4. Airbag deployment kinematics

An airbag is not a simple balloon. It is a folded membrane structure driven by a pyrotechnic or hybrid inflator, and its deployment is a strongly coupled fluid–structure interaction (FSI) event that unfolds over roughly 20–50 ms. For reduced-order modeling (ROM), this stage is where the input space explodes: fold topology, fabric material, venting, inflator mass flow, and occupant proximity all interact nonlinearly. This chapter explains the physical mechanisms that generate that high-dimensional input space, so you can decide which parameters to sample, which to fix, and which to represent through latent variables.

1. Folding Patterns and Their Mechanical Consequences

Folding is the process of compacting the cushion into the module so it fits behind the cover and deploys along a controlled trajectory. The pattern determines the initial contact sequence between fabric layers, the unfolding path, and the local stress concentrations that drive early-time behavior. Common patterns include:

• Accordion (concertina) fold: the cushion is pleated in alternating layers. It unfolds axially with low lateral spread and is common in driver-side modules.
• Rolled fold: the cushion is rolled from the outer edge inward. It produces a sweeping, rotational deployment and is often used for passenger-side or side airbags.
• Z-fold / reverse fold: a folded flap that reverses direction, used to control the timing of specific regions (e.g., delaying the lower chamber).
• Petal / star fold: radial pleats that open outward symmetrically, typical for side curtains and some knee airbags.

From a ROM perspective, the fold pattern is not a single scalar. It is a topological input: the number of pleats, pleat depth, fold-line positions, and the order in which layers separate all matter. Two cushions with identical geometry but different fold sequences can produce measurably different early pressure traces and occupant contact timing. In practice, you encode the fold as a set of parameters (pleat count, fold angle, layer ordering) or, more compactly, as a low-dimensional latent vector obtained from a shape descriptor of the folded state.

2. Inflator Gas Dynamics

The inflator supplies mass flow and enthalpy. Two families dominate: pyrotechnic inflators (solid propellant, fast burn, high gas temperature) and hybrid inflators (stored compressed gas plus a small pyrotechnic heater). The governing behavior is captured by the mass and energy balance of the cushion control volume:

dm/dt = m_in(t) - m_out(t)
d(mu)/dt = m_in*h_in - m_out*h_out - p*dV/dt + Q_loss

where m_in(t) is the inflator mass flow rate, m_out(t) is venting and leakage, h is specific enthalpy, p*dV/dt is the work done expanding the cushion, and Q_loss accounts for heat transfer to fabric and ambient. The mass flow itself follows a choked-flow relation through the inflator throat:

m_in(t) = Cd * A_throat * p_tank / sqrt(R*T) * f(gamma, p_ratio)

For a pyrotechnic inflator, m_in(t) is a burn-rate-driven pulse with a rise time of a few milliseconds; for a hybrid, it is a decaying blowdown from the stored pressure. The key ROM-relevant point is that the inflator is a time-varying boundary condition, not a constant. Sampling only peak mass flow misses the rise-time and duration, which control the pressure ramp and therefore the fabric strain rate.

Venting is equally important. Vent holes, fabric porosity, and seam leakage all contribute to m_out. A common modeling error is to treat venting as a fixed orifice; in reality the vent area changes as the cushion inflates and the fabric stretches, so venting is coupled to the deformation state.

3. The High-Dimensional Input Space

Combining folding, inflator, material, and environment yields a large parameter set. A representative enumeration for a driver-side airbag ROM study is shown below. The "ROM role" column indicates how each parameter is typically handled.

Even this trimmed set gives 10+ active dimensions. A full-factorial design is infeasible, and naive random sampling wastes runs in regions where the response is flat. The standard remedy is a two-stage approach: first, a screening design (e.g., Latin hypercube or Sobol sequence) to rank parameter influence via variance-based sensitivity; second, a focused design on the top 4–6 parameters, with the rest fixed at nominal values or represented through a single lumped parameter.

4. From Input Space to Reduced Representation

The central insight for ROM construction is that the high-dimensional input space is not uniformly important. Deployment kinematics are dominated by a low-dimensional manifold: the interaction between inflation rate and unfolding resistance. You can exploit this by projecting the input vector onto a small set of latent variables—for example, a "deployment aggressiveness" index combining peak mass flow and rise time, and an "unfolding resistance" index combining pleat depth and fabric stiffness. The ROM then maps these latent variables to output quantities of interest (pressure trace, contact force, bag shape at key instants).

The figure below summarizes how the physical inputs flow through the coupled mechanisms into the outputs that the ROM must reproduce.

5. Practical Screening Workflow

A reproducible screening procedure for a driver-side airbag ROM looks like this. Assume a baseline model with the nominal values from the table above. The goal is to identify the 4–6 parameters that dominate the pressure trace and contact force.

1. Define the output metrics. Extract from each simulation: time to 50% peak pressure, peak pressure, time to full deployment (bag reaches 95% of final volume), and peak contact force. These are the quantities the ROM must reproduce.
2. Build a screening design. Use a 256-run Latin hypercube over the 10 active parameters. Keep the fold layer ordering fixed at nominal to avoid categorical explosion in the first pass.
3. Run and post-process. For each run, compute the four metrics. Normalize each metric to its nominal value so sensitivities are comparable across units.
4. Rank influence. Compute first-order Sobol indices (or a cheaper Morris screening if runs are expensive). Expect peak mass flow and rise time to dominate pressure metrics, and pleat depth plus fabric stiffness to dominate deployment time.
5. Refine. Fix the bottom 4–5 parameters at nominal, and run a focused 128-run design on the top 5. This becomes the training set for the ROM.

6. What the ROM Must Capture

A reduced-order representation of deployment kinematics is only useful if it preserves the causal chain: inflator pulse → pressure ramp → unfolding rate → shape evolution → occupant contact. If your ROM reproduces the pressure trace but not the deployment timing, it will fail when coupled to an occupant model, because contact force depends on when the bag reaches the occupant, not just how hard it pushes. The latent-variable formulation described above is attractive precisely because it preserves this chain: both latent axes are defined in terms of the physical mechanisms, so the ROM's outputs remain interpretable and extrapolate more gracefully than a black-box fit over raw parameters.

When validating, compare not only scalar metrics but the full time histories at a few design points held out from training. A ROM that matches peak values but shifts the rise time by 3–4 ms is not ready for occupant coupling, even if its aggregate error looks acceptable.