3. Reduced-order modeling taxonomy for restraint systems

Restraint system simulation sits at an awkward intersection: the physics is highly nonlinear (belt slack, retractor locking, airbag deployment, contact), yet the design loop demands hundreds or thousands of runs for occupant safety optimization. A full explicit finite element model of a belted occupant with a deploying airbag can take 30–120 minutes per run on a compute node, which makes direct DOE or optimization impractical. Reduced-order models (ROMs) close this gap, but "ROM" is not one method — it is a family of techniques with very different assumptions, costs, and failure modes. Choosing the wrong member of that family is the most common reason a ROM project fails to deliver. This chapter builds a working taxonomy of the four dominant approaches used in restraint system engineering: static condensation, modal truncation, proper orthogonal decomposition (POD), and data-driven surrogates.

What distinguishes the four families

Two axes organize the taxonomy. The first is where the reduction happens: inside the governing equations (physics-based reduction) or outside them (input–output mapping). Static condensation and modal truncation reduce the structural matrices themselves; POD reduces the solution space of a parameterized simulation; surrogates replace the solver entirely with a fitted function. The second axis is what is preserved: physical interpretability, time-history fidelity, or only scalar outputs of interest (OOI) such as HIC, chest deflection, or belt force.

These axes matter because restraint engineers rarely need the full field. They need peak values, time histories at a handful of channels, and trends across a design space. A method that preserves the full displacement field but costs 20 minutes per run is worse than a surrogate that predicts HIC within 5% in 10 milliseconds — provided the surrogate's validity domain is respected.

Static condensation: exact at the boundary, approximate inside

Static condensation (Guyan reduction, and its dynamic extension Craig–Bampton) partitions the degrees of freedom into a retained set b (boundary/interface DOF) and an eliminated set i (interior DOF). The interior displacement is approximated as a static response to the boundary motion:

K_bb  K_bi   u_b     f_b
K_ib  K_ii   u_i  =  f_i

u_i ≈ -K_ii^-1 K_ib u_b
K_red = K_bb - K_bi K_ii^-1 K_ib
M_red = T^T M T,   T = [ I ; -K_ii^-1 K_ib ]

In restraint work this is the natural choice for the seat structure and belt anchorages: the mounting points to the vehicle body are retained exactly, while the seat frame interior is condensed. A typical seat frame model with 180,000 DOF can be reduced to 4,000–8,000 DOF with boundary DOF preserved, cutting a linear eigenvalue solve from minutes to seconds. The critical caveat: the reduction is exact only for static loading. Under crash-level inertia, the neglected inertia of the interior DOF produces errors that grow with frequency. Craig–Bampton adds fixed-interface normal modes to the basis to recover part of this, but the method still assumes linearity — it cannot represent belt slack or retractor lock-up, which are the very phenomena restraint engineers care about. Use static condensation for the structure, not for the restraint interaction.

Modal truncation: a basis chosen by eigen-analysis

Modal truncation projects the equations onto the lowest n eigenvectors of the undamped system, Φ_n, giving diagonal reduced mass and stiffness. The engineering decision is the mode count. A practical rule for belt and seat structures is to retain modes up to 1.5× the highest excitation frequency of interest; for a 50 ms crash pulse with meaningful content to ~80 Hz, retaining modes to 120 Hz typically requires 30–60 modes for a seat frame. The reduced model then integrates in modal coordinates, which is cheap and numerically stable.

The failure mode is well known: modal truncation is a linear reduction. It handles the belt webbing's elastic stretch and the seat frame's bending well, but it cannot capture the unilateral contact between occupant and airbag, the friction in the retractor, or the geometric stiffening during belt spool-out. In practice, modal truncation is used for the substructure (seat, belt segments modeled as nonlinear springs) and coupled to a nonlinear occupant model through interface forces — a hybrid that keeps the linear reduction where it is valid.

Proper orthogonal decomposition: learning the solution space

POD takes a different route. Instead of deriving a basis from the operator, it derives one from data: run the full simulation at m parameter samples, collect snapshots u(x, t; μ_k), assemble the snapshot matrix S, and compute its singular value decomposition. The leading left singular vectors form the reduced basis. The rank r is chosen by an energy criterion, e.g. retain the smallest r such that the captured energy exceeds 99.9%:

S = [u(t1;μ1) ... u(tN;μM)]        # columns = snapshots
S = U Σ V^T
energy(r) = sum(σ_1..σ_r)^2 / sum(σ_all)^2  ≥ 0.999
u(x,t;μ) ≈ Σ_{j=1..r} a_j(t;μ) φ_j(x)

For a belt-pull or airbag-deployment study, a snapshot set of 40–80 full runs across the parameter space (belt stiffness, anchor position, inflator mass flow) typically yields r = 15–40 modes capturing 99.9% of the displacement energy. The reduced coefficients a_j are then obtained by projecting the governing equations (Galerkin POD) or by interpolating them across parameters. POD's strength is that it captures the actual solution manifold, including nonlinear effects present in the training runs. Its weakness is equally clear: the basis is only valid inside the training envelope. Extrapolating a POD model to a belt stiffness 30% outside the sampled range can produce non-physical oscillations, because the basis has no information about that regime.

Data-driven surrogates: mapping inputs to occupant injury metrics

The final family abandons field reconstruction entirely. A surrogate fits a function from design inputs (belt force limiter level, anchor coordinates, airbag vent size, seatback angle) to scalar outputs (HIC₁₅, chest deflection, neck force, belt peak force). Common choices are Gaussian process regression (good for 10–100 samples, provides uncertainty), radial basis functions, polynomial response surfaces, and neural networks (need hundreds to thousands of samples).

A reproducible example: sample five inputs with a Latin hypercube of 60 runs, hold out 12 runs for validation, fit a GP with a Matérn 5/2 kernel, and report R² and normalized RMSE on the hold-out set. Acceptance criteria commonly used in occupant safety are R² > 0.95 and NRMSE < 5% for HIC and chest deflection. If the hold-out error exceeds this, the usual causes are: too few samples for the input dimensionality (60 samples in 5D is marginal), a discontinuous response (retractor lock timing creates step changes that GP smooths away), or an input range that spans a regime change (airbag bottoming-out). The diagnostic is to plot predicted vs. actual and residual vs. each input; a funnel shape indicates heteroscedastic noise, a step in residuals indicates a missed regime boundary.

Choosing among the four

The taxonomy is not a hierarchy of quality — it is a map of trade-offs. Static condensation and modal truncation buy speed by assuming linearity; POD buys accuracy inside a known envelope by paying for a snapshot campaign; surrogates buy the most speed by giving up field information entirely. The next chapters apply these methods to specific restraint subsystems, beginning with the belt and retractor, where the choice between modal and POD reduction is decided by how strongly the force limiter nonlinearity dominates the response.