3. Fundamentals of projection-based reduced order modeling

Projection-based reduced order modeling (ROM) replaces a high-fidelity crashworthiness model with a small set of generalized coordinates. The full-order model (FOM) may contain millions of shell and solid elements, while the ROM retains perhaps tens to a few hundred degrees of freedom. The construction pipeline has three stages: generate snapshots from full-order simulations, extract a low-dimensional basis with Proper Orthogonal Decomposition (POD), and project the governing equations onto that basis using Galerkin projection. This chapter develops each stage with concrete automotive impact examples.

3.1 Snapshot Generation

A snapshot is a full-order solution vector sampled at a specific time or parameter value. For crashworthiness, the displacement field u(t) ∈ R^n is stored at selected time steps, where n is the number of unconstrained degrees of freedom. A snapshot matrix collects these vectors column-wise:

X = [u(t₁), u(t₂), ..., u(t_m)] ∈ R^(n×m)

Snapshot quality determines ROM accuracy more than any algorithmic choice downstream. Three practical rules apply:

• Parameter coverage. If the ROM must predict multiple impact velocities, barrier angles, or material grades, snapshots must span that parameter space. A common automotive example is a frontal offset crash at 40, 56, and 64 km/h with a 40% overlap barrier.
• Temporal resolution. Crash pulses contain high-frequency content during the first 20–40 ms. Sampling every 0.5 ms captures the primary deformation modes; coarser sampling may miss buckling transitions.
• State variables. Displacement alone is often insufficient. For elastoplastic crash, the snapshot vector should include displacement, velocity, and internal state variables such as equivalent plastic strain at integration points.

In LS-DYNA or Radioss, snapshots are exported via *DATABASE_BINARY_D3PLOT at controlled intervals. A typical B-pillar intrusion study might produce 80 snapshots across 40 ms, yielding a matrix of 80 columns. Memory cost is n × m × 8 bytes; for n = 2×10⁶ and m = 80, this is approximately 1.28 GB in double precision, which is manageable but motivates compression before POD.

3.2 Proper Orthogonal Decomposition

POD finds an orthonormal basis Φ = [φ₁, φ₂, ..., φ_r] that minimizes the mean-square projection error of the snapshots. The optimal basis vectors are the left singular vectors of the centered snapshot matrix, obtained from the singular value decomposition (SVD):

X_centered = U Σ Vᵀ
Φ = U(:, 1:r)

The singular values σ₁ ≥ σ₂ ≥ ... ≥ σ_m ≥ 0 quantify the energy captured by each mode. The cumulative energy ratio determines the truncation rank r:

E(r) = Σᵢ₌₁ʳ σᵢ² / Σᵢ₌₁ᵐ σᵢ²

For crashworthiness, a common target is E(r) ≥ 0.999, but energy alone is misleading. A mode carrying 0.1% energy may still govern a local failure trigger. Engineers should inspect the spatial content of retained modes and verify that critical regions—spot welds, hem flanges, and material transitions—are represented.

When n ≫ m, the method of snapshots computes the m × m Gram matrix G = X_centeredᵀ X_centered, solves its eigenproblem, and recovers spatial modes as φᵢ = X_centered vᵢ / σᵢ. This reduces cost from O(nm²) to O(m³) and is essential for automotive models with millions of DOFs.

3.3 Galerkin Projection

The full-order semi-discrete equation of motion for crash dynamics is:

M ü + C u̇ + f_int(u) = f_ext(t)

where M is the mass matrix, C is damping, f_int is the nonlinear internal force from elastoplastic elements, and f_ext is the external crash load. The ROM approximates the solution as u ≈ ū + Φ a(t), where a(t) ∈ R^r are generalized coordinates. Substituting and enforcing orthogonality of the residual to the basis yields the reduced system:

ΦᵀMΦ ä + ΦᵀCΦ ȧ + Φᵀ f_int(ū + Φa) = Φᵀ f_ext(t)

The reduced mass and damping matrices, M_r = ΦᵀMΦ and C_r = ΦᵀCΦ, are r × r and can be precomputed. The nonlinear term Φᵀ f_int(ū + Φa) cannot be precomputed because f_int depends on the current state. Evaluating it requires reconstructing the full displacement field, computing element forces, and projecting back—an O(n) operation that dominates ROM runtime.

3.4 Worked Example: Frontal Rail ROM

Consider a front rail section modeled with 120,000 shell elements, loaded by a rigid barrier at 56 km/h. The simulation runs 50 ms with 100 snapshots. POD retains 25 modes capturing 99.95% energy. The reduced system has 25 DOFs instead of 720,000.

The intrusion error is computed at the B-pillar base node. A 3.1% error is acceptable for concept-phase design but not for certification. The speedup comes primarily from the larger stable time step and the smaller linear solve, though the nonlinear term evaluation still scales with n.

3.5 Practical Considerations and Failure Modes

Three failure modes recur in crash ROM development:

1. Insufficient snapshot diversity. A basis built only from 56 km/h frontal impacts will fail to predict 64 km/h offset crashes. Symptom: large projection error on validation snapshots. Remedy: add snapshots from the missing parameter regime and re-run POD.
2. Mode truncation too aggressive. Retaining 10 modes instead of 25 may capture 99% energy but miss local buckling. Symptom: ROM predicts smooth deformation while FOM shows localized folding. Remedy: increase r or use a weighted energy criterion that emphasizes critical regions.
3. Nonlinear term instability. Galerkin projection of f_int can produce spurious oscillations. Symptom: generalized coordinates grow unboundedly. Remedy: apply DEIM/ECSW hyper-reduction or add a stabilization term.

The next stage of the pipeline—hyper-reduction and online evaluation—builds directly on the basis and projection framework established here. The key takeaway is that snapshot design and basis truncation are engineering decisions, not automatic ones; they determine whether the ROM is a useful design tool or a misleading approximation.