_i)}, the posterior predictive distribution at a new point x* is Gaussian with:

μ(x*) = m(x*) + k*ᵀ (K + σ²I)⁻¹ (y − m)
σ²(x*) = k(x*, x*) − k*ᵀ (K + σ²I)⁻¹ k*

Here K is the n × n kernel matrix, k* is the vector of covariances between x* and the training points, and σ² is the noise variance (set small for deterministic FE simulations). The kernel encodes assumptions about smoothness and length scale. For crashworthiness responses such as intrusion or acceleration, the Matérn 5/2 kernel is a common choice because it balances smoothness with the ability to capture local nonlinearity:

k(x, x') = σ_f² (1 + √5 r/ℓ + 5r²/(3ℓ²)) exp(−√5 r/ℓ)

where r = ||x − x'|| and ℓ is the characteristic length scale. Hyperparameters (σ_f, ℓ, σ²) are typically estimated by maximizing the log marginal likelihood. In practice, for a 6-variable crash model (e.g., thicknesses of the front rail, bumper, shotguns, and B-pillar), a GP with an anisotropic Matérn 5/2 kernel and automatic relevance determination (ARD) can be fitted in seconds, making the surrogate essentially free compared to the FE simulation.

Acquisition functions: turning uncertainty into decisions

The acquisition function is where the exploration–exploitation trade-off is explicitly encoded. It uses the GP posterior to assign a score to every candidate x. Three widely used acquisition functions are:

In the formulas above, Φ and φ are the standard normal CDF and PDF, f_best is the best observed value so far, and ξ is a small non-negative parameter that encourages exploration when improvement is uncertain. For crashworthiness minimization (e.g., minimizing mass subject to intrusion constraints), f_best is the lowest feasible objective value found so far.

The exploration–exploitation trade-off in engineering design

Exploration means sampling where the surrogate uncertainty σ(x) is large, which improves the model globally and reduces the risk of missing a better design region. Exploitation means sampling where the predicted mean μ(x) is favorable, which refines the current best solution. A purely exploitative strategy can converge prematurely to a local optimum; a purely explorative strategy wastes expensive simulations on regions that are unlikely to contain the optimum.

In vehicle crashworthiness, this trade-off has a concrete physical interpretation. Exploration corresponds to testing design regions where the FE model has not been exercised — for example, a combination of a very thin front rail and a thick B-pillar that has never been simulated. Exploitation corresponds to fine-tuning a promising configuration, such as adjusting the front rail thickness from 1.8 mm to 1.9 mm when the current best design already meets intrusion targets. The acquisition function automatically balances these two impulses: EI, for instance, is large when either the predicted improvement is substantial (exploitation) or the uncertainty is high (exploration).

Practical workflow and parameter guidance

When applying BO to a crashworthiness problem, the following concrete steps are typical:

1. Define variables and bounds. For a frontal impact model, select 4–8 thickness variables (e.g., front rail t₁ ∈ [1.2, 2.4] mm, bumper beam t₂ ∈ [1.5, 3.0] mm, B-pillar t₃ ∈ [1.0, 2.2] mm). Normalize all variables to [0, 1] before fitting the GP.
2. Generate the initial DOE. Use a Latin Hypercube sample of n₀ = 10–15 points. Run the FE simulations and record the objective (e.g., mass) and constraints (e.g., maximum intrusion ≤ 150 mm, HIC ≤ 700).
3. Fit the GP. Use a Matérn 5/2 kernel with ARD. Optimize hyperparameters by maximizing the log marginal likelihood. In Python, scikit-learn provides GaussianProcessRegressor; for constrained problems, use a constrained EI variant or penalty-based EI.
4. Select the next point. Maximize EI using a multi-start L-BFGS-B or a fine grid search over the normalized design space. For constrained problems, multiply EI by the probability of feasibility: α_cEI(x) = EI(x) · P(g(x) ≤ 0).
5. Run the simulation and update. Add the new point to the dataset and refit. Repeat until the budget (e.g., 50–80 FE runs) is exhausted or EI falls below a threshold (e.g., 10⁻³ of the current best objective).

Why BO is well-suited to crashworthiness

Crashworthiness optimization typically involves 5–15 design variables, a simulation budget of 30–100 runs, and a mix of continuous and categorical variables (e.g., material grades, trigger types). BO excels in exactly this regime: it is sample-efficient, handles black-box responses, and naturally incorporates constraints through acquisition function modifications. Compared to a full factorial or Taguchi DOE, BO can reduce the number of FE simulations by 40–60% while achieving comparable or better designs — a decisive advantage when each simulation costs hours of cluster time.

The next section will build on this foundation by examining how to choose and tune acquisition functions for constrained, multi-objective crashworthiness problems, and how to validate the resulting optimum against the FE model.