3. Crash simulation model setup

Occupant injury sensitivity analysis is only as trustworthy as the model it interrogates. Before a single parameter is perturbed, the crash simulation must be assembled so that every input that will later be varied is exposed as a controllable variable, and every output that will later be ranked is recorded at a well-defined location and in a well-defined coordinate frame. This section describes how to build that model: the restraint system representation, the dummy model, and the boundary conditions that make a sensitivity study both physically meaningful and numerically stable.

Coordinate and Sign Conventions

All quantities in this chapter follow the vehicle coordinate system: X positive forward, Y positive to the vehicle's right, Z positive upward, with the origin at the vehicle datum. Occupant kinematics are reported in this frame unless stated otherwise. Accelerations are positive along the positive axis direction, so a frontal impact produces a negative X acceleration at the vehicle tunnel. Injury criteria follow the standard definitions (HIC36, chest deflection in millimetres, Nij, and femur force in kilonewtons); keeping these conventions fixed across every run is essential, because a sign flip in a single output channel will silently corrupt the ranking.

Restraint System Models

The restraint system is where most sensitivity parameters live, so it deserves the most careful modelling. A typical frontal model contains a seat belt with retractor, slip ring, buckle, and anchor points; a driver airbag; a passenger airbag; a steering column; and a seat structure. Each of these must be parameterised rather than hard-coded.

For the belt, the retractor is modelled with a spool inertia and a force-limiting torsion bar. The two parameters that dominate occupant response are the force limiter level (typically 2–5 kN) and the pretensioner firing time and stroke. In the model these appear as a load-curve definition and a prescribed retraction, for example:

*ELEMENT_SEATBELT_RETRACTOR
$  sid     pid    sid2   pull   llr    elatch
   101     201     0     1.0    0.0    0
$  load curve for force limiter (time [ms], force [kN])
   301
   0.0    0.0
   5.0    3.5
  60.0    3.5
 120.0    3.5

The airbag is represented by a uniform-pressure gas model with a mass-flow-in curve and a vent-hole area. The parameters that matter for sensitivity are vent hole area, inflator mass flow scaling, and bag volume. A common mistake is to tune the airbag to match a single sled test and then treat the tuned values as fixed; for sensitivity work the vent area and mass-flow multiplier should be exposed as design variables with realistic ranges (for example, vent area 0.8–1.2× nominal, mass-flow multiplier 0.9–1.1×).

The steering column contributes through its collapse force and stroke. These are modelled as a translational joint with a force-displacement curve. Because the column interacts with the airbag and the belt, its parameters are not independent of the restraint parameters, which is precisely why a global sensitivity method is needed rather than one-at-a-time perturbation.

Dummy Models

The dummy model choice determines which injury criteria can be evaluated and how much run time each sample costs. For frontal sensitivity studies the Hybrid III 50th percentile male is the baseline; the THOR-50M is preferred when chest deflection and rib-strain-based criteria are of interest, at roughly two to three times the computational cost. The dummy must be positioned with a documented procedure: H-point location, seat back angle, pelvis angle, and steering-wheel-to-chest distance all affect the response and must be held constant across the design space, or they become uncontrolled noise.

Two practical rules apply. First, the dummy's contact definitions must be checked against the restraint geometry at the nominal position; a belt that grazes a shoulder surface in the nominal run may penetrate it in a perturbed run, producing a spurious force spike that dominates the ranking. Second, output channels should be requested at the standard measurement locations (head CG accelerations, upper and lower chest accelerations, chest deflection, femur loads) and written at a fixed sampling rate, typically 10 kHz, so that HIC integration is consistent between runs.

Boundary Conditions

The boundary conditions define the loading the occupant sees and the numerical environment in which the solver operates. For a full-vehicle frontal model, the barrier impact speed, the pulse at the rocker, and the seat mounting stiffness are the dominant boundary parameters. For a sled-based sensitivity study, the sled pulse itself becomes a boundary condition and should be parameterised by its peak deceleration and duration rather than by a fixed table.

Numerical boundary conditions matter as much as physical ones. The model should use a consistent time step, with the belt and airbag meshes sized so that the critical time step is not dominated by a single small element. Contact thickness and penalty stiffness should be set explicitly for the dummy-to-restraint interfaces. Gravity and initial velocity must be applied consistently; a model that initialises the dummy at rest but applies a pulse starting at t = 0 will show a non-physical pre-load.

Preparing the Model for Parameter Sweeps

The final setup step is to make the model scriptable. Every parameter destined for the study should be referenced through a single include file or a parameter block, so that a sampling script can rewrite values without editing the main deck. A minimal parameter block for a frontal study might look like this:

$ sensitivity_parameters.inc
$ belt force limiter level [kN]
#define FL_LEVEL   3.5
$ pretensioner firing time [ms]
#define PT_TIME    12.0
$ airbag vent area scale [-]
#define VENT_SCALE 1.00
$ steering column collapse force [kN]
# define COL_FORCE  4.0
$ pulse peak deceleration [g]
# define PULSE_PEAK  35.0

With this structure, a Latin hypercube or Sobol sampling script can generate N parameter sets, write each to the include file, launch the solver, and archive the output channels. The nominal run should be retained as the reference against which every perturbed run is compared, and the output archive should record the parameter vector alongside the injury metrics so that the ranking stage has a complete, traceable dataset.

Worked Setup Example

Consider a frontal sensitivity study with five parameters: belt force limiter level (2.5–4.5 kN), pretensioner firing time (8–16 ms), airbag vent area scale (0.8–1.2), steering column collapse force (3–5 kN), and pulse peak (30–40 g). A Latin hypercube sample of 64 runs plus the nominal case is generated. Each run writes the parameter vector and the four injury metrics to a CSV archive. The nominal run is checked first: if the belt slips or the energy error exceeds 5%, the run is discarded and the setup is corrected before any ranking is attempted. Only when all 65 runs pass the pre-run checks is the archive handed to the sensitivity analysis stage, where variance-based indices will be computed for each parameter–metric pair.

The setup described here is deliberately conservative: it fixes everything that is not a design variable, parameterises everything that is, and validates the nominal case before sampling. That discipline is what allows the later ranking to be interpreted as genuine parameter influence rather than as a mixture of physical sensitivity and modelling noise.